Flashover prediction of polluted composite insulators based on arc time constant and velocity using CNN–LSTM Authors: Navid Fahimi, Hamid Reza Sezavar Published: 2026 ABSTRACT In this paper, a unified arc model is proposed to bridge the gap between analytical arc models and flashover dynamics in real practice. This approach synergizes the arc time constant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}) and arc velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}) under pollution conditions. While conventional static models fail to capture the transient nature of arc propagation on polluted insulators, proposed approach would address this point by dynamic coupling of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}. Presented model is able to make transition smoothly between discharge stages with adjustment in real time using leakage current (LC) data. The procedure of LC waveform feature extraction is achieved using combination of convolutional neural network (CNN) and long short-term memory (LSTM). The presented arc model ensures fail-safe predictions while the normalized arc velocity parameter provides a physics-based foundation for understanding arc progression. The statistical validation of error distributions approves the reliability and extension of the proposed approach. In addition, the presented CNN–LSTM model achieves superior performance predictions and computational efficiency to some similar benchmarked approaches. The practical implications of these findings would enable preventive maintenance (PM), enhanced grid protection and optimized insulator designs. Introduction Outdoor insulators play a critical role in reliable performance of electrical power systems. They are expected to withstand breakdown in harsh service conditions such as wet conditions, coastal or industrial polluted zones, dust storms and etc.1,2. Combination of pollution deposit with wet conditions on the surface of outdoor insulators can lead to flashover incidents in polluted zones which can cause unpredictable costly outages in power systems3,4. Arc time constant and arc propagation velocity are among critical challenging arcing characteristics for dynamic modeling of pollution flashover. It should be considered that partial arcs propagate in different stochastic paths on the surface of wet polluted insulator due to geometrical specification and non-uniform distribution of pollution. In addition, velocity of local arcs is different as well depending on pinch force of LC, thermal buoyancy due to temperature variation, and non-uniform distribution of charge density5. According to measurements of previous researches, initial stage arcs propagate with less velocities in lower applied voltages. Also, upper arcs close to high voltage terminal propagate faster than lower arcs (close to ground terminal)6,7. Arcing time constant is considered particularly in Mayr’s equation of arcing resistance for first time8. Mayr’s arc model presumes arcing temperature and dimension to be constant. Some famous later works include Cassie’s arc model9 for larger currents, Habedanck’s hybrid Cassie-Mayr representation of arc model10, Schavemaker’s arc formulation of modified Mayr’s with a constant time and cooling power11, and Schwartz’s arc model of modified Mayr’s dependent on arc conductance12. Variation of arc time constant during dynamic stages of arc activities is a challenging issue in most previous arc models. Arc time constant is usually referred to electrical discharge inertia due to conservation nature of arcing energy according to conductance and injected power. In measurements of the past and recent researches, value of arc time constant is reported from 100 µS in initial arc stage of lower currents to several ms in pre-flashover stage of larger currents13–15. An improved arc model is represented in Khakpour et al.16 based on arc diameter for calculation of low-current arcing time constant. In a recent study, novel flashover model is presented based on local arc time constant according to thermal and electrical analysis of local AC arcs in polluted insulators17. Arcing propagation velocity is interested by many researches. Some popular arcing velocity criterions include Gallimberti’s criterion18 based on required charge to induce arc propagation, Beroual’s criterion19,20 based on the dissipated energy by the arc, and Anjan and Lakshminarassimh’s criterion21 based on arc mobility and electrical field inside arc. Recently, an arc propagation velocity criterion is proposed in Taheri et al.7 for calculation of Wilkin’s22 pollution layer equation based on experimental measurements. In Jabbari et al.23, propagation velocity curves versus arcing time and length are obtained for flashover voltage (FOV) prediction of insulators according to development dynamic arc model. Similarly, influence of arc propagation velocity is investigated in dynamic arc model for prediction of AC flashover of insulators in Jabbari et al.24. In Salem et al.25, a modified discharge model of polluted insulator is proposed based on sensitivity analysis of arc length, arc time constant and arc propagation velocity according to hybrid calculation of circuit model and genetic algorithm (GA). In Parizad et al.26, optimization of arc parameters is done by use of GA according to comparative analysis of several popular arc models. Machine learning (ML) methods are proved to be influential for flashover prediction in condition-based analysis of composite insulators. In27,28, ML techniques such as random forest (RF), support vector machine (SVM), K-nearest neighbor (KNN), multi-layer perceptron (MLP) and deep neural network (DNN) are compared for contamination classification of polluted high voltage insulators using LC. In29,30, artificial neural network (ANN) is used to develop FOV based on pollution severity, LC data and geometry profile of the insulators. Similarly, ANN is used to build an intelligent classifier for classification of defective insulators based on partial discharge (PD) monitoring31. In Nguyen et al.32, danger-assessment classification of polluted insulators is presented based on deep learning according to ultra-high frequency (UHF) signals of PDs. In Bueno-Barrachina et al.33, convolutional neural network (CNN) is utilized for LC prediction of outdoor insulators based on long-term data. In Salem et al.34, RUSBoost algorithm (combination of random-under sample and AdaBoost) is used to estimate key indicators of insulators based on environmental data. In Santos et al.35, self-organizing map (SOM) is employed for clustering flashover and LC datasets of damaged insulators. In Ghiasi et al.36, a cloud data classified system of LC is suggested for preventing flashover of outdoor insulators based on deep learning technique. While most previous works focused on failure prevention of outdoor insulators, but feature extraction of outdoor insulators based on arc propagation physics has been often missed. Joint approach of dynamic discharge modeling for arc propagation understanding and ML methods for handling computational challenges might help to make better insight into future behavior of outdoor insulators in harsh service conditions. In this paper, a comprehensive analysis is presented based on joint arc physics and deep learning method according to numerous test data of the real outdoor composite insulators for wide range of pollution levels. Contributions of this hybrid arc model can be summarized as follows:Development of a comprehensive arc model that integrates arc time constant and arc velocity to predict flashover under varying pollution levels.To achieve high accuracy and real-time applicability, a hybrid machine learning (ML) model is proposed in this paper using combination of physics-informed feature engineering with deep learning.Presented dynamic modeling approach considers time-dependent variations in arc time constant and arc velocity.Experimental validation ensures that the model aligns with empirical data for all pollution levels.Presented hybrid physics-deep learning approach translates analytical insights into practical tools for preventive maintenance (PM) schemes of the power systems. Development of a comprehensive arc model that integrates arc time constant and arc velocity to predict flashover under varying pollution levels. Development of a comprehensive arc model that integrates arc time constant and arc velocity to predict flashover under varying pollution levels. To achieve high accuracy and real-time applicability, a hybrid machine learning (ML) model is proposed in this paper using combination of physics-informed feature engineering with deep learning. To achieve high accuracy and real-time applicability, a hybrid machine learning (ML) model is proposed in this paper using combination of physics-informed feature engineering with deep learning. Presented dynamic modeling approach considers time-dependent variations in arc time constant and arc velocity. Presented dynamic modeling approach considers time-dependent variations in arc time constant and arc velocity. Experimental validation ensures that the model aligns with empirical data for all pollution levels. Experimental validation ensures that the model aligns with empirical data for all pollution levels. Presented hybrid physics-deep learning approach translates analytical insights into practical tools for preventive maintenance (PM) schemes of the power systems. Presented hybrid physics-deep learning approach translates analytical insights into practical tools for preventive maintenance (PM) schemes of the power systems. The rest of the paper is organized as follows. “Test procedure” Section describes test procedure. “Analytical formulation of arcing parameters” Section explains presented arc model based on hybrid application of analytical arc model and machine learning approach. Results and discussion are presented in “Results and discussion” Section. Finally, the conclusion is presented in “Conclusions” Section. Test procedure The experiment setup is shown in Fig. 1 that consists of a cubic chamber with equal dimensions of 2 m, a step-up transformer (100 kV, 5 kVA), capacitive voltage divider (with a voltage ratio of 10000:1) and a shunt resistor (LC measurement) in series with the insulator. The output file of the digital data acquisition system for all measured variables can be processed by MATLAB software. Samples are made from high temperature vulcanized (HTV) silicone rubber (SiR) with aluminum trihydrate (ATH) filler. Rated voltage of the samples is 20 kV. The five-year-old samples exhibit a hydrophobicity class (HC) of 3 according to IEC62073. This classification indicates that the polymeric surface of the insulators is neither completely hydrophobic (distinctly water-repellent) nor hydrophilic (fully water-accepting), resulting in an intermediate level of hydrophobicity. This is a common state for aged or slightly polluted polymeric insulators in real practices. In addition, this would allow for less deviation or modeling errors in further development of the intelligent model based on test data. Images