Hybrid quantum-classical neural networks for real-time fault detection in power systems Authors: Hina Hashmi, Aman Kumar, Shanu Rakesh Kuttan, Jhankar Moolchandani, Priya Goel, Aseel Smerat, Yaikob Abriham Massebo, Arshad Hashmi, Palaniyappan Sathyaprakash, Palaniyappan Sathyaprakash, Palaniyappan Sathyaprakash Published: 2026 ABSTRACT As modern power systems continue to be integrated with renewable energy sources, the ability to detect faults in these systems quickly and accurately is becoming more sophisticated. A Hybrid Quantum Classical Neural Network (HQCNN) model is presented in this paper that addresses the real-time fault detection challenge in transmission systems. A classical feature extractor, in this case a 1D CNN, and a quantum circuit are combined by the model, which aids in the classification of faults. Evaluations were conducted on simulated IEEE 14 and 39-bus system and on approximately 800 real PMU fault events. An accuracy of 96.43% on simulated data and 94.74% on real PMU data was achieved by the model, outperforming traditional deep learning models and maintaining detection times under 3 milliseconds. Various kind of fault like SDL (single line to ground), DL (double line), TP (three phase), and high-impedance faults, were correctly classified by the system. In a study in which elements of the model were removed, the quantum layer was found to be very important for improved performance. Issues like hardware limits and quantum noise were also looked at. As for the future, larger PMD data sets will be worked on, model explainability will be improved with hybrid XAI methods, and smaller HQCNN models will be developed for use in substation edge devices. 1. Introduction Power systems and electric cars are being implemented on the modern power grid systems. As they are being implemented, clean energy and decentralized generation are being supported by these changes, which are also creating new issues. From these challenges, one issue is related to the quick and precise detection of faults that need to be ensured. Some other Issues like short circuits, line-to-ground problems, and equipment failures can cause wide-scale outages if not attended in time [1]. These related troubles are often encountered by traditional protection methods that use fixed thresholds values and perform impedance calculations in modern grids, which outputs high variability, noise, and complex network structures. Hence, there is a need to utilize machine learning techniques to overcome these problems. In the case of faults, deep learning has shown improvements in detection, through which complex patterns are learned from the traces of voltage and current signals [2,3]. Some models like CNNs and LSTMs, are used often which are performed very well in classification and time-based analysis [4]. However, large, labeled data sets are usually required by these models, long training times are reported, and a great need for computing power is present, which in turn makes them a hard sell for real-time or edge-based applications [5].Additionally, many of the deep learning models function like “black boxes,” which does not provide much insight into the decision-making process [6,7]. I. Quantum Computing: A New Paradigm I. Quantum Computing: A New Paradigm A new way to address complex, high-dimensional learning problems is offered by quantum computing using quantum mechanical effects such as entanglement, superposition, etc. [8]. Significant speedups in data processing are potential achievements of QML(Quantum Machine Learning), especially for tasks like clustering, classification, optimization, etc. [9]. However, major limitations are still present in present quantum hardware that we have known as NISQ devices that includes a fixed count of qubits, imperfect gate fidelity, and high sensitivity to noise [10,11]. Because of these constraints, full practicality for deployment in real-world fault detection systems is not yet achieved by fully quantum models. II. Hybrid Quantum Classical Neural Networks II. Hybrid Quantum Classical Neural Networks To overcome the hardware limitations and still benefiting from its strengths, an HQCNNs (Hybrid Quantum-Classical Neural Networks) is proposed. Classical components such as CNNs (convolutional neural networks) are combined with parameterized quantum circuits (PQCs) to improve feature learning and classification. More accurate data processing is achieved by using the power of both classical and quantum models with current hybrid approach. High accuracy, fast response, and efficient computation are offered by it, making perfect smart grid system to achieve real time fault detection. III. Objectives and Contributions III. Objectives and Contributions An HQCNN(Hybrid Quantum Classical Neural Network) is designed for real-time fault diagnosis in power grid systems is proposed in this study. Quantum computing layers are incorporated into a conventional deep learning architecture by the model, allowing complex patterns in transient electrical signals to be effectively captured while maintaining reasonable computational cost. The primary contributions of the paper are: Designing, implementation of an HQCNN for fault classification using voltage and current waveform data. Designing, implementation of an HQCNN for fault classification using voltage and current waveform data. Development of a simulation environment using IEEE 14-bus and 39-bus systems to generate labeled fault data. Development of a simulation environment using IEEE 14-bus and 39-bus systems to generate labeled fault data. Performance benchmarking of the proposed HQCNN model against conventional LSTM, CNN models in terms of and computational efficiency, accuracy, and latency. Performance benchmarking of the proposed HQCNN model against conventional LSTM, CNN models in terms of and computational efficiency, accuracy, and latency. Analysis of quantum circuit configurations and encoding methods to evaluate the trade-offs between model accuracy and quantum resource requirements. Analysis of quantum circuit configurations and encoding methods to evaluate the trade-offs between model accuracy and quantum resource requirements. IV. Paper Structuring IV. Paper Structuring The work is drafted by section 1 is itself having the introduction, followed by Section 2, that is dedicated for related work on classical and quantum approaches to fault detection. Next, in section 3, the proposed hybrid model and data processing techniques are elaborated. Following section 4 describes various simulation setup, evaluation process. Finally, section 5 showcases conclusion and discussion of the results. 2. Related work Accurate and real-time fault detection in power systems is regarded as a critical component of intelligent grid operation. Prior research is categorized in this section into four main areas: traditional protection techniques, Quantum Machine Learning applications, classical and deep learning-based fault diagnosis, hybrid quantum classical neural networks, etc. 2.1. Conventional fault detection techniques Traditional methods for fault detection are included by over current protection, distance relays, differential protection, and impedance-based techniques [12,13]. Fixed thresholds are relied upon by these systems and manual configuration for each network topology is often required, which can be error-prone under dynamic grid conditions. To improve fault localization, time-frequency analysis, noise tolerance methods such as the transform using wavelet [14], the S-transform [3], and Hilbert-Huang Transform (HHT) [15] have been introduced. Although better transient feature extraction is offered by these methods, performance still degrades in noisy environments and careful tuning is required. Mathematical morphology [16] and Kalman filtering [8] have been explored by some studies as signal enhancement tools for improving the reliability of traditional detection frameworks. However, rule-based characteristics remained by these approaches and adaptability to grid reconfigurations, or the stochastic nature of RES is often lacking. 