A predictive model for flow index performance of pit drip irrigation emitters using BP-PSO algorithm Authors: Weixiong Xu, Jiatong Jiang, Tianyu Xu Published: 2026 ABSTRACT It is crucial to accurately obtain the flow index in designing and developing labyrinth drip irrigation emitters. This study designed a pit drip irrigation emitter based on plant biomimetic principles, and created training-testing datasets (160 data points) and external validation datasets (25 data points) through experiments. The improved backpropagation neural network (BP) regression algorithm based on particle swarm optimization (PSO) is used to predict the flow index of the pit drip irrigation emitter. The input parameters include the pit depth l, the inner and outer boundary spacing h, the flow channel angle θ, and the pit aperture j. There were four combinations with the different hidden layer settings, namely BP1, BP2, BP1-PSO, BP2-PSO. 85% of the training-testing dataset was used to train the developed model (using 5-fold cross-validation), of which 15% was used for testing. The reliability of the prediction model was evaluated using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE), uncertainty at 95%(U95), coefficient of determination (R2), and global performance indicator (GPI). The results indicated that the optimal structures of BP1 and BP (input-hidden layer-output) were 4-7-1 and 4-4-9-1. The BP2-PSO model had the best prediction accuracy and stability, MAE, MSE, RMSE, MAPE, U95, R², and GPI values were 1.98 × 10− 3, 5.49 × 10− 6, 2.34 × 10− 3, 4.09 × 10− 3, 6.46 × 10− 3, 0.9456, and 3.74 × 10− 3, respectively. the evaluation values were 6.46 × 10− 3, 0.9456 and 3.74 × 10− 3 respectively. The GPI of BP1, BP2, BP1-PSO and BP2-PSO were 6.35 × 10− 3, 5.96 × 10− 3, 4.88 × 10− 3 and 3.74 × 10− 3 respectively, indicating good prediction accuracy. The ranking of prediction accuracy was as follows: BP2-PSO > BP1-PSO > BP2 > BP1. The four models were tested on the external validation dataset, the R² values of all four models were higher than 0.85, indicating good correlation. Through SHAP analysis, the importance of structural parameters to the flow index is ranked as: j > θ > h > l. In order to verify the validity and extensive superiority of BP-PSO model in solving drip irrigation emitter flow index prediction problem, it was compared with six representative machine learning models, including classical models (GR and MLR), integrated models (CatBoost and LSBoost) and hybrid models (MLP-SVM and MLP-DT). The BP2-PSO and BP1-PSO models had the highest prediction accuracy in the training-testing process among ten machine learning methods, and BP2-PSO model had the highest prediction accuracy in the external validation process, which indicated that particle swarm optimization can improve the accuracy of neural network models. The BP2-PSO model was used to predict the flow index had greater advantages during the design and development stages of pit drip irrigation emitters. Introduction The rational utilization of water resources is an important measure to promote agricultural development1–3. Drip irrigation is core technology in the field of water-saving irrigation, it has been widely applied in global agricultural water management4. The performance of drip irrigation emitters can reflect the working efficiency of the drip irrigation system5,6. The flow index is an important indicator for measuring the hydraulic performance of drip irrigation emitters, and its value ranges be-tween 0 and 17. The smaller the flow index has the lower the sensitivity of drip irrigation emitter to the working pressure8. The structure of the flow channel can lead to some differences in the flow index. It is crucial to accurately and quickly establish the relationship between the flow channel structure parameters and the flow index during the design, development, and optimization stages. At present, most scholars adopted single-factor or multi-factor orthogonal experiments and genetic algorithms to explore the relationship between the flow channel structure parameters and the flow index. ZHANG et al.9 employed the single-factor experimental method to explore the impact of the inter-tooth angle of the labyrinth flow channel on the hydraulic performance of drip irrigation emitters. The results showed the inter tooth angle increased, the flow quantity inside the drip irrigation emitter increased. When the inter tooth angle was 70°, the hydraulic performance was optimal. Single factor experiments had certain advantages for preliminary exploration and simple causal relationship analysis, but multi factor experiments is crucial to study of complex systems. Hou et al.10 utilized orthogonal experiments to study the influence of flow channel geometric parameters on hydraulic performance, and obtained the equation of the relationship between the flow index and these parameters through multiple linear regression analysis. The advantage of multi-factor experiments was that they can explore and explain the behavior of complex systems more comprehensively. However, the complexity of such experimental designs and the difficulty of data analysis were relatively high. Hu et al.11 designed the structural parameters of the flow channel using genetic algorithms and analyzed its hydraulic performance using CFD tool. The result shown through