of the samples and their specification are presented in Fig. 2 and Table 1 respectively. Test procedure is performed according to IEC 60507, IEC 60243-1. Initial value of the source voltage is set to 10 kV; then the voltage is stepped up 0.5 kV per minute for 20 min, while the voltage and LC data are saved by data acquisition system. Pre-contamination process is achieved according to IEC 60815. Pollution is sprayed into the test chamber with the fixed equivalent salt deposit density (ESDD) of the standard solution by distilled water and NaCl. ESDD value is controlled by measurement of solution conduction. Surface conduction of the pollution solution (µS/cm) can be converted into equivalent ESDD (mg/cm2) using surface polluted area of the insulator. Stages of surface discharge on sample A for very heavy pollution level are shown in Fig. 3.Fig. 1Experiment Setup.Fig. 2Photo of the samples.Table 1Specification of the test samples.IDArcing distance (mm)Num. of sheds (Big/Small)Diameter of sheds (mm)Creepage distance (mm)Age (year)HCA3409 (5/4)110/8084053B3206 (4/2)110/8072053C300314056053D3006 (4/2)100/9058053E240513068053F30010 (4/6)75/5566053Fig. 3Stages of surface discharges on sample A for very heavy pollution level. Experiment Setup. Photo of the samples. Table 1Specification of the test samples.IDArcing distance (mm)Num. of sheds (Big/Small)Diameter of sheds (mm)Creepage distance (mm)Age (year)HCA3409 (5/4)110/8084053B3206 (4/2)110/8072053C300314056053D3006 (4/2)100/9058053E240513068053F30010 (4/6)75/5566053 Specification of the test samples. Stages of surface discharges on sample A for very heavy pollution level. Analytical formulation of arcing parameters Calculated parameters In this part, numerical value of arc time constant and arc velocity is presented for the test sample according to LC data of the polluted insulator based on different pollution levels. It is worth to note that calculation process of arc time constant is done through the presented method in Zhang et al.37. In this work, development of surface arc is investigated for insulators and arc time constant of the test sample is calculated based LC data using a novel analytical approach. In order to calculate arc velocity, numerical value of arc time constant and arc length can be adopted using the following (1),1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}=\frac{{l}_{arc}}{{\tau}_{arc}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} are arc velocity, arc length and arc time constant. In this paper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is calculated from the optimization model that is presented in Fahimi et al.38. In this work, a modified Wilkins’ model22 is adopted according to variations of pollution layer conduction and pollution layer thickness for calculation of arc length based on LC data of polluted insulators. Both presented models in Zhang et al.37 and Fahimi et al.38 adopt a dynamic approach where variation of test conditions is included for final output variables. LC data for different pollution levels are presented in Fig. 4. Examples of surface discharges in LC are illustrated in this figure for both positive and negative half cycles as well. It can be seen that maximum discharge values (according to vertical axes of the Fig. 4a–d are increased due to pollution level from light to very heavy. Oscillatory nature of the LC has been discussed in many previous researches. In fact, two physical factors can be responsible for the LC content on the surface of a polluted insulator. First, a wet condition is applied continuously during test period. Therefore, wet pollution layer due to wetting rate and dry bands due to evaporation because of LC heat are both present on the surface of insulator during all test time. In addition, as it was explained in previous section, base input voltage is adjusted to 10 kV and gradually increased by time-step of 0.5 kV/min. Hence, if transient oscillations are observed in trends of LC content39,40, it is due to ceaseless dynamic condition of local discharges because variable domains of wet polluted bands and dry bands that results in oscillatory nature of the LC content. Consequently, this would result in oscillating trends for the calculated parameters of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}.Fig. 4Surface discharges in positive and negative half cycles of the LC for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. Surface discharges in positive and negative half cycles of the LC for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} (colorful dots) and correspondent fitted curves (by MATLAB) for different zones of pollution levels (from light to very heavy levels) based on the experiment results are presented in Figs. 5 and 6, respectively. It can be seen that respective values are calculated for positive half cycles, negative have cycles and complete cycles. As it was discussed above, oscillations in all variable trends that are extracted from LC data can cause ambiguities in further analyses and interpretation. Consideration of the following notes would be helpful to mitigate such obscurities in further analyses:LC content has oscillatory nature in experimental pollution flashover test; because pollution bands and dry bands are dynamically changing due to continuation of wetting condition and heat of discharge activities. However, general trend of LC and discharge numbers are increasing due to voltage step-up. This point was explained above with correspondent references to previous researches.It is tried to obtain smoother trends by average calculation for each 1000 ms time widow (equal to fifty time-cycles in a 50 Hz power system). The largest discharge in each half cycle (for calculation of parameters in negative/positive half cycles) and largest discharge in each full cycle (for calculation of parameters in full cycle) is selected for averaging operation in fifty cycles. This would make further analysis less challenging41.Calculations of the correspondent parameters in Fig. 5 and 6 for full cycle are done independent of similar calculations for negative and positive have cycles. It means that calculated values (colorful dots in Fig. 5 and 6) for full cycle parameters are not average of correspondent negative and positive half cycles.The final fitted curve for each parameter is obtained using the Gaussian model by MATLAB curve fitting tool. The purpose is to cover oscillations in order to better understand the general trend (rate and intensity) for arc time constant and arc velocity. Due to transient nature of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}, Gaussian function42 is utilized to keep the contained information of the test data and cover unnecessary oscillations for long-term study.Fig. 5Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy.Fig. 6Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. LC content has oscillatory nature in experimental pollution flashover test; because pollution bands and dry bands are dynamically changing due to continuation of wetting condition and heat of discharge activities. However, general trend of LC and discharge numbers are increasing due to voltage step-up. This point was explained above with correspondent references to previous researches. LC content has oscillatory nature in experimental pollution flashover test; because pollution bands and dry bands are dynamically changing due to continuation of wetting condition and heat of discharge activities. However, general trend of LC and discharge numbers are increasing due to voltage step-up. This point was explained above with correspondent references to previous researches. It is tried to obtain smoother trends by average calculation for each 1000 ms time widow (equal to fifty time-cycles in a 50 Hz power system). The largest discharge in each half cycle (for calculation of parameters in negative/positive half cycles) and largest discharge in each full cycle (for calculation of parameters in full cycle) is selected for averaging operation in fifty cycles. This would make further analysis less challenging41. It is tried to obtain smoother trends by average calculation for each 1000 ms time widow (equal to fifty time-cycles in a 50 Hz power system). The largest discharge in each half cycle (for calculation of parameters in negative/positive half cycles) and largest discharge in each full cycle (for calculation of parameters in full cycle) is selected for averaging operation in fifty cycles. This would make further analysis less challenging41. Calculations of the correspondent parameters in Fig. 5 and 6 for full cycle are done independent of similar calculations for negative and positive have cycles. It means that calculated values (colorful dots in Fig. 5 and 6) for full cycle parameters are not average of correspondent negative and positive half cycles. Calculations of the correspondent parameters in Fig. 5 and 6 for full cycle are done independent of similar calculations for negative and positive have cycles. It means that calculated values (colorful dots in Fig. 5 and 6) for full cycle parameters are not average of correspondent negative and positive half cycles. The final fitted curve for each parameter is obtained using the Gaussian model by MATLAB curve fitting tool. The purpose is to cover oscillations in order to better understand the general trend (rate and intensity) for arc time constant and arc velocity. Due to transient nature of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}, Gaussian function42 is utilized to keep the contained information of the test data and cover unnecessary oscillations for long-term study. The final fitted curve for each parameter is obtained using the Gaussian model by MATLAB curve fitting tool. The purpose is to cover oscillations in order to better understand the general trend (rate and intensity) for arc time constant and arc velocity. Due to transient nature of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}, Gaussian function42 is utilized to keep the contained information of the test data and cover unnecessary oscillations for long-term study. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. It is worth noting that the choice of the Gaussian model for smoothing the transient and oscillatory nature of LC data is made with the intention of creating a robust AI prediction model in subsequent sections. Capturing all transients and oscillations with similar fitting tools could increase the risk of overfitting phenomena in the final prediction model. In fact, the Gaussian model has been employed first for data pre-processing and then for statistical characterization (feature engineering) that would result in a robust pipeline for further AI training. The core advantages are:By using a Gaussian filter to smooth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} data, the primary risk of overfitting is mitigated. This preprocessing step removes high-frequency transients and oscillations that lead to targeted noise reduction, allowing the AI model to learn from the stable, underlying physical trends that are crucial for reliable prediction.The smoothed outputs serve as higher-quality input features. They reduce the variance and complexity that could impose challenges to the AI model by simplifying the learning problem. Consequently, this would lead to faster training convergence and more stable learning processes. In addition, the performance of the developed final model better generalizes to unseen validation and test data. By using a Gaussian filter to smooth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} data, the primary risk of overfitting is mitigated. This preprocessing step removes high-frequency transients and oscillations that lead to targeted noise reduction, allowing the AI model to learn from the stable, underlying physical trends that are crucial for reliable prediction. By using a Gaussian filter to smooth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} data, the primary risk of overfitting is mitigated. This preprocessing step removes high-frequency transients and oscillations that lead to targeted noise reduction, allowing the AI model to learn from the stable, underlying physical trends that are crucial for reliable prediction. The smoothed outputs serve as higher-quality input features. They reduce the variance and complexity that could impose challenges to the AI model by simplifying the learning problem. Consequently, this would lead to faster training convergence and more stable learning processes. In addition, the performance of the developed final model better generalizes to unseen validation and test data. The smoothed outputs serve as higher-quality input features. They reduce the variance and complexity that could impose challenges to the AI model by simplifying the learning problem. Consequently, this would lead to faster training convergence and more stable learning processes. In addition, the performance of the developed final model better generalizes to unseen validation and test data. Further observation of the presented results in Fig. 5 and 6 reveals the following points:Increasing the pollution level from light to very heavy (correspondent conduction of pollution solution to 10,000 µS/cm, 20,000 µS/cm, 30,000 µS/cm and 40,000 µS/cm) would result is increasing trend for target variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}. While wet pollution condition is established during test time, step-up of input voltage would intensify discharge activities with more energy injection. Correspondent growing LC would expand dry bands both in number and area. This would lead to more powerful arcs with longer durations (because of more energy) and higher velocities.Comparison between the negative/positive half cycle and full cycle values for the same parameter reveals that negative discharges have higher arc time constants and higher velocities. This is in agreement with previous researches43–45. The inception of streamers (localized, high-density electrical discharges) can trigger flashover process in polluted insulators. It is shown that voltage polarity is correlated to the initiation and propagation of streamers that would result in impacting the flashover voltage46. Studies have shown that the negative half-cycle discharges of polluted insulators usually occur in lower voltages than the similar positive ones. This effect is more remarkable for cases of non-uniform pollution47.It can be seen that increasing level of pollution (heavy and very heavy levels) would result in change of arc time constant trends. In other words, values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} would approach to maximum possible values in a power–frequency half-cycle at middle times of test duration. The presented arc time constant values and trends are in accordance with previous researches17,23–25,37. According to (1), this trend would mean higher sharp rates of velocity values for heavy and very heavy pollution levels particularly in final quarter of the test time. This observation is well in agreement with previous researches23,24. Increasing the pollution level from light to very heavy (correspondent conduction of pollution solution to 10,000 µS/cm, 20,000 µS/cm, 30,000 µS/cm and 40,000 µS/cm) would result is increasing trend for target variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}. While wet pollution condition is established during test time, step-up of input voltage would intensify discharge activities with more energy injection. Correspondent growing LC would expand dry bands both in number and area. This would lead to more powerful arcs with longer durations (because of more energy) and higher velocities. Increasing the pollution level from light to very heavy (correspondent conduction of pollution solution to 10,000 µS/cm, 20,000 µS/cm, 30,000 µS/cm and 40,000 µS/cm) would result is increasing trend for target variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}. While wet pollution condition is established during test time, step-up of input voltage would intensify discharge activities with more energy injection. Correspondent growing LC would expand dry bands both in number and area. This would lead to more powerful arcs with longer durations (because of more energy) and higher velocities. Comparison between the negative/positive half cycle and full cycle values for the same parameter reveals that negative discharges have higher arc time constants and higher velocities. This is in agreement with previous researches43–45. The inception of streamers (localized, high-density electrical discharges) can trigger flashover process in polluted insulators. It is shown that voltage polarity is correlated to the initiation and propagation of streamers that would result in impacting the flashover voltage46. Studies have shown that the negative half-cycle discharges of polluted insulators usually occur in lower voltages than the similar positive ones. This effect is more remarkable for cases of non-uniform pollution47. Comparison between the negative/positive half cycle and full cycle values for the same parameter reveals that negative discharges have higher arc time constants and higher velocities. This is in agreement with previous researches43–45. The inception of streamers (localized, high-density electrical discharges) can trigger flashover process in polluted insulators. It is shown that voltage polarity is correlated to the initiation and propagation of streamers that would result in impacting the flashover voltage46. Studies have shown that the negative half-cycle discharges of polluted insulators usually occur in lower voltages than the similar positive ones. This effect is more remarkable for cases of non-uniform pollution47. It can be seen that increasing level of pollution (heavy and very heavy levels) would result in change of arc time constant trends. In other words, values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} would approach to maximum possible values in a power–frequency half-cycle at middle times of test duration. The presented arc time constant values and trends are in accordance with previous researches17,23–25,37. According to (1), this trend would mean higher sharp rates of velocity values for heavy and very heavy pollution levels particularly in final quarter of the test time. This observation is well in agreement with previous researches23,24. It can be seen that increasing level of pollution (heavy and very heavy levels) would result in change of arc time constant trends. In other words, values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} would approach to maximum possible values in a power–frequency half-cycle at middle times of test duration. The presented arc time constant values and trends are in accordance with previous researches17,23–25,37. According to (1), this trend would mean higher sharp rates of velocity values for heavy and very heavy pollution levels particularly in final quarter of the test time. This observation is well in agreement with previous researches23,24. Investigation of variation rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} can be useful for better understanding of arc propagation dynamics. Hence, curves of first and second derivatives for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} according to Gaussian fit for different zones of pollution levels (from light to very heavy levels) are presented in Figs. 7, 8, 9 and 10 as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document}, respectively. Observation of the presented results in Figs. 7, 8, 9 and 10 reveals the following points:There is an extreme point as a convexity in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} trend for all pollution levels in Fig. 7. It can be observed that the value of this extremum point increases from 200 to 600 with increase of pollution levels. In addition, time of this convexity occurrence is shifted to earlier times from 17th minute in light pollution level to 12th minute in very heavy pollution level.Similar investigation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} trend in Fig. 8 for all pollution levels reveals the possibility of occurrence of two extremum points. It can be seen from Fig. 8 that the first extremum point appears as a convexity for all pollution levels. However, the second extremum point appears only in heavy and very heavy pollution levels as a concave.Investigation of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} trend in Fig. 9 implies that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} would not tend to grow for light and medium pollution levels. Further observation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} verifies this point. Analytically, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} was defined through (1) as a ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}. It was seen from Figs. 7 and 8 that there is a limit for growth of arc time constant (theoretically, it is lower than half cycle). Therefore, the answer to reason behind decreasing trend of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} for lower pollution levels at ending times of the test returns to variation of arc length (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document}). In other words, rate of growth of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} has stopped sooner than rate of growth of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is defined as the expected value of equivalent length of arc in a single pollution layer-single arc layer representation for flashover modeling of polluted insulators48. Upper limit of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is the creepage distance of the insulators.Physically, continuation of the wetting rate combined with pollution and growth of the dry-band zones (in number and area) due to LC growth is the key of transfer from stable arcing stage to unstable arcing stage and flashover in polluted insulators. Above observation regarding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} can be justified that while local discharges exist with remarkable duration at the surface of the polluted insulators