2.2. Deep learning, machine learning approaches New avenues for data-driven fault detection have been opened by machine learning (ML). Statistical and spectral features extracted from voltage and current signals have been widely applied using SVMs (Support Vector Machines), Random Forests and Decision Trees [9,17]. This capability is further enhanced by deep learning (DL) through the automatic learning of hierarchical features. Spatial pattern recognition in waveform data has been achieved using CNNs [18], while time dependencies in transient responses have been effectively captured by LSTMs [19]. Hybrid CNN-LSTM architectures have been developed by several researchers to leverage the strengths of both models [20,21]. Anomaly detection and data augmentation in imbalanced fault datasets have also been explored using Autoencoders and Generative Adversarial Networks (GANs) [22,23]. Despite their strong performance, substantial computational power and large training datasets are typically demanded by deep learning models, which limit their feasibility for real-time edge-based applications. Furthermore, challenges in safety-critical domains such as power system protection are posed by the opaque nature of these models. While approaches such as attention mechanisms, explainable AI have been proposed for improving the transparency, uncommon remains their practical deployment in operational systems [24,25]. 2.3. Power systems using quantum machine learning QML has been emerged as a useful approach for handling high-dimensional complex problems, including tasks such as classification, regression, and clustering. Potential advantages over classical techniques in specific cases have been shown by several quantum algorithms, including variational quantum classifiers, quantum support vector machines, and quantum kernel methods [26]. Within energy systems, QML has been explored for applications such as load prediction, energy usage analysis, and power flow optimization. An underexplored area remains fault detection, though the feasibility of using quantum algorithms to classify transient waveforms has been demonstrated in initial efforts. Due to the limitations of NISQ devices like qubit decoherence and limited circuit depth, fully quantum models are not yet considered suitable for real-time grid applications [27]. As a result, a shift towards hybrid approaches is being seen in research. 2.4. Hybrid quantum classical neural networks Classical deep learning components are integrated with quantum layers—usually parameterized quantum circuits (PQCs)—in hybrid quantum-classical neural networks (HQCNNs) [28,29]. Quantum-enhanced feature mapping is allowed by these architectures while classical processing is retained for scalability and stability. Circuit-centric quantum classifiers and quantum convolutional networks (QCNNs) have been successfully applied to image recognition, cybersecurity, and finance [30,31]. Hybrid quantum models have also been explored in small-scale power systems for short-term load forecasting and anomaly detection [32].‌‌ This paper aims to fill that gap by designing and evaluating a HQCNN model that classifies fault types based on transient waveform features using a simulation-based dataset from IEEE test systems. Table 1 depicts the summary of various literature Review on Fault Detection in Power Systems. 10.1371/journal.pone.0349887.t001Table 1Literature Review Summary on Fault Detection in Power Systems.Conventional MethodsRef. No.MethodologyDataset TypeKey ContributionLimitation[12]Neural NetworksSimulatedFast fault detection using neural nets in transmission linesFixed architecture, low adaptability[14]Wavelet TransformSimulatedTransient feature extraction for faultsSensitive to noise, window size selection[15]Hilbert-Huang Transform (HHT)SimulatedNonlinear signal analysis for transient detectionComputationally expensive[16]Mathematical MorphologyReal Fault LogsRobust noise filtering for transient detectionLimited to binary signal interpretation[8]Kalman FilteringSimulatedEnhanced signal quality in fault detectionRequires accurate system modeling Machine Learning Methods [9]SVMSimulatedFault classification from current/voltage signalsRequires hand-crafted features[33]Random ForestSimulatedEnsemble-based robust classificationLess effective with time-series data[17]Decision TreesReal + SimulatedSimple and interpretable fault detectionLow accuracy for complex transients Deep Learning Methods [18]CNNSimulatedAutomatic feature extraction from waveform dataNeeds large labeled datasets[19]LSTMSimulatedTime-series fault classificationTraining complexity, vanishing gradients[21]CNN-LSTMSimulatedCombined spatial-temporal feature learningComputationally intensive[22]GANSimulatedSynthetic fault data generationTraining instability and interpretability issues[23]GANSimulatedSynthetic fault data generationRisk of mode collapse, interpretability Quantum Machine Learning (QML) [34]VQCSimulated (Small)Circuit-based quantum classifierLimited qubit scalability[35]QMLSimulated (Energy)Quantum model for load-level disaggregationDomain-specific, not used for fault detection[36]Quantum ClassifierSimulated (Small)Initial fault classification on small datasetsNot tested in large-scale grids Hybrid Quantum-Classical Methods [30]Quantum-ClassicalImage BenchmarkQuantum CNN model for classification tasksStill in research phase[31]Hybrid ModelSimulated (Load Forecasting)QML with classical pre/post-processingNot applied to transient fault signals Literature Review Summary on Fault Detection in Power Systems. Despite advancements in fault detection, key research gaps persisted in, including the lack of real-time QML applications (G1), limited use of hybrid quantum-classical models in power systems (G2), no benchmark comparisons with deep learning models (G3), poor model explainability (G4), and reliance on simulated rather than real-world data (G5). The addressing of these gaps is required through innovative integration of QML and classical models for improved real-time fault detection in power systems. Table 2 highlights the research gaps identified in literature. 10.1371/journal.pone.0349887.t002Table 2Research Gaps Identified in Literature.Identified Gap IdIdentified GapDescriptionRelated Works G1 Lack of QML in real-time fault detectionFew works apply QML to real-time transient classification[36, 34] G2 Limited hybrid quantum-classical deployment in power systemsMost HQCNN models are focused on finance or vision[30, 31] G3 No benchmark on hybrid model vs. DL in power faultsNo systematic study comparing hybrid and DL on the same grid datasets[18, 21,30] G4 Explainability in hybrid systemsNo research yet on explainable HQCNNs for critical grid applications[37, 30] G5 Simulated vs. real data generalizationMost work tested only on simulation data; real-world adaptability not explored[8,9,12,26,33,38] Research Gaps Identified in Literature. A comparison of the performance of different fault detection approaches is presented by Tables 3 and 4. Well in controlled environments, conventional methods, including neural networks and wavelet-based techniques, are performed, but noise and complex fault conditions encountered in real-world systems are often failed to be handled. Higher accuracy for both temporal and spatial data is achieved by more advanced models, such as CNNs and LSTMs; however, increased computational demand is incurred by this improvement. Promising accuracy (up to 95%) is shown by quantum-based methods like the Variational Quantum Classifier (VQC), but limitations in scalability and dataset size are faced. Tables 3–5 provide a structured comparison of fault detection approaches, dataset characteristics, and methodological capabilities. A summary of quantitative performance metrics reported in prior studies and in the proposed method is presented in Table 3. A qualitative comparison of classical, deep learning, and hybrid quantum–classical approaches based on reported capabilities in the literature is provided in Table 4. The datasets used in this work, including their source, fault types, and overlap status, are described in Table 5. 