regression equations that the tooth staggering value was the most important factor affecting hydraulic performance. The genetic algorithm can provide effective optimization solutions in solving complex, nonlinear, nonconvex, combinatorial optimization and other problems. Its disadvantage was prone to prematurely converge to a local optimal solution, failing to find the global optimal solution. The above experiment combined with regression analysis model indirectly obtained the relationship between flow index and flow channel structural parameters. This method only discussed the independent influence of a single structural parameter. It was important to consider the interaction effects among the structural parameters12,13. Developing more reasonable, simple, and direct prediction methods was crucial for the design and development of drip irrigation emitters. When it comes to the multi-objective optimization, non-linearity and prediction problems faced in the de-sign of drip irrigation emitters, the use of machine learning methods has significant advantages14–16. Guo et al.17 developed a flow quantity prediction model for drip irrigation emitters by using support vector machine (SVM). The average error between this model and the experimental samples was 2.25%, demonstrating the high accuracy of the model. Mattar et al.18 employed the artificial neural network (ANN) model and gene expression programming (GEP) model to predict the flow quantity variations of drip irrigation emitters. They found that the ANN model had a lower error compared with the GEP model, indicating that the ANN model had a significant advantage in predicting the hydraulic performance of drip irrigation emitters. Chen et al.19 predicted the flow index of labyrinth drip irrigation emitters by using the response surface method (RSM) in combination with the ANN and support vector regression (SVR) models. The results showed that the SVR model had the highest prediction accuracy. The results showed that the SVR model had the highest prediction accuracy. Li et al.20 developed extreme learning machine (ELM), backpropagation neural network (BPNN) and traditional multiple linear regression (MLR) models to predict the flow index of SWRLC drip irrigation emitters. The comparison results showed that the ELM model performed the best, followed by the BPNN model. Machine learning method had achieved good results in the hydraulic performance prediction of drip irrigation emitters. The single model prediction method had limitations. There were many parameters or the nonlinear degree is high, the prediction results will be difficult to obtain the optimal solution. In hybrid algorithms, Achite et al.21 compared the predictive performance of standard SVR against two novel hybrid models (HBA-SVR and COOT-SVR) for monthly pan evaporation prediction in Algeria. Results demonstrated that employing the COOT optimization algorithm to identify optimal SVR parameter configurations significantly enhanced model performance when addressing complex prediction tasks involving multivariate, strongly non-linear relationships. Samantaray et al.22 pioneered the integration of a novel Sparrow Search Algorithm into SSL for optimizing SVM hyperparameters, thereby constructing an SVM-SSA hybrid forecasting model. Research demonstrated that the SSA algorithm effectively optimizes SVM parameters, enhancing its performance in complex non-linear forecasting tasks. Mirzania et al.23 proposed a novel hybrid model (AIG-SVR) for forecasting monthly water level variations in Lake Michigan and Lake Superior within the Great Lakes of North America. Results demonstrated that compared to a standalone SVR model, AIG-SVR exhibited superior generalization capabilities, successfully enhancing prediction accuracy. This provides water resource managers with a reliable tool to support decision-making. Combining different machine learning methods with optimization models (hybrid algorithms) is more conducive to model prediction. BP neural networks have non-linear fitting, self-learning and generalization capabilities. In order to predict the flow index of perforated drip irrigation emitters, the best hidden layer of the BP model is determined through five-fold cross verification to improve the prediction accuracy. Most of the existing hybrid model research focuses on internal verification and comparison, and lacks strict testing of its generalization ability and the applicability of engineering algorithms. This study conducts external verification based on the generalization ability of the algorithm to obtain the best flow index prediction accuracy. PSO is a population optimization algorithm that uses “particles” to simulate each bird in the flock and optimize the thresholds and input weights of the BP neural network. It has certain advantages in terms of model stability, convergence accuracy and speed24,25. Six different machine learning algorithms (Categorical boosting (CatBoost), Least squares boosting (LSBoost), Multilayer perceptron-support vector machine (MLP-SVM), Multi-layer perceptron-determination tree (MLP-DT), Gaussian regression (GR), and Multiple linear regression (MLR))) were simultaneously compared to determine the overall performance of the BP-PSO model. The purpose of this study was to: (1) construct a PSO-BP hybrid structure to solve the problem that BP neural