in light and medium pollution levels, but final transition to unstable arc stage leading to flashover would not happen because the dry band activities does not grow anymore. In other words, LC due to remained polluted bands cannot enhance dry band activities because of lower pollution levels. This can be traced through trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document}.Fig. 7Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy.Fig. 8Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy.Fig. 9Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy.Fig. 10Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. There is an extreme point as a convexity in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} trend for all pollution levels in Fig. 7. It can be observed that the value of this extremum point increases from 200 to 600 with increase of pollution levels. In addition, time of this convexity occurrence is shifted to earlier times from 17th minute in light pollution level to 12th minute in very heavy pollution level. There is an extreme point as a convexity in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} trend for all pollution levels in Fig. 7. It can be observed that the value of this extremum point increases from 200 to 600 with increase of pollution levels. In addition, time of this convexity occurrence is shifted to earlier times from 17th minute in light pollution level to 12th minute in very heavy pollution level. Similar investigation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} trend in Fig. 8 for all pollution levels reveals the possibility of occurrence of two extremum points. It can be seen from Fig. 8 that the first extremum point appears as a convexity for all pollution levels. However, the second extremum point appears only in heavy and very heavy pollution levels as a concave. Similar investigation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} trend in Fig. 8 for all pollution levels reveals the possibility of occurrence of two extremum points. It can be seen from Fig. 8 that the first extremum point appears as a convexity for all pollution levels. However, the second extremum point appears only in heavy and very heavy pollution levels as a concave. Investigation of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} trend in Fig. 9 implies that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} would not tend to grow for light and medium pollution levels. Further observation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} verifies this point. Analytically, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} was defined through (1) as a ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}. It was seen from Figs. 7 and 8 that there is a limit for growth of arc time constant (theoretically, it is lower than half cycle). Therefore, the answer to reason behind decreasing trend of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} for lower pollution levels at ending times of the test returns to variation of arc length (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document}). In other words, rate of growth of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} has stopped sooner than rate of growth of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is defined as the expected value of equivalent length of arc in a single pollution layer-single arc layer representation for flashover modeling of polluted insulators48. Upper limit of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is the creepage distance of the insulators. Investigation of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} trend in Fig. 9 implies that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} would not tend to grow for light and medium pollution levels. Further observation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} verifies this point. Analytically, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} was defined through (1) as a ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}. It was seen from Figs. 7 and 8 that there is a limit for growth of arc time constant (theoretically, it is lower than half cycle). Therefore, the answer to reason behind decreasing trend of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} for lower pollution levels at ending times of the test returns to variation of arc length (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document}). In other words, rate of growth of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} has stopped sooner than rate of growth of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is defined as the expected value of equivalent length of arc in a single pollution layer-single arc layer representation for flashover modeling of polluted insulators48. Upper limit of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document} is the creepage distance of the insulators. Physically, continuation of the wetting rate combined with pollution and growth of the dry-band zones (in number and area) due to LC growth is the key of transfer from stable arcing stage to unstable arcing stage and flashover in polluted insulators. Above observation regarding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} can be justified that while local discharges exist with remarkable duration at the surface of the polluted insulators in light and medium pollution levels, but final transition to unstable arc stage leading to flashover would not happen because the dry band activities does not grow anymore. In other words, LC due to remained polluted bands cannot enhance dry band activities because of lower pollution levels. This can be traced through trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document}. Physically, continuation of the wetting rate combined with pollution and growth of the dry-band zones (in number and area) due to LC growth is the key of transfer from stable arcing stage to unstable arcing stage and flashover in polluted insulators. Above observation regarding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} can be justified that while local discharges exist with remarkable duration at the surface of the polluted insulators in light and medium pollution levels, but final transition to unstable arc stage leading to flashover would not happen because the dry band activities does not grow anymore. In other words, LC due to remained polluted bands cannot enhance dry band activities because of lower pollution levels. This can be traced through trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document}. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. Calculated values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document} in sample A for different pollution levels: (a) light, (b) medium, (c) heavy and (d) very heavy. According to (1) and correspondent above-mentioned explanations, increasing/decreasing trends and the emergence of extreme points in Figs. 5, 6, 7, 8, 9 and 10 originates from dynamic behavior of pollution layer resistance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document}) that has correlation with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{arc}$$\end{document}. In fact, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} is driven by two sequential physical events on the surface of the insulators under continuous wet polluted conditions during high-voltage stresses:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} experiences an initial decrease due to combination of pollution layer with water that forms a conductive solution. Consequently, surface conductivity increases that leads to growth in partial discharge (PD) activities.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} experiences a subsequent increase due to intensive PD activities that causes local heating and formation of dry bands on the surface. In this state, water evaporates from these bands that results in increasing resistance locally and overall. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} experiences an initial decrease due to combination of pollution layer with water that forms a conductive solution. Consequently, surface conductivity increases that leads to growth in partial discharge (PD) activities. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} experiences an initial decrease due to combination of pollution layer with water that forms a conductive solution. Consequently, surface conductivity increases that leads to growth in partial discharge (PD) activities. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} experiences a subsequent increase due to intensive PD activities that causes local heating and formation of dry bands on the surface. In this state, water evaporates from these bands that results in increasing resistance locally and overall. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}$$\end{document} experiences a subsequent increase due to intensive PD activities that causes local heating and formation of dry bands on the surface. In this state, water evaporates from these bands that results in increasing resistance locally and overall. This progression continues until the discharge path resistance surpasses that of the pollution layer. This shift in the dominant resistance is identified as a critical pre-flashover condition in polluted insulators. Throughout the test, the surface state is determined by the competition between the continuous wetting from the test environment and evaporation driven by the LC from the applied high voltage. In other words, PD stage to flashover stage can be understood as a transition between energy injection from the pollution layer and energy dissipation from the arc. In light pollution, the energy input is low, and arcs remain localized and stable. As pollution level increases, the enhanced LC widens the dry bands, reducing the effective arcing distance and allowing the arc to bridge larger gaps. Once the arc covers a critical portion of the creepage path, the field distortion becomes severe enough to trigger unstable, accelerating arc propagation, leading to flashover. Intense overstress caused by pollution levels and wetting rate sets off an irreversible chain of events where increasing LC and dry-band expansion ultimately result in complete discharge (flashover). The critical turning point where the evaporation of surface water continues until flashover can be identified by extreme points in Figs. 5, 6, 7, 8, 9 and 10. In other words, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} and their derivatives are assisting performance indexes in this research, reflecting dynamic behavior of pollution layer and discharge layer on wet polluted surface of the insulators:The convex peak in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} (Fig. 7) corresponds to the transition from a thermally limited arc to one that is dominated by field-driven propagation. As pollution increases, this peak shifts to earlier times because the higher surface conductivity accelerates the energy accumulation process.Specific convex-concave transition in heavy pollution