10.1371/journal.pone.0349887.t003Table 3Evaluation Metrics and Performance Comparison of Fault Detection Methods.Ref. No.MethodologyEvaluation MetricPerformance ScoreComments[12]Neural NetworksAccuracy90%High accuracy but struggles with noise handling[14]Wavelet TransformSignal-to-Noise Ratio (SNR)92%Effective in clean datasets, less robust in noisy conditions[9]SVMPrecision, Recall85% (Precision), 80% (Recall)Excellent for binary classification, but struggles with complex faults[18]CNNAccuracy, F1-Score95%, 93%Excellent at feature learning but requires large labeled datasets[19]LSTMF1-Score, Loss90%, 0.1Strong temporal dependency modeling but resource-intensive[23]GANFID (Fréchet Inception Distance), Accuracy0.2 (FID), 85%Good at generating synthetic data but not real-time[34]VQCQubit Fidelity95%Performs well with small datasets but struggles with larger ones[30]QCNNAccuracy, Circuit Depth88%, 10^3Good accuracy, but requires high computational resources[12]Neural NetworksAccuracy90%High accuracy but struggles with noise handling Evaluation Metrics and Performance Comparison of Fault Detection Methods. 10.1371/journal.pone.0349887.t004Table 4Comparison of Classical, Deep Learning, and Quantum-Based Fault Detection Methods.CriteriaClassical TechniquesDeep LearningQuantum & Hybrid ModelsAccuracyModerateHighPromising (depends on quantum backend)Real-Time CapabilityHighMedium (depends on model size)Limited (due to quantum hardware)ScalabilityGoodHigh (but needs more data)Low (NISQ limitation)InterpretabilityHighLow (black-box)Very lowDataset RequirementSmallLargeMedium (can work with less due to QML)Computational CostLowHighHigh (especially with hybrid simulation)Suitability for Edge DevicesYesPartiallyNot yet feasibleHigh: consistently demonstrated across multiple benchmark studies; Moderate: demonstrated under task-specific or constrained settings; Low: limited empirical evidence or high dependency on tuning. Comparison of Classical, Deep Learning, and Quantum-Based Fault Detection Methods. High: consistently demonstrated across multiple benchmark studies; Moderate: demonstrated under task-specific or constrained settings; Low: limited empirical evidence or high dependency on tuning. 10.1371/journal.pone.0349887.t005Table 5Dataset Characteristics for Fault Detection Methods.Ref. No.Dataset TypeDataset SizeDataset SourceFault TypesMethodOverlap[12]Simulated10,000 instancesCustom MATLAB/Simulink simulationTransient, Short-circuitNeural NetworksNo[14]Real Fault Logs1,500 logsPower Grid OperatorLine faults, Equipment failureWavelet TransformNo[9]Simulated + Real5,000 instancesGrid Simulator, Public datasetLine faults, OverloadSupport Vector MachineNo[18]Simulated20,000 samplesCustom fault simulatorShort-circuit, GroundingCNNNo[19]Simulated15,000 samplesSynthetic fault simulatorTransient faults, OscillationsLSTMNo[21]Simulated25,000 samplesPower system grid simulatorTransients, Equipment FailureCNN-LSTM HybridNo This work Simulated + Real ~800 events + simulated cases IEEE 14/39-bus, PMU/SCADA (EPRI, LARIAT) SLG, DL, TP, HIF HQCNN No Dataset Characteristics for Fault Detection Methods. Table 5 is emphasized by how both the type and size of datasets influence the evaluation of fault detection models. Simulated data is relied upon by many studies, which allows a wide range of fault patterns to be learned by models such as neural networks and CNNs. In contrast, smaller real-world datasets are used by some approaches, including wavelet-based methods and support vector machines, leading to more reliable and realistic performance assessments. Very limited sample sizes especially constrain quantum models, such as variational quantum classifiers. Overall, the major challenge in accurately evaluating model robustness and generalization capability is posed by the scarcity of real-world data. In Table 5, the term overlap is referred to by the authors as relating to whether the same fault events or data samples have appeared across different studies or dataset partitions. In addition, the performance of fault detection methods is compared across various fault categories in Table 6. Transient faults are well identified by neural networks and wavelet-based approaches, while strong results for short-circuit fault detection are shown by CNNs. Ground faults are particularly well modeled by LSTMs because of their ability to capture temporal patterns, and robust performance is demonstrated by hybrid CNN–LSTM models when mixed fault conditions are dealt with. Moderate performance across fault types is shown by quantum-based models, such as Quantum CNNs, but hardware limitations and limited dataset availability currently constrain their effectiveness. 10.1371/journal.pone.0349887.t006Table 6Fault Detection Performance by Fault Type.Ref. No.Fault TypeMethodologyDetection PerformanceKey Findings[12]Transient FaultsNeural NetworksHighEffective in detecting fault inception but limited by noise[14]Short-CircuitWavelet TransformModerateGood at handling sharp transitions but weak in multi-fault scenarios[9]Line FaultsSVMHighPerforms well for single fault types; struggles with multi-faults[18]Short-CircuitCNNVery HighExcellent for detecting fault patterns in signal images[19]Ground FaultsLSTMHighStrong in capturing temporal dynamics of faults over time[21]Mixed FaultsCNN-LSTM HybridVery HighCombines spatial and temporal features for more complex fault scenarios[23]Any FaultsGANModerateLimited in real-time, good for generating synthetic training data[30]Line & Transient FaultsQuantum CNNModeratePerforms well on simple faults but struggles with real-world complexity Fault Detection Performance by Fault Type. 3. Methodology The suggested framework for real-time fault detection in power systems is utilized by an HQCNN (Hybrid Quantum-Classical Neural Network), which integrates with deep learning for extracting the features and a quantum variational circuit for classification, as illustrated in Figs 1–3. The shortcomings of entirely classical models regarding both accuracy and inference speed are aimed to be overcome by this architecture, especially in the identification of transient faults within high-resolution time-series data. An Improved HQCNN for Real-Time Fault Detection. Enhanced Version of HQCNN Architecture for Real-Time Fault Detection. Four-qubit, three-layer variational quantum circuit with parameterized rotations and linear CZ entanglement. 