in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} (Fig. 8) indicates the point at which dry band formation becomes rapid and widespread, leading to unstable arc elongation.The decline in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} for light and medium pollution levels (Fig. 9) indicates that arc elongation stops before the arc duration reaches saturation. This occurs because the LC is not enough to sustain dry band expansion across the entire creepage distance, leaving residual wet bands that hinder further arc propagation. The convex peak in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} (Fig. 7) corresponds to the transition from a thermally limited arc to one that is dominated by field-driven propagation. As pollution increases, this peak shifts to earlier times because the higher surface conductivity accelerates the energy accumulation process. The convex peak in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document} (Fig. 7) corresponds to the transition from a thermally limited arc to one that is dominated by field-driven propagation. As pollution increases, this peak shifts to earlier times because the higher surface conductivity accelerates the energy accumulation process. Specific convex-concave transition in heavy pollution in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} (Fig. 8) indicates the point at which dry band formation becomes rapid and widespread, leading to unstable arc elongation. Specific convex-concave transition in heavy pollution in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document} (Fig. 8) indicates the point at which dry band formation becomes rapid and widespread, leading to unstable arc elongation. The decline in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} for light and medium pollution levels (Fig. 9) indicates that arc elongation stops before the arc duration reaches saturation. This occurs because the LC is not enough to sustain dry band expansion across the entire creepage distance, leaving residual wet bands that hinder further arc propagation. The decline in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document} for light and medium pollution levels (Fig. 9) indicates that arc elongation stops before the arc duration reaches saturation. This occurs because the LC is not enough to sustain dry band expansion across the entire creepage distance, leaving residual wet bands that hinder further arc propagation. Above interpretation regarding observation of the Figs. 5, 6, 7, 8, 9 and 10 can be used to distinguish arcing stage transitions in a polluted insulator. This point is shown in Table 2. As it can be seen from the table, tracing the extremum points in correspondent trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} can be used to develop a condition-assessment scheme to judge the discharge activity level in surface of the polluted insulators.Table 2Arcing mode transition according to extremum points in correspondent trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document}Transition mode––ConvexConvexPartial dischargesConvexConvex––Partial discharges to stable arcConvexConvex–concave––Stable arc to unstable arc Table 2Arcing mode transition according to extremum points in correspondent trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{\tau}_{arc}}{dt}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{\tau}_{arc}}{d{t}^{2}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d{v}_{arc}}{dt}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{d}^{2}{v}_{arc}}{d{t}^{2}}$$\end{document}Transition mode––ConvexConvexPartial dischargesConvexConvex––Partial discharges to stable arcConvexConvex–concave––Stable arc to unstable arc Arcing mode transition according to extremum points in correspondent trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}. In order to investigate robustness of the presented analysis according to the Gaussian smoothing method, a comparison is made against raw unsmoothed data and five other smoothing techniques including moving average, Savitzky-Golay, median filter, exponential moving average, and wavelet denoising. The results are summarized in Table 3. It is observed that the Gaussian filter with a standard deviation of two samples presents the lowest mean absolute error (MAE) of 0.30 plus or minus 0.02 min and the highest F1 score of 87.3% across all pollution levels. This means a 40% improvement compared to non-smoothing approach. A sensitivity analysis is conducted as well regarding this point. The results in Table 4 demonstrates that the optimal settings with a standard deviation of two samples would provide an effective kernel width. This means the filter cuts down the 50 Hz oscillations while leaving the slower arc dynamics intact. Those arc dynamics have time constants above 100 ms, so they remain unaffected. The improvement holds steady between 39.2 and 42.3% across pollution levels from 10,000 to 40,000 µS/cm, which indicates that the benefit does not depend on the pollution severity. Beyond accuracy, smoothing also cut false negatives from 27.1 down to 15.0% and false alarms from 19.0 down to 10.1% (flashover events). The biggest gain showed up in oscillatory pre-flashover situations, where the signal tends to get noisy. Five-fold cross-validation returned p values under 0.001 for every fold that confirms the improvement is statistically significant. Overall, these results show that Gaussian smoothing effectively cleans up the high-frequency LC oscillations while preserving the meaningful physical trends in arc time constant and arc velocity. In addition, no systematic bias is introduced and the detection of critical transitions would not be delayed.Table 3Performance of smoothing techniques for heavy pollution level (30,000 µS/cm).Smoothing techniqueAccuracy (%)F1 score (%)MAE (min)Training epochs to convergeInference time (ms)No smoothing81.378.90.52 ± 0.071423.89Moving average (10-point)85.783.40.38 ± 0.041183.92Savitzky-Golay (order 3, frame 11)86.284.10.34 ± 0.031044.18Median filter (k = 5)82.980.30.41 ± 0.051313.95Exponential moving average (α = 0.1)84.581.90.37 ± 0.041253.91Wavelet denoising (db4, level 4)87.485.60.32 ± 0.03965.23Gaussian (proposed, σ = 2)89.187.30.30 ± 0.02873.95Table 4Sensitivity of Gaussian smoothing parameter (σ) for heavy pollution level.σ (samples)Effective Kernel width (ms)MAE (min)F1 score (%)Physical interpretation0.530.4879.2Under-smoothed; retains 50 Hz oscillations1.060.3883.7Partial smoothing2.0120.3087.3Optimal; removes LC oscillations, preserves arc dynamics3.0180.3385.1Over-smoothed; delays transition detection5.0300.4181.9Excessive smoothing; loses pre-flashover acceleration Table 3Performance of smoothing techniques for heavy pollution level (30,000 µS/cm).Smoothing techniqueAccuracy (%)F1 score (%)MAE (min)Training epochs to convergeInference time (ms)No smoothing81.378.90.52 ± 0.071423.89Moving average (10-point)85.783.40.38 ± 0.041183.92Savitzky-Golay (order 3, frame 11)86.284.10.34 ± 0.031044.18Median filter (k = 5)82.980.30.41 ± 0.051313.95Exponential moving average (α = 0.1)84.581.90.37 ± 0.041253.91Wavelet denoising (db4, level 4)87.485.60.32 ± 0.03965.23Gaussian (proposed, σ = 2)89.187.30.30 ± 0.02873.95 Performance of smoothing techniques for heavy pollution level (30,000 µS/cm). Table 4Sensitivity of Gaussian smoothing parameter (σ) for heavy pollution level.σ (samples)Effective Kernel width (ms)MAE (min)F1 score (%)Physical interpretation0.530.4879.2Under-smoothed; retains 50 Hz oscillations1.060.3883.7Partial smoothing2.0120.3087.3Optimal; removes LC oscillations, preserves arc dynamics3.0180.3385.1Over-smoothed; delays transition detection5.0300.4181.9Excessive smoothing; loses pre-flashover acceleration Sensitivity of Gaussian smoothing parameter (σ) for heavy pollution level. It is noteworthy that inclusion of other experimental design choices such as pollution levels, pollution materials, variable wetting rates and durations, or different insulator materials would enrich the available dataset for the further development of the physics-induced AI training. Anyhow, the presented experiment works were performed with the purpose of recording all stages of the surface discharges in polluted insulators in specific patterns through LC trends and the introduced arc constants. In other words, all possible LC patterns and arc constant patterns are recorded. Therefore, variation of the experiment parameters would be reflected as one of the elaborated LC patterns, arc constant patterns and transition in discharge stages that are considered in the proposed model. It would be shown that the proposed physics-induced AI approach is adaptive in case of field input data. This is due to normalized footprint of dynamic behavior of arc constants in any different cases. In other words, variation of experimental conditions, sample selections or any other uncertainties in real outdoor implications would be reflected in one of the investigated arc constant trends. Therefore, the presented analysis is independent of the input data and would achieve to trace critical pre-flashover thresholds in any cases. This means that changing experiment or filed conditions would not make challenge in in model training or the ability of the developed model to predict transition stages of the surface discharge activities in polluted insulators. Proposed arc model It was explained that development of the predictive algorithm for pollution flashover of the outdoor insulators would be achieved based on trends of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}​ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}. This procedure can be summarized as follows according to the Table 2:Extraction of extremum points: First and second derivatives of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}​ (Figs. 7, 8, 9 and 10) is used to identify critical points (convexity/concavity) that indicate transitions between discharge stages (Table 2).Defining thresholds: Thresholds for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}​ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} values can be established that correlate with flashover risk (e.g., very heavy pollution levels show sharp increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}​ near flashover).Discharge classification: Table 2 is operationalized to classify discharge modes (partial discharges, stable arc, and unstable arc) in real-time monitoring systems. Extraction of extremum points: First and second derivatives of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}​ (Figs. 7, 8, 9 and 10) is used to identify critical points (convexity/concavity) that indicate transitions between discharge stages (Table 2). Extraction of extremum points: First and second derivatives of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}​ (Figs. 7, 8, 9 and 10) is used to identify critical points (convexity/concavity) that indicate transitions between discharge stages (Table 2). Defining thresholds: Thresholds for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}​ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} values can be established that correlate with flashover risk (e.g., very heavy pollution levels show sharp increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}​ near flashover). Defining thresholds: Thresholds for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}​ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} values can be established that correlate with flashover risk (e.g., very heavy pollution levels show sharp increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}​ near flashover). Discharge classification: Table 2 is operationalized to classify discharge modes (partial discharges, stable arc, and unstable arc) in real-time monitoring systems. Discharge classification: Table 2 is operationalized to classify discharge modes (partial discharges, stable arc, and unstable arc) in real-time monitoring systems. In order to achieve high accuracy and real-time applicability, a hybrid machine learning (ML) model is proposed in this paper using combination of physics-informed feature engineering with deep learning. Step-by-step methodology of this procedure including model selection, training, and validation would be presented in the following subsections. Hybrid CNN–LSTM with attention In this paper, proposed ML model is a hybrid CNN–LSTM with attention. A convolutional neural network (CNN) is a type of deep learning algorithm primarily used for analyzing, extraction and hierarchical learning of features from input data49. CNN can be used to extracts local patterns in LC waveforms32,33,50. Long short-term memory (LSTM) is a type of recurrent neural network (RNN) designed to handle sequential data such as time series with beneficial feature of capturing long-term dependencies within sequences that overcomes traditional RNNs. LSTMs remember past information and discard irrelevant details by selection51. This particular feature can help to capture temporal dependencies in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}​ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} trends. In deep learning, the attention mechanism is a technique that enables a model to focus on different parts of the input data when making predictions. It helps understanding most relevant parts of the input for the final objective52. It can be used to focus on critical time steps such as pre-flashover instability. Specification of the applied ML model is represented in Table 5.Table 5Architecture specification of the proposed hybrid CNN-LSTM with attention.Layer typeOutput shapeParametersInput layer(None, T, 8)01D Conv + MaxPool (Filters = 64, Kernel = 3)(None, T/2, 64)1664Bidirectional LSTM (Units = 128)(None, T/2, 256)263,168Attention layer (Self-Attention)(None, T/2, 256)65,792Flatten + Dense (ReLU)(None, 64)16,448Output layer (Sigmoid)(None, 1)65 Table 5Architecture specification of the proposed hybrid CNN-LSTM with attention.Layer typeOutput shapeParametersInput layer(None, T, 8)01D Conv + MaxPool (Filters = 64, Kernel = 3)(None, T/2, 64)1664Bidirectional LSTM (Units = 128)(None, T/2, 256)263,168Attention layer (Self-Attention)(None, T/2, 256)65,792Flatten + Dense (ReLU)(None, 64)16,448Output layer (Sigmoid)(None, 1)65 Architecture specification of the proposed hybrid CNN-LSTM with attention. Dataset is split for further process: 70% of the data is used to train the model, 15% of the data is used to tune the hyperparameters (model validation) and 15% of the data is considered for final evaluation (model test). Since the final goal is to predict flashover, loss function is defined as binary cross-entropy: 0 = No Flashover and 1 = Flashover. Adam optimizer is enabled with learning rate equal to 0.001. Batch size is 32 for balance between running speed and model stability. Results and discussion Presented results in Table 6 provides a comprehensive analysis of insulator performance under heavy pollution level (30,000 µS/cm). In can be seen that the proposed model demonstrates a high degree of accuracy in predicting transition times between different arc phases according to consistent mean absolute error (MAE) of 0.30 ± 0.02 min for all samples. This error margin is due to an intentional bias that is incorporated into the model to ensure fail-safe operation. This bias originates from the dynamic arc model where the arc time constant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}) is intentionally overestimated by 10% to cover measurement uncertainties in real practices as specified by IEC 60,815 standards. It is observed that final predictions of the proposed model are lower than actual observed values. This is a safety margin that prevents underestimation of flashover risks. Further observation of Table 6 reveals that only 50% of the tested samples (A, C and E) have progressed to complete discharge (flashover) while the remaining samples (B, D and F) were remained at the stable arc phase. It can be seen that the presented model is able to identify these cases by monitoring the second derivative of the arc time constant. This point ensures that only sustained arc acceleration would lead to flashover prediction. In addition, the normal distribution of errors (Shapiro–Wilk p = 0.42) validates reliability of the model. It confirms that parametric statistical methods can be utilized for data analysis. Furthermore, the lack of correlation between prediction errors and geometric factors such as arcing distance (ρ = − 0.08, p = 0.72) and shed configuration (ρ = 0.05, p = 0.82) highlights robustness of the presented model. Independency to profile factors is due to the use of a normalized arc velocity parameter (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}/{l}_{arc}$$\end{document}) that effectively separates the arc progression dynamics from the geometrical dimensions of the insulator. The results approve the influential role of pollution severity over insulator geometry in flashover behavior which provides valuable insights for maintenance and design optimization.Table 6Transition performance for heavy pollution level (30,000 µS/cm).SampleArcing distShed configPartial dischargesPartial → Stable arcStable → Unstable arcUnstable → Flashover(mm)P/Act/Err/σP/Act/Err/σP/Act/Err/σP/Act/Err/σA3409 (5/4)2.1/2.4/0.3/0.055.2/5.5/0.3/0.0410.5/10.8/0.3/0.0315.1/15.4/0.3/0.02B3206 (4/2)2.0/2.3/0.3/0.065.1/5.4/0.3/0.0510.4/10.7/0.3/0.04–C30031.9/2.2/0.3/0.044.9/5.2/0.3/0.039.8/10.1/0.3/0.0214.7/15.0/0.3/0.02D3006 (4/2)2.0/2.3/0.3/0.055.0/5.3/0.3/0.0410.0/10.3/0.3/0.03–E24051.8/2.1/0.3/0.034.7/5.0/0.3/0.029.5/9.8/0.3/0.0214.2/14.5/0.3/0.01F30010 (4/6)2.1/2.4/0.3/0.065.2/5.5/0.3/0.0510.3/10.6/0.3/0.04– Table 6Transition performance for heavy pollution level (30,000 µS/cm).SampleArcing distShed configPartial dischargesPartial → Stable arcStable → Unstable arcUnstable → Flashover(mm)P/Act/Err/σP/Act/Err/σP/Act/Err/σP/Act/Err/σA3409 (5/4)2.1/2.4/0.3/0.055.2/5.5/0.3/0.0410.5/10.8/0.3/0.0315.1/15.4/0.3/0.02B3206 (4/2)2.0/2.3/0.3/0.065.1/5.4/0.3/0.0510.4/10.7/0.3/0.04–C30031.9/2.2/0.3/0.044.9/5.2/0.3/0.039.8/10.1/0.3/0.0214.7/15.0/0.3/0.02D3006 (4/2)2.0/2.3/0.3/0.055.0/5.3/0.3/0.0410.0/10.3/0.3/0.03–E24051.8/2.1/0.3/0.034.7/5.0/0.3/0.029.5/9.8/0.3/0.0214.2/14.5/0.3/0.01F30010 (4/6)2.1/2.4/0.3/0.065.2/5.5/0.3/0.0510.3/10.6/0.3/0.04– Transition performance for heavy pollution level (30,000 µS/cm). Similarly, transition performance of the samples under very heavy pollution level (40000 µS/cm) is presented in Table 7. Compared to the heavy pollution scenario, significant differences can be observed. Unlike the 50% of flashover rate that was observed in Table 4, all samples have progressed to flashover in this case that illustrates the key role of pollution severity in arc development. The model maintains its high predictive accuracy with a MAE equal to 0.30 ± 0.01 min. It is consistent with the systematic bias that was observed in heavy pollution level. This consistency confirms that the model errors are primarily due to calibration margins rather than environmental or geometric variations. The accelerated transition times under very heavy pollution level is particularly remarkable. For instance, flashover time for sample A is decreased from 15.4 min in heavy pollution level to 12.4 min in very heavy pollution level that reflects the increased surface conductivity at higher pollution levels. This behavior aligns with the governing exponential scaling law for flashover dynamics where the time-to-flashover is inversely proportional to the pollution layer conductivity. Presented model catches this correlation which accounts for the dynamic changes in arc conduction as a function of applied voltage and power losses. Further observation of the normal error distribution (Shapiro–Wilk p = 0.51) and minimal standard deviations (≤ 0.04 min) approve the model precision and reliability as well. The 100% transition rate for flashover under very heavy pollution level verifies the capability of the proposed model to distinguish between critical conditions where arc progression is energy-limited (heavy pollution level) and those where it is conductivity-driven (very heavy pollution level). This distinction is crucial for operational decision-making frameworks as it allows power utilities to schedule maintenance and protection measures based on real-time pollution monitoring. The results also highlight the effectiveness of the model for optimization of insulator design as compact configurations (samples C and E) that demonstrate comparable performance to larger units.Table 7Transition performance for very heavy pollution level (40,000 µS/cm).SampleArcing distShed configPartial dischargesPartial → Stable ArcStable → Unstable ArcUnstable → Flashover(mm)P/Act/Err/σP/Act/Err/σP/Act/Err/σP/Act/Err/σA3409 (5/4)1.7/2.0/0.3/0.044.2/4.5/0.3/0.038.5/8.8/0.3/0.0212.1/12.4/0.3/0.02B3206 (4/2)1.6/1.9/0.3/0.034.0/4.3/0.3/0.028.3/8.6/0.3/0.0211.9/12.2/0.3/0.01C30031.5/1.8/0.3/0.023.8/4.1/0.3/0.028.0/8.3/0.3/0.0111.7/12.0/0.3/0.01D3006 (4/2)1.6/1.9/0.3/0.034.0/4.3/0.3/0.028.2/8.5/0.3/0.0211.8/12.1/0.3/0.01E24051.4/1.7/0.3/0.023.6/3.9/0.3/0.017.8/8.1/0.3/0.0111.5/11.8/0.3/0.01F30010 (4/6)1.7/2.0/0.3/0.044.1/4.4/0.3/0.038.4/8.7/0.3/0.0212.0/12.3/0.3/0.02 Table 7Transition performance for very heavy pollution level (40,000 µS/cm).SampleArcing distShed configPartial dischargesPartial → Stable ArcStable → Unstable ArcUnstable → Flashover(mm)P/Act/Err/σP/Act/Err/σP/Act/Err/σP/Act/Err/σA3409 (5/4)1.7/2.0/0.3/0.044.2/4.5/0.3/0.038.5/8.8/0.3/0.0212.1/12.4/0.3/0.02B3206 (4/2)1.6/1.9/0.3/0.034.0/4.3/0.3/0.028.3/8.6/0.3/0.0211.9/12.2/0.3/0.01C30031.5/1.8/0.3/0.023.8/4.1/0.3/0.028.0/8.3/0.3/0.0111.7/12.0/0.3/0.01D3006 (4/2)1.6/1.9/0.3/0.034.0/4.3/0.3/0.028.2/8.5/0.3/0.0211.8/12.1/0.3/0.01E24051.4/1.7/0.3/0.023.6/3.9/0.3/0.017.8/8.1/0.3/0.0111.5/11.8/0.3/0.01F30010 (4/6)1.7/2.0/0.3/0.044.1/4.4/0.3/0.038.4/8.7/0.3/0.0212.0/12.3/0.3/0.02 Transition performance for very heavy pollution level (40,000 µS/cm). A sensitivity analysis is conducted to quantify the effect of Gaussian smoothing on prediction performance. As shown in Table 8, the relative smoothing effect remains stable between 39.2 and 42.3% across all four pollution levels from 10,000 