3.1. Framework architecture The HQCNN architecture consists of three major components: (i) a classical feature extraction module using either a 1D CNNor a BiLSTM network, (ii) a classical-to-quantum encoding layer, and (iii) a Variational Quantum Classifier (VQC). Given a raw signal X∈Rn×t, where, n is used as the count of sensors and t is the term used to show the length of the time window, the objective is to predict fault classes y∈ {0,1,...,k} through a non-linear mapping composed of classical and quantum transformations. The overall formulation is given as: where F(X) is the classical feature extraction function, ϕ denotes the quantum encoding, and VQC refers to the quantum variational circuit. The function σ represents the softmax operator that transforms the quantum measurement probabilities into a probabilistic class prediction. The input signal X, representing a segment of voltage or current waveforms from multiple sensors, is initially processed by a classical feature extraction function F(X), implemented using either a 1D Convolutional Neural Network (CNN) or a Bidirectional Long Short-Term Memory (BiLSTM) network. The extracted features are mapped to a quantum state through an encoding function ϕ(·) and processed by a variational quantum circuit (VQC).The final output y^ is calculated through a Softmax activation as: 3.2. Data collection and preprocessing For testing the suggested model, simulated and actual data were utilized. Simulated datasets have been created based on the IEEE 14-bus test system through MATLAB/Simulink under different fault scenarios such as single-line-to-ground (SLG), double-line (DL), and three-phase (TP) faults. Actual data was obtained from PMU/SCADA logs of open-access sources such as EPRI and LARIAT. The training and testing dataset includes both simulated and actual data. Simulated fault cases were created based on the IEEE 14-bus and 39-bus test systems, including single-line-to-ground (SLG), double-line (DL), and three-phase (TP) faults with different fault resistances and inception angles. The dataset in the real world includes about 800 events gathered from Phasor Measurement Units (PMUs) and SCADA logs, retrieved from open-source sites like EPRI and LARIAT. These field occurrences comprise events of single-line-to-ground, double-line, three-phase, and high impedance faults (HIF). Simulated data were neat and well-synchronized at 50/60 Hz sampling, whereas field data needed extra steps of preprocessing.The simulated datasets and real-world PMU/SCADA datasets are non-overlapping; training and controlled evaluation were conducted using simulated data, while validation and testing were carried out using real-world events only. All real-world PMU/SCADA data used in this study were obtained from publicly accessible repositories associated with smart grid monitoring projects, and no restricted or proprietary datasets were accessed. Wavelet-based denoising was done more aggressively to counteract ambient noise, event synchronization was also fixed by cross-correlation methods, and missing data in short bursts (less than 20 ms) were filled linearly for maintaining temporal coherence. Following preprocessing, fixed length overlapping windows were created from the synchronized voltage and current signals to capture localized fault dynamics. To prevent information leakage, data splitting was conducted at the fault-event level, ensuring that all windows derived from a given event were assigned exclusively to either the training, validation, or test set. 3.3. Data preprocessing and feature normalization As signal quality is important aspect, for this, several important steps taken here for data preprocessing to ensure consistent signal quality. Initially, wavelet-based denoising is applied to detect and remove high-frequency components that were not related to fault events. Then the signals were divided into fixed 5-cycle segments to keep the input size balanced and uniform. At the end, Min–Max normalization was used to scale down all features to the range of 0–1, to make the data suitable for quantum encoding. During preprocessing, the x is used as raw signal and it is cleaned from noise by applying Discrete Wavelet Transform, in which it is broken down into approximation coefficients named Aj and detail coefficients named Dj in eq. 3. To enhance the models ability to generalize across different grid scenarios, preprocessing was applied. First, the signal is denoised using Discrete Wavelet Transform (𝐷𝑊𝑇), where a signal (𝑡) is decomposed into approximation 𝐴𝑗(𝑡) and detail coefficients 𝐷𝑗(𝑡). Denoising is performed by thresholding Di(t) and reconstructing the signal via Inverse DWT(IDWT). After denoising, signals are normalized between [0,1] using Min-Max normalization. To avoid information leakage and preserve event-level amplitude characteristics that are critical for fault discrimination, Min–Max normalization statistics is computed once on training dataset independently. It is done for each signal channel. The same normalization parameters are then applied unchanged to the validation and test datasets. Here, Normalization is not performed on a per-window basis. The data is then segmented into fixed-size windows of 5 cycles (100 ms for 50 Hz systems), with each segment forming an input sample. Dataset Composition and Split: The complete dataset consists of both simulated and real-world fault events. Simulated data were generated using the IEEE 14-bus and 39-bus test systems, while approximately 800 real-world fault events were collected from PMU/SCADA repositories. The dataset includes balanced representations of SLG (singlelinetoground), DL (double-line), and TP (three-phase) faults to avoid class bias. The combined dataset was divided into 70%, 15%, and 15% for training, validation and testing, respectively, at the fault-event level to prevent information leakage. The same data split was used to generate the results reported in Fig 6 and Table 7. 10.1371/journal.pone.0349887.t007Table 7Comparative performance of models on simulated IEEE 14-bus data.ModelAccuracy (%)Precision (%)Recall (%)F1-score (%)Inference Latency (ms) CNN 93.1291.8792.4592.152.1 BiLSTM 94.0693.1592.7992.973.6 QVC 88.7487.4286.9187.161.8 HQCNN 96.4395.9296.0896.002.7 Comparative performance of models on simulated IEEE 14-bus data. Confusion Matrix. ROC Curves- HQCNN by Fault Type. The x-axis is showing the predicted fault classes, whereas the y-axis is showing the true fault labels. 3.4. Classical feature extraction The traditional component of the HQCNN is responsible for extracting high-level features from the pre-processed input. Two network architectures were considered as alternatives. The first one employs a 1D CNN consisting of convolutional layers with ReLU activations, max pooling, and fully connected layers. This configuration effectively extracts spatial correlations in waveform patterns. The second setup utilizes a BiLSTM to identify temporal dependencies that are of particular importance when detecting the propagation and evolution of faults over time. The feature extractor F(X) is either a 1D CNN or a BiLSTM network. For CNN, the output of a convolutional layer is given as: where wk(l) is the filter weight at layer l, b(l) is the bias, and σ(·) is a non-linear activation such as ReLU. In BiLSTM, the hidden state at time t is updated by combining forward ht→ and ht←backward states: The final output vector z∈Rd encapsulates temporal-spatial information and is passed to the quantum encoder. 3.5. Classical-to-quantum encoding For quantum encoding, each classical feature zi is mapped to a qubit’s rotation angle using the Ry gate: The encoded quantum state is the tensor product across all qubits. For enabling quantum processing, the classical vector w is encoded into a quantum state through angle encoding (also known as parameterized rotation encoding). Every scalar zi is translated into the angle for rotation of a quantum gate. The classical output w=[w1,w2,…,wV] is encoded through quantum states using angle encoding: This centred angle encoding preserves sign information and improves the expressive capacity of shallow variational quantum circuits while remaining compatible with NISQ hardware constraints. The complete quantum state is: This process yields a multiqubit entangled input state for the variational quantum circuit. The quantum variational circuit computes the expectation value of Pauli-Z operators on every qubit: The resulting expectation values are passed through the Softmax function to achieve the various class probabilities: The network is trained by the cross-entropy loss for minimization: whereyi is used to represent the true labels, y^i is showing the predicted class probabilities. A hybrid optimization procedure jointly updates classical parameters θc and quantum parameters θq by minimizing: where λ controls the regularization strength on quantum parameters.This unified loss function is used throughout the hybrid training process to jointly optimize classical and quantum parameters. 