to 40,000 µS/cm, despite the fact that the peak-to-peak noise amplitude increased more than fivefold from 0.8 to 4.2 mA. This stability confirms that the Gaussian smoothing parameter of two samples is robustly tuned to the signal-to-noise ratio characteristics of LC measurements rather than to absolute noise levels. The smoothing effect is also examined across the four arc transition stages previously defined in Table 2. The stable-to-unstable arc transition, which is the most critical operational decision point, exhibits the largest relative improvement at 42.3%. In comparison, the partial discharges and partial to stable arc transitions shows improvements of approximately 39%. The Smoothing Sensitivity Index which is defined as the absolute difference between smoothed and raw predictions divided by the standard deviation of raw bootstrap predictions, remains below one for all transitions, with values of 0.67 for the stable to unstable transition and 0.73 for the unstable to flashover transition. This point indicates that smoothing would reduce prediction variance without systematic shifting of the mean prediction beyond natural data variability. According to Table 9, when flashover events are categorized into three temporal patterns based on LC characteristics, the benefit of smoothing is the best for Type II oscillatory pre-flashover scenarios, which accounted for 35% of flashover events. For these challenging cases, smoothing improved the F1 score from 71.5 to 86.7%, while reducing the false negative rate from 65.3 to 16.3%. For Type I (gradual scenarios) and Type III (abrupt scenarios), smoothing provides modest improvements of 2.2% and 2.1%, respectively, confirming that smoothing does not degrade performance for any scenario type. Five-fold cross-validation confirms that the observed improvements are statistically significant with p-values less than 0.001 across all folds. These results establish that Gaussian smoothing is particularly valuable for the most challenging field conditions where oscillatory LC patterns would otherwise obscure the underlying arc propagation dynamics.Table 8Relative smoothing effect across pollution levels and transition phases.ConditionRaw MAE (min)Gaussian MAE (min)Relative Smoothing Effect (%)Noise Amplitude (mA, p-p)Pollution levelLight (10,000 µS/cm)0.480.2939.60.8Medium (20,000 µS/cm)0.510.3139.21.5Heavy (30,000 µS/cm)0.520.3042.32.8Very heavy (40,000 µS/cm)0.490.2940.84.2Transition phase Partial discharges0.480.2939.6– Partial → Stable Arc0.510.3139.2– Stable → Unstable Arc0.520.3042.3– Unstable → Flashover0.490.2940.8–Table 9Smoothing performance by flashover scenario type.Scenario typeFrequencyRaw F1 (%)Smooth F1 (%)Improvement (%)Raw false negative rate (%)Smooth false negative rate (%)Type I (Gradual acceleration)45%86.288.42.218.311.2Type II (Oscillatory pre-flashover)35%71.586.715.265.316.3Type III (Abrupt flashover)20%83.185.22.122.515.8 Table 8Relative smoothing effect across pollution levels and transition phases.ConditionRaw MAE (min)Gaussian MAE (min)Relative Smoothing Effect (%)Noise Amplitude (mA, p-p)Pollution levelLight (10,000 µS/cm)0.480.2939.60.8Medium (20,000 µS/cm)0.510.3139.21.5Heavy (30,000 µS/cm)0.520.3042.32.8Very heavy (40,000 µS/cm)0.490.2940.84.2Transition phase Partial discharges0.480.2939.6– Partial → Stable Arc0.510.3139.2– Stable → Unstable Arc0.520.3042.3– Unstable → Flashover0.490.2940.8– Relative smoothing effect across pollution levels and transition phases. Table 9Smoothing performance by flashover scenario type.Scenario typeFrequencyRaw F1 (%)Smooth F1 (%)Improvement (%)Raw false negative rate (%)Smooth false negative rate (%)Type I (Gradual acceleration)45%86.288.42.218.311.2Type II (Oscillatory pre-flashover)35%71.586.715.265.316.3Type III (Abrupt flashover)20%83.185.22.122.515.8 Smoothing performance by flashover scenario type. Analysis of model performance across respective pollution levels reveals distinct signatures in arc parameters that correlate with flashover results. As summarized in Table 10, the proposed model maintains consistent accuracy across all pollution levels, with MAEs ranging from 0.28 to 0.31 min. A critical arc velocity threshold of approximately 30 cm/s is identified. Below this threshold, the flashover occurrence probability is 17%, while the probability increases to 92% above it. The peak value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{\mathrm{d}{\uptau}_{\mathrm{a}\mathrm{r}\mathrm{c}}}{\mathrm{d}\mathrm{t}}$$\end{document} is able to distinguish non-flashover from flashover samples, with values of 180–220 ms/min for non-flashover heavy pollution, 280–350 ms/min for flashover heavy pollution, and 450–600 ms/min for very heavy pollution. The time from the peak of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{\mathrm{d}{\uptau}_{\mathrm{a}\mathrm{r}\mathrm{c}}}{\mathrm{d}\mathrm{t}}$$\end{document} to flashover is consistently 3.5 ± 0.3 min for heavy pollution and 2.8 ± 0.2 min for very heavy pollution, providing a reliable early warning window for preventive actions.”Table 10Granular model performance across pollution levels.Pollution levelFlashover occurredMAE (min)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} peak (ms)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} peak (cm/s)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{\mathrm{d}{\uptau}_{\mathrm{a}\mathrm{r}\mathrm{c}}}{\mathrm{d}\mathrm{t}}$$\end{document} peak (ms/min)True positive rateSpecificityLightNo0.291.2880–120N/A92.4%MediumNo0.312.515120–150N/A89.7%HeavyYes (3/6)0.304.8 (F) / 3.2 (NF)38 (F) / 22 (NF)280–350 (F) / 180–220 (NF)88.9%85.7%Very heavyYes (6/6)0.296.282450–60092.1%N/A Table 10Granular model performance across pollution levels.Pollution levelFlashover occurredMAE (min)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document} peak (ms)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document} peak (cm/s)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{\mathrm{d}{\uptau}_{\mathrm{a}\mathrm{r}\mathrm{c}}}{\mathrm{d}\mathrm{t}}$$\end{document} peak (ms/min)True positive rateSpecificityLightNo0.291.2880–120N/A92.4%MediumNo0.312.515120–150N/A89.7%HeavyYes (3/6)0.304.8 (F) / 3.2 (NF)38 (F) / 22 (NF)280–350 (F) / 180–220 (NF)88.9%85.7%Very heavyYes (6/6)0.296.282450–60092.1%N/A Granular model performance across pollution levels. Detailed prediction errors are presented for different transition phases under heavy and very heavy pollution levels in Table 11. The results exhibit no significant variation in MAE between these pollution levels (t-test p values > 0.87). This suggests that the model error is not dependent to pollution level. This is a direct consequence of the intentional bias into the model which makes consistent overestimation of transition times regardless of environmental conditions. In fact, the role of time constant overestimation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc\_model} = 1.1 \times {\tau}_{arc\_actual}$$\end{document}) is a calibration factor which compensates for uncertainties in arc dynamics rather than consideration of random variations. The 0.3 min error for all transition phases including partial discharges, partial to stable arc, stable to unstable arc and unstable to flashover approves the model effectiveness in trace of arc development. This consistency would be crucial for operational applications because it ensures reliable predictions throughout all discharge stages. In addition, non-existing correlation between model error and pollution levels emphasizes that the accuracy of the model is limited by its parametric design architecture rather than external variables. Hence, the presented model can be extended for wider range of pollution conditions and possibly other unseen environmental factors that makes it a competing candidate for condition-assessment schemes.Table 11Error analysis for all transition phases.PhaseHeavy Pollution MAE (min)Very Heavy MAE (min)t-test (p-value)Error reduction (%)Partial Discharges0.30 ± 0.020.30 ± 0.010.870Partial → Stable Arc0.30 ± 0.020.30 ± 0.010.920Stable → Unstable Arc0.30 ± 0.010.30 ± 0.010.950Unstable → Flashover0.30 ± 0.010.30 ± 0.010.890 Table 11Error analysis for all transition phases.PhaseHeavy Pollution MAE (min)Very Heavy MAE (min)t-test (p-value)Error reduction (%)Partial Discharges0.30 ± 0.020.30 ± 0.010.870Partial → Stable Arc0.30 ± 0.020.30 ± 0.010.920Stable → Unstable Arc0.30 ± 0.010.30 ± 0.010.950Unstable → Flashover0.30 ± 0.010.30 ± 0.010.890 Error analysis for all transition phases. In addition, statistical evaluation of error distributions (Shapiro-Wilk p > 0.4) verifies that parametric methods can be employed for performance analysis of the presented model. It is worth noting that the output results of predictions can be used in probabilistic risk assessment frameworks and decision-making schemes. The results also illustrate the ability of the model for further refinement, recalibration and tuning as the intentional bias setting could be utilized according to other operational data. Therefore, higher accuracy could be achieved without compromising safety margins. Comparative analysis of key performance metrics is presented in Table 12 for heavy (30,000 µS/cm) and very heavy (40,000 µS/cm) pollution levels. A 19.9% reduction can be seen in average flashover time from 15.1 min in heavy pollution level to 12.1 min in very heavy pollution level. This point is due to the exponential relationship between pollution level and arc propagation velocity. This is described through the scaling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{flashover}\propto {\sigma}_{pollution}^{-n}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{pollution}$$\end{document} is the surface conductivity and n≈ 1.2. It can be seen that the presented model is able to reflect this relationship that allows for precise predictions of flashover risks under varying pollution levels. Although flashover times are different, MAE for prediction remains unchanged at 0.30 min. This is a validation for the true calibration of the proposed model. It is worth noting that fail-safe operation is preferred over absolute accuracy in the proposed approach. Furthermore, the transition metric represents a remarkable improvement from 50% under heavy pollution level to 100% in very heavy pollution level. This improvement illustrates the capability of the presented approach to distinguish between energy-limited and conductivity-driven arc propagation.Table 12Comparison of performance metrics for heavy and very heavy pollution levels.MetricHeavy (30 k µS/cm)Very Heavy (40 k µS/cm)ImprovementAvg. Flashover Time15.1 min12.1 min19.9%Prediction MAE0.30 min0.30 min0%Transition Completeness50%100% + 50% Table 12Comparison of performance metrics for heavy and very heavy pollution levels.MetricHeavy (30 k µS/cm)Very Heavy (40 k µS/cm)ImprovementAvg. Flashover Time15.1 min12.1 min19.9%Prediction MAE0.30 min0.30 min0%Transition Completeness50%100% + 50% Comparison of performance metrics for heavy and very heavy pollution levels. Further analysis of variance (ANOVA) results (F(1,34) = 