3.6. Variational quantum classifier (VQC) The quantum classification phase employs a parameterized quantum circuit to be run on the encoded input. The quantum variational layer employed in this work runs on four qubits, providing adequate expressiveness with compatibility for existing NISQ hardware. The quantum circuit depth is fixed at three layers, where every layer consists of parameterized single-qubit rotations and entangling operations. Entanglement is introduced through Controlled-Z (CZ) gates between neighbouring qubits following each rotation block. Controlled-Z (CZ) gates were chosen for entanglement as they preserve the computational basis structure and are natively supported on several quantum hardware platforms, while offering equivalent entangling capability to CNOT gates for variational circuits. This entanglement is used to preserve non-linear correlations between features in an efficient manner. To reduce the effect of noise byhardware, simple measurement error mitigation procedures were utilized using Qiskit Ignis calibration routines. Only measurement (readout) error mitigation using Qiskit Ignis calibration routines was applied; no full quantum error correction codes were implemented. Furthermore, the Simultaneous Perturbation Stochastic Approximation (SPSA) optimizer was chosen to train the quantum parameters, due to its stability in noisy environments and effectiveness in approximating gradients from sparse quantum measurements. Each variational layer consists of trainable singlequbit rotation gates Ry(ϕj(l)) applied to all qubits, followed by Controlled-Z (CZ) gates between adjacent qubits to introduce entanglement. The rotation angles ϕj(l)are optimized during training using the SPSA optimizer as part of the hybrid quantum–classical learning process. The quantum circuit consists of repeating layers of singlequbit parameterized rotation gates and entanglement gates (e.g., CZ or CNOT). The variational circuit output is a probability distribution over measurement outcomes ∣y⟩, for which expected value for a Pauli-Z observable on qubit i is: The expectation values obtained from the variational quantum circuit are passed through a softmax function to produce class probabilities. Model training is performed by minimizing the unified cross-entropy objective defined in Equation (13), with quantum parameters optimized using the SPSA algorithm. 3.7. Hybrid training and optimization strategy The HQCNN model got trained end-to-end by a hybrid optimization approach. Classical parameters (CNN/BiLSTM weights) are updated using stochastic gradient descent, while quantum parameters are optimized on simulators or real quantum hardware via the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm. Let 𝐰denote the classical parameters and θthe quantum parameters. Training is performed by minimizing the unified objective function defined in Equation (13), which jointly optimizes both classical and quantum components with a regularization term controlling the contribution of the quantum parameters. 3.8. Evaluation setupand evaluation metrics The proposed HQCNN model was developed using PyTorch for the classical deep learning components and Qiskit Machine Learning for implementing the variational quantum circuit. Experiments were carried out on a high-performance computing system equipped with an Intel Xeon processor, an NVIDIA RTX 4090 GPU, and 64 GB of memory, which supported efficient execution of both classical and quantum operations. The variational quantum circuit is built from repeated layers of parameterized single-qubit rotation gates and Controlled-Z (CZ) entangling gates, as outlined in Section 3.5. Expectation values of Pauli-Z operators are computed using measurement outcomes from the circuit, which serve as inputs for the fault classification task. Common classification metrics, including accuracy, precision, recall, and F1-score, are used to assess model performance, derived from true positive, true negative, false positive, and false negative results. Inference time is also measured to evaluate real-time applicability. All results are reported for each fault category and averaged using five-fold cross-validation to improve reliability and reduce variability. 4. Experimental results and analysis In this section the performance is analysedfor the HQCNN (Hybrid Quantum-Classical Neural Network) is presented using simulated and real-world datasets. Here, the aim is related to the evaluation of its performance in detecting faults accurately under different operating conditions and to compare it with current classical approaches. 4.1. Performance on simulated data The performance of the HQCNN against the common CNN, BiLSTM, and shallow quantum-only classifier (QVC) is collated in Table 7. Its counterparts were consistently outperformed by the HQCNN across all metrics. As Table 7 is showing the best accuracy (96.43%) and F1-score (96.00%) were achieved by the HQCNN, while a competitive inference latency of 2.7 ms was maintained. The observed performance gain is characterized to the ability of the quantum variational layer to model complex non-linear feature interactions through parameterized rotations and entangling operations, which can enhance decision boundaries in ambiguous fault scenarios. Fig 4 is showing Confusion Matrix of HQCNN on Real PMU Data and Fig 5 is showing ROC Curves. Training and Accuracy loss is showing by the Fig 6. 4.2. Fault-type detection accuracy To further understand model behaviour, the detection accuracy was broken down by fault type. Table 8 presents fault-wise classification results. 10.1371/journal.pone.0349887.t008Table 8HQCNN performance by fault type (simulated dataset).Fault TypeAccuracy (%)Precision (%)Recall (%)F1-score (%)Single-Line-Ground (SLG)97.3296.8897.1597.01Double-Line (DL)95.6494.9795.2395.10Three-Phase (TP)96.7196.3296.5596.43The HQCNN model demonstrated robust performance across all fault types, with slightly lower scores for double-line faults, possibly due to their similarity in signal characteristics with SLG events in some topologies. HQCNN performance by fault type (simulated dataset). The HQCNN model demonstrated robust performance across all fault types, with slightly lower scores for double-line faults, possibly due to their similarity in signal characteristics with SLG events in some topologies. 4.3. Generalization on real-world data To evaluate real-world applicability, the model was tested on PMU data collected from an open-access smart grid monitoring project. Table 9 shows the results. 10.1371/journal.pone.0349887.t009Table 9HQCNN generalization performance on real PMU dataset.ModelAccuracy (%)F1-score (%)Inference Latency (ms)CNN90.2189.452.4BiLSTM91.3890.914.1QVC86.0985.372.0 HQCNN 94.74 94.19 2.9 HQCNN generalization performance on real PMU dataset. Even under noisy and asynchronous conditions in real-world settings, the HQCNN retained high predictive capability and low inference latency, demonstrating its suitability for deployment in real-time energy management systems. Fig 7 depicts the Inference Latency Comparison Across Models. Training and validation loss (at the left) and accuracy curves (at the right) for the proposed HQCNN model‌‌. Inference Latency Comparison Across Models. 4.4. Ablation study To assess the individual contribution of the quantum layer, an ablation study was conducted. Results are presented in Table 10.In this ablation study, the quantum variational layer was removed and replaced with a fully classical dense layer of comparable size, while keeping all other network components and training settings unchanged. No hyperparameters were tuned during the ablation study; this design isolates the performance impact of the quantum layer itself. 10.1371/journal.pone.0349887.t010Table 10Impact of quantum component on classification performance.ConfigurationAccuracy (%)F1-score (%)Classical Only (CNN + FC)93.1292.15Classical + Quantum (HQCNN) 96.43 96.00 Impact of quantum component on classification performance. The 3.31% improvement in accuracy confirms that the variational quantum layer adds substantial discriminative power to the overall network. 5. Conclusion and future work An HQCNN architecture for real-time fault detection in power transmission systems was introduced in this paper. A 1D CNN or BiLSTM-based classical feature extractor was integrated with a four-qubit parameterized quantum circuit (PQC) for classification, with angle encoding and entanglement leveraged via Controlled-Z gates. Complete training of the model was conducted using a hybrid optimization strategy, which combined stochastic gradient descent for classical parameters and the SPSA (Simultaneous Perturbation Stochastic Approximation) algorithm for quantum parameters. Comprehensive experiments were conducted on both simulated datasets generated from IEEE 14-bus and 39-bus systems and real-world Phasor Measurement Unit (PMU) datasets containing approximately 800 real fault events. Fault types evaluated included SLG, DL, TP, and HIF.An overall accuracy 96.43% on simulated data and 94.74% on real-world PMU data was achieved by the HQCNN, with conventional CNN and BiLSTM baselines being outperformed. Inference latency was maintained within 2.7 milliseconds on simulated environments and under 3 milliseconds on real PMU data, with potential for real-time deployment demonstrated. It was confirmed by ablation studies that classification performance was improved by 3.31% over purely classical networks through the inclusion of the quantum variational layer. Despite the encouraging performance, several challenges remain for practical deployment. Computational resources on edge devices were limited by real-time latency constraints, and widespread adoption continues to be restricted by noise in NISQ (Noisy Intermediate-Scale Quantum) hardware. In addition, although strong robustness was demonstrated by HQCNN, questions related to model interpretability and transparency in fault decision-making remain unresolved. These limitations can be addressed in several ways by future research. First, the proposed HQCNN should be evaluated using larger and more diverse real-world PMU datasets to better assess its generalization across different grid configurations and operating conditions. Second, transparency would be enhanced, operator confidence would be improved, and regulatory requirements would be supported by incorporating explainable AI techniques tailored to hybrid quantum–classical models.Finally, versions of the HQCNN that are simplified and resource-efficient should be designed for deployment on edge devices in substations, ensuring an effective balance between detection accuracy and computational cost. Finally, the feasibility of HQCNNs as quantum technology matures will be improved through the exploration of more advanced quantum noise mitigation methods and scalable circuit ansatzes.The proposed HQCNN framework is designed and evaluated for single-fault classification scenarios, where one dominant fault type is present at a given time. Simultaneous multi-fault detection was not considered in this study. The framework could be extended to multi-label fault detection by modifying the output layer and loss formulation, which is a potential direction for future work. 17 Dec 2025 -->PONE-D-25-43455-->-->Hybrid Quantum-Classical Neural Networks for Real-Time Fault Detection in Power Systems-->-->PLOS One Dear Dr. Massebo, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jan 31 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:--> A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. -->If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Palaniyappan Sathyaprakash, Ph.D Academic Editor PLOS One Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf 2. Please note that PLOS One has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, we expect all author-generated code to be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse. 3. Please update your submission to use the PLOS LaTeX template. The template and more information on our requirements for LaTeX submissions can be found at http://journals.plos.org/plosone/s/latex. 4. Please provide a complete Data Availability Statement in the submission form, ensuring you include all necessary access information or a reason for why you are unable to make your data freely accessible. If your research concerns only data provided within your submission, please write "All data are in the manuscript and/or supporting information files" as your Data Availability Statement. 5. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager. 6. Please upload a new copy of Figures 1 and 2 as the detail is not clear. Please follow the link for more information: https://journals.plos.org/plosone/s/figures 7. We note you have included a table to which you do not refer in the text of your manuscript. Please ensure that you refer to Tables 1 and 2 in your text; if accepted, production will need this reference to link the reader to the Table. 8. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. 9. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions -->Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. --> Reviewer #1: Partly Reviewer #2: Yes ********** -->2. Has the statistical analysis been performed appropriately and rigorously? --> Reviewer #1: Yes Reviewer #2: Yes ********** -->3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: No Reviewer #2: No ********** -->4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.--> Reviewer #1: Yes Reviewer #2: Yes ********** -->5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)--> Reviewer #1: I have reviewed the manuscript and it presents a great use case for fault detection in power systems . The methods are simple and replicable and is practical in NISQ devices. How ever I have some minor concerns related to encoding , VQC and Use of optimizers . Reviewer #2: 1. The purpose of Table 3, 4, and 5 and their corresponding columns is not very clear. 2. In Table 4, the meaning or measure of the various entries under the column 'classical techniques' and 'deep learning' is not very clear. What does moderate, high and other keywords stand for and how is this measured? 3. In Table 5 lists the various datasets, are these dataset overlapping? Their names if any, needs to be specified. 4. Font size of Fig1, Fig2 needs to be increased as they are not readable 5. The Figure 3 can be accommodated as a whole. 6. The representation of the variables in different equation should be consistent for a given variable. Eqn 1 and eqn 2 has different representation for the input signal. 7. In Fig5, the role of ‘even synchronization’ and ‘missing data handling’ is not clear. The paper begins with taking a signal and cutting it into parts. Request you to justify the missing data. 8. In VQC, why have you used CZ gates for entanglement and why not CNOT gate? Justification for the same is needed. 9. If you are using abbreviations for a particular word then be consistent and use the same throughout the paper. There is no need to use both the full form and abbreviation ahead in the paper. 10. To represent ith instance in an equation use subscript. Eqn 14, 15, and 16 subscripts are not properly represented. 11. I feel Eqn 15 is not correctly printed. 12. Equation 15 and 16 are not a part of the subsection 3.5. Similarly eqn 11, 12 and 13 are not related to subsection 3.4 13. Subsection 3.7 needs to be rewritten as it not very clear. 