156.2, p < 0.001) confirms that pollution level accounts for 89% of the variance in transition times while the effect of profile factors is less than 3%. This point shows that the pollution severity can be considered as the key stimulant for flashover dynamics. It is worth to note that the practical implications of these results are profound since it can lead to development of preventive maintenance (PM) frameworks and protection strategies based on real-time pollution monitoring. For example, insulators can be cleaned through scheduled washing programs when predicted transition times reach to critical pre-defined thresholds (for example less than 10 min for heavy pollution level or less than 8 min for very heavy pollution level). This would ensure optimal performance and minimize flashover risks. The performance comparison plot in Fig. 11a demonstrates that the CNN-LSTM model offers superior accuracy (89.1%) and F1 score (87.3%) compared to some similar benchmark approaches. It can be seen that the statistical (logistic regression) model performs the worst (78.2% accuracy) that shows its weakness to follow temporal patterns in LC data. While the Transformer model shows competitive accuracy (86.9%), but its computational inefficiency is a disadvantage according to Fig. 11b where its inference time (13.1 ms) is 3.3 times slower than the CNN-LSTM (3.95 ms). The artificial neural network (ANN) provides the fastest predictions (0.52 ms) but suffers from 7% lower accuracy than the CNN-LSTM. This point illustrates the trade-off between speed and performance in simpler architectures. The gated recurrent unit (GRU) performs intermediately but is still outperformed by the CNN-LSTM by 3.4% in F1 score. Again, this point emphasizes the benefits of combining convolutional and recurrent layers for insulator flashover prediction.Fig. 11Comparison for different model performance: (a) prediction performance and (b) computational efficiency. Comparison for different model performance: (a) prediction performance and (b) computational efficiency. It is worth noting that to present a reliable interpretability and reproducibility comparison between the proposed CNN-LSTM model and benchmark approaches including Transformer, ANN, GRU, and logistic regression, all models are trained and evaluated using the same dataset. The dataset is divided identically for each model into training, validation, and test subsets corresponding to 70%, 15%, and 15% of the total data, respectively. Each model is trained for one hundred epochs with an early stopping patience of ten epochs to prevent overfitting phenomena. The Adam optimizer is employed for all models with a fixed learning rate of 0.001. A batch size of thirty-two is maintained as well. All models receive the same eight-dimensional input features that are derived from the LC waveforms, which include the arc time constant and arc velocity trends. In addition, validation metrics are computed according to the same test data. A binary classification threshold of 0.5 is applied for all comparison models. Computational efficiency is assessed through the average inference time per sample. It is measured on the same hardware system with specification an Intel Core i7 7th generation CPU and 32 GB DDR4 RAM. The Transformer model is implemented with four attention heads, two encoder layers, and a feed-forward network dimension of 256. The developed ANN consists of three layers with 128, 64, and 32 units, respectively and a regularization value of 0.2. The GRU model is built up with two GRU layers with 128 units that are followed by a dense layer of 64 units. The logistic regression model uses L2 regularization with the regularization strength parameter C that is set to 1.0. These settings would ensure that the presented performance comparison in Figure 11 is fair, transparent and reproducible. The receiver operating characteristic (ROC) curves in Fig. 12 reveal that the CNN-LSTM presents the highest area under curve (AUC). It is worth to note that AUC of the ROC is a common metric to evaluate the overall performance of a numerical model. It can be seen from Fig. 12 that AUC of the presented model is 0.924 which indicates superior distinction between flashover and non-flashover events for all threshold settings. Further observation of Fig. 12 reveals that the statistical model (AUC=0.812) represents the poorest performance as its correspondent curve approaches the diagonal line of randomness for high false-positive rates (FPRs). The Transformer (AUC = 0.892) and GRU (AUC = 0.881) exhibit similar performance but tend to be fined earlier than the CNN-LSTM which shows less sensitivity to sensitive arc propagation patterns. In addition, correspondent curve of CNN-LSTM maintains a sharp climb in the low false-positive region (FPR < 0.2) which proves its reliability in safety-critical predictions where a potential failure could lead to malfunction or severe damage in power system applications.Fig. 12ROC curve analysis for the presented model and similar approaches. ROC curve analysis for the presented model and similar approaches. Finally, the confusion matrix is shown in Fig. 13 for the presented model. This table evaluates the balanced performance of the CNN-LSTM with 142 true negatives (86.5%) and 119 true positives (85.0%) according to the input dataset. It can be seen that the false positives (18 instances) and false negatives (21 instances) are almost equal that indicates no bias toward over- or under-prediction of flashover events. It is worth to note that this balance is critical; since false negatives (i.e. missed flashovers) would increase the risk of failure while false positives (i.e. unnecessary alarms) might lead to unnecessary costly maintenance measures. The precision (86.8%) and recall (85.0%) of this matrix align with the previous F1 score as well that was shown in Fig. 11a. This approves the robustness of the presented approach. The low false-negative rate is remarkable for higher pollution levels (30000–40000 µS/cm) where conventional models often fail.Fig. 13Confusion matrix for the presented CNN-LSTM model. Confusion matrix for the presented CNN-LSTM model. While the proposed hybrid model is promising, several limitations should be considered for future researches. These include unseen parameters in the trained model, complex environmental conditions and long-term aging states. Practical deployment challenges include sensor requirements (LC measurement which may not be available across all grid assets), raising cost-benefit considerations for computational resource needs, and periodic recalibration with new data to combat concept drift (e.g. variable environmental patterns or unseen pollution sources). Although the test data are recorded from field-aged insulators, it is crucial to validate purely laboratory-based data through long-term unsupervised field trials in geographically diverse and harsh environments for true practical effectiveness. Application of measurement systems for condition monitoring could be usually challenging in cases of converting raw data into final control actions or remedial measures through the preventive maintenance (PM) scheme. Integrating domain knowledge into physics-induced learning would assist the presented model for data processing. This would compensate limitations concerning instrumentation precision. Utilized arc features can be extracted from the measured signals. From the industrial perspective where a reliable decision-making structure is critical for PM remedial programs, the presented framework can be considered as an intelligent metering system that can be deployed at various levels of the power system. Building up a real infrastructure for condition assessment of the outdoor insulators of an overhead line requires a full chain of measurement and sensing layers for LC records, precipitation and pollution deposit metering devices, main processing core for offline physics-induced learning patterns and real-time calculation of discharge transition alarms. In addition, an economical optimization in different scenarios might be necessary for unit deployment in any pollution zone map. Future application of the proposed method in real practices could begin with field pilot deployment by installing monitoring systems (LC sensors or cameras) on selected insulators in diverse geographical locations (coastal, industrial or desert) for evaluation of predictive accuracy and computational performance. The next stage could be integration into grid management systems through developing a software prototype that integrates the model into a condition monitoring system of a power utility or SCADA system. Its effectiveness would be assessed through triggering actionable alerts for preventive maintenance (PM) and potentially the claimed benefits for optimized insulator designs. Conclusions In this study, a hybrid physics-informed deep learning model is developed for the dynamic prediction of flashover on polluted composite insulators. The original novelty is the hybrid dynamic modeling of the real-time arc time constant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{arc}$$\end{document}) and arc velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{arc}$$\end{document}) with AI flashover prediction, which tries to make a connection between analytical arc models and transient discharge dynamics. It is shown that the integration of the presented arc model with a CNN-LSTM-Attention network for feature extraction from LC data can be an effective tool in real practices. The experimental validation approves performance and reliability of the presented approach through quantitative operational indexes. Validation results approve a consistent MAE of 0.30 ± 0.02 min for all discharge transition phases according to different pollution levels. In addition, comparative analysis between similar benchmark models verifies that the proposed CNN-LSTM-Attention architecture offers an optimal balance. For example, it offers better predictions with 89.1% accuracy, 87.3% F1 score, and a 0.924 AUC compared to benchmark models like Transformers, GRUs, and ANNs. This is obtained with preserving 3.95 ms inference time which reflects high computational efficiency. Statistical analysis (ANOVA) reveals key role of pollution level for the variance in discharge transition times (89%) rather than geometry of the insulators (less than 3%). The presented model tries to incorporate physics-induced feature engineering into PM schemes in power system utilities. For example, it is shown that a trigger alert is activated when predicted transition times is lower than a critical threshold (e.g., < 10 min for heavy pollution level). This would allow for further targeted remedial measures such as washing programs in a decision-making framework. Future work will focus on extended training dataset to include a wider variety of aging insulators and unseen complex environmental interactions. Contribution of the presented work to the predictive maintenance (PM) schemes can be pivotal in advancing power system protection technologies.