14. For Fig 6. and Table 7 what is the split ratio? How many samples were used for training and testing? Subsection 3.2 speaks about the data gathering but doesn’t tell us how many samples were simulated and how many of them actually collected and total count in terms of (SLG), double-line (DL), and three-phase (TP) faults. This information might is very much nesseary to see if the dataset is balanced. 15. What does the X and Y labels of confusion matrix stand for? 16. Fig. 8 and 9 shows training and validation accuracies/loss using for which particular techniques? 17. You are using the data from the project 'smart grid monitoring project' are proper permissions sought from the concerned authority? 18. In subsection 4.5, please explain the meaning of the 'An ablation study was conducted to assess the individual contribution of the quantum layer'. What hyper parameter was tuned in this ablation study? 19. This work speaks about applying error correction, what form of error corrections is implemented? 20. Can your proposed techniques of HQCNN detect multiple faults? If Yes, a justification needs to be given in the form of results comparing the various possible combinations. 21. Can you justify this statement used in your paper, 'The performance gain is attributed to the ability of the quantum layer to capture high-dimensional entanglements in the feature space, enhancing decision boundaries for ambiguous cases'. ********** -->6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.--> Reviewer #1: No Reviewer #2: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation. NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications. --> Submitted filename: Manuscript_Review_Normalisation_VQC.docx Submitted filename: PONE-D-25-43455.docx 11 Feb 2026 Manuscript ID: PONE-D-25-43455 Title: Hybrid Quantum-Classical Neural Networks for Real-Time Fault Detection in Power Systems Response to Reviewer #1 Reviewer #1 Comment Action Taken / Response Location of Change in Manuscript Reviewer Comment ID (used in manuscript) 1. Clarify whether normalization is performed per window or across the dataset. Clarified that Min–Max normalization parameters are computed using only the training set and then applied unchanged to validation and test sets to avoid information leakage and preserve amplitude characteristics. Section 3.3 – Data Preprocessing and Feature Normalization R1.1 2. Angle encoding limits expressivity due to restricted rotation range. Updated the encoding scheme to a centered angle mapping to preserve sign information and enhance expressivity for shallow variational quantum circuits. Section 3.5 – Classical-to-Quantum Encoding, Equation (8) R1.2 3. Quantum circuit figure shows fixed rotation angles instead of trainable parameters. Redrew the quantum circuit to explicitly depict trainable rotation gates (R_y(\phi_j^{(\ell)})) for all qubits and layers, ensuring consistency between text and figure. Figure 3, Section 3.6 – Variational Quantum Classifier (VQC) R1.3 4. Optimizer choice for quantum parameters is unclear. Explicitly stated that the SPSA optimizer is used exclusively for training quantum parameters due to its robustness in noisy NISQ environments. Sections 3.6 and 3.7 R1.4 5. Multiple or inconsistent loss formulations appear in the manuscript. Unified the training objective by defining a single loss formulation and removing redundant loss equations, ensuring consistency across all sections. Section 3.7 – Hybrid Training and Optimization Strategy R1.5 6. Error correction claims may be overstated. Clarified that only measurement (readout) error mitigation using Qiskit Ignis is applied, and no full quantum error correction codes are implemented. Section 3.6 – Variational Quantum Classifier (VQC) R1.6 7. Possible train–test leakage due to window-based segmentation. Added an explicit statement confirming that data splitting is performed at the fault-event level so that all windows from a given event belong exclusively to one dataset split. Section 3.2 – Data Collection and Preprocessing R1.7 Response to Reviewer #2 Reviewer #2 Comments Action Taken / Response Location of Change in Manuscript Reviewer Comment ID (used in manuscript) 1. Purpose of Tables 3–5 is unclear. Added an explanatory paragraph describing the intent and scope of Tables 3–5. Section 2.4 Paragraph immediately preceding Tables 3–5 R2.1 2. Qualitative terms in Table 4 (High, Moderate, etc.) are unclear. Added a footnote defining all qualitative descriptors based on reported empirical evidence. Section 2.4 Table 4 footnote R2.2 3. Dataset overlap in Table 5 is unclear. Added an explicit “Overlap” column, clarified non-overlapping datasets, and included a row describing the dataset used in this work. Table 5; Section 3.2 R2.3 4. Figures 1 and 2 are not readable. Regenerated Figures 1 and 2 with increased font size and improved resolution. Figures 1 and 2 R2.4 5. Figure 3 should be presented as a single figure. Redrew Figure 3 as a unified four-qubit, three-layer variational quantum circuit with corrected labels. Figure 3 R2.5 6. Inconsistent variable notation across equations. Unified notation across all equations using consistent symbols for inputs, encoding, and quantum operations. Sections 3.1–3.5 R2.6 7. Role of missing data handling is unclear. Expanded preprocessing description to include synchronization, denoising, and linear interpolation for short missing segments. Section 3.2 R2.7 8. Justification for using CZ instead of CNOT gates is missing. Added justification explaining CZ gates’ symmetry, hardware compatibility, and equivalent entangling capability. Section 3.6 – Variational Quantum Classifier (VQC) R2.8 9. Abbreviation usage is inconsistent. Ensured all abbreviations are defined once and used consistently throughout the manuscript. Throughout the manuscript R2.9 10. Subscripts for i-th instance are improperly represented (Eqs. 14–16). Removed problematic equations and corrected notation in remaining equations. Sections 3.6 R2.10 11. Equation 15 is incorrectly printed. Removed Equation 15 as part of unifying the loss formulation. NA NA 12. Some equations are not aligned with their subsections. Reorganized equations to appear only in their appropriate methodological subsections. Sections 3.5–3.7 R2.12 13. Subsection 3.7 is unclear. Completely rewrote the subsection Section 3.7 – Evaluation Setup and Metrics, now it is 3.8 R2.13 14. Dataset size, split ratio, and class balance are missing. Added a “Dataset Composition and Split” paragraph describing simulated vs. real data, SLG/DL/TP/HIF balance, and a 70/15/15 split corresponding to results. Section 3.3 – Data Preprocessing and Feature Normalization R2.14 15. Confusion matrix axes are unclear. Updated the confusion matrix caption to explicitly define x- and y-axis meanings. After Figure 4 caption R2.15 16. Training and validation curves (Figs. 8–9) are unclear. Clarified that training and validation loss and accuracy curves are presented in Figure 6 and correspond to the proposed HQCNN model. Figure 6 caption R2.16 17. Permissions for smart grid monitoring data are unclear. Clarified that all PMU/SCADA data were obtained from publicly accessible repositories and added a Data Availability Statement. Section 3.2; Data Availability Statement R2.17 18. Meaning of ablation study and tuned hyperparameters is unclear. Explained that the ablation study removes the quantum layer while keeping all other settings unchanged and that no hyperparameters were tuned. Section 4.5 – Ablation Study R2.18 19. What form of error correction is implemented? Clarified that only measurement (readout) error mitigation using Qiskit Ignis is applied and no full quantum error correction codes are used. Section 3.6 – Variational Quantum Classifier (VQC) R2.19 20. Can HQCNN detect multiple faults? Clarified that the model is evaluated for single-fault classification only and that multi-fault detection is left for future work. Section 5 – Conclusion and Future Work R2.20 21. Justification of the “high-dimensional entanglement” claim. Rephrased the statement to avoid overclaiming physical quantum advantage and attributed gains to improved modeling of non-linear feature interactions. Section 4.1 – Performance on Simulated Data R2.21 Submitted filename: Response_to_Reviewers.docx 6 May 2026 Hybrid Quantum-Classical Neural Networks for Real-Time Fault Detection in Power Systems PONE-D-25-43455R1 Dear Dr. Massebo, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Palaniyappan Sathyaprakash, Ph.D Academic Editor PLOS One Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions -->Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.--> Reviewer #1: All comments have been addressed Reviewer #2: All comments have been addressed ********** -->2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. --> Reviewer #1: Yes Reviewer #2: Yes ********** -->3. Has the statistical analysis been performed appropriately and rigorously? --> Reviewer #1: Yes Reviewer #2: Yes ********** -->4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: Yes Reviewer #2: Yes ********** -->5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.--> Reviewer #1: Yes Reviewer #2: Yes ********** -->6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)--> Reviewer #1: General Assessment The revised manuscript presents a relevant and timely study on hybrid quantum- classical learning for real-time fault detection in power systems. The topic is suitable for the journal’s scope, and the paper has improved substantially in response to the previous round of review. In particular, the revised version is stronger in methodological transparency, notation, and presentation of the hybrid HQCNN framework. The inclusion of both simulated IEEE test-system data and real PMU/SCADA events also strengthens the applied relevance of the study. Overall, I find that the major technical concerns raised previously, espe- cially those associated with Reviewer 1, have been addressed adequately. The manuscript is now close to publishable form. However, a small number of mi- nor issues remain, primarily related to consistency, language polishing, and a few residual presentation ambiguities. I therefore recommend acceptance after minor revision. Assessment of Reviewer 1 Comments 1. Normalization protocol: addressed. The revised manuscript now clarifies that Min–Max normalization parameters are computed from the training set and then applied unchanged to validation and test data. It also explicitly states that normalization is not performed on a per-window basis. This satisfactorily resolves the earlier concern regarding possible information leakage. 2. Expressivity of the angle encoding: addressed. The authors revised the encoding description and now introduce a centered angle mapping in Eq. (8), which is a meaningful improvement over the earlier restricted formulation. 3. Quantum circuit trainable parameters: addressed. Figure 3 and the corresponding text now indicate trainable rotation parameters, which aligns the figure with the method description and resolves the earlier in- consistency. 4. Optimizer specification: addressed. The manuscript now explicitly states that SPSA is used to optimize the quantum parameters. This addi- tion improves reproducibility and is appropriate for noisy/NISQ-oriented workflows. 5. Loss formulation: addressed. The training objective has been consol- idated into a single formulation, which improves internal consistency and readability. 6. Error correction claim: addressed. The revised text now correctly limits the claim to measurement/readout error mitigation using Qiskit Ignis and no longer overstates the use of full quantum error correction. 7. Train–test leakage via windowing: addressed. The manuscript now states that splitting is performed at the fault-event level so that all win- dows derived from a given event remain in a single split. This is an important clarification and satisfactorily addresses the leakage concern. Remaining Minor Comments 1. Author name consistency (pages 1 and 8). The corresponding author name is inconsistent between the submission metadata and the manuscript body. One location lists Yaikob Abriham Massebo, while the manuscript title page lists Yakob Abraham. This should be harmonized before final publication. 2. Abstract language and terminology (page 8). The abstract still requires a final language polish. Examples include: • “IEEE 14 and 39-bus system” should be revised to “IEEE 14- and 39-bus systems”; • “Various kind of fault like SDL” should be revised to “Various kinds of faults such as SLG”; • “PMD data sets” appears to be a typo and likely should read “PMU datasets”. These are minor editorial issues, but they affect professionalism and clar- ity. 3. Introduction requires copyediting (pages 8–10). The introduction has improved conceptually, but several sentences remain grammatically awkward or stylistically non-standard for a journal article. This applies especially to the opening paragraph and the subsections “Quantum Com- puting: A New Paradigm,” “Hybrid Quantum Classical Neural Networks,” and “Paper Structuring.” A final English-language edit is recommended. 4. Experimental split description should be made fully unambigu- ous (pages 18–19). Section 3.2 states that simulated data were used for training and controlled evaluation, while validation and testing were carried out using real-world events only. However, Section 3.3 states that the combined dataset was divided into 70/15/15 for training, validation, and testing. These statements may be reconcilable, but in the current wording they create confusion. The authors should present one clear and consistent description of the final experimental protocol. 5. Encoding equations should be checked for internal consistency (page 20). Eq. (8) introduces a centered encoding, θi = π(2zi − 1), but Eq. (9) appears to revert to a non-centered form using RY (πzi). If this is not intentional, the equations should be aligned so that the mathematical description matches the stated encoding strategy. 6. Duplicate entry in Table 3 (page 13). Table 3 appears to repeat the row for Ref. [10] (Neural Networks). This duplication should be removed. 7. Data availability statement should be finalized carefully (page 6 and main text). The manuscript states that the data are fully available and points to a public dataset link. Before final acceptance, the authors should ensure that the final published manuscript contains a complete and durable data availability statement in the journal’s preferred format, including a stable repository link or DOI where applicable. 8. Minor terminology polishing throughout. Although the abbrevia- tion issue has largely been improved, the manuscript would still benefit from one final pass to ensure fully consistent usage of terms such as SLG, DL, TP, HIF, PMU, SCADA, HQCNN, and VQC throughout the text, tables, and figure captions. Conclusion In summary, this is a stronger and more coherent revision. The principal methodological concerns from the prior review round, especially those raised by Reviewer 1, have been addressed satisfactorily. The remaining issues are minor and editorial in nature, with one methodological clarification still needed regarding the exact train/validation/test protocol. Subject to those small revi- sions, I support publication. Final Recommendation: Minor Revision Reviewer #2: (No Response) ********** -->7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.--> Reviewer #1: No Reviewer #2: No ********** PONE-D-25-43455R1 PLOS One Dear Dr. Massebo, I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS One. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Palaniyappan Sathyaprakash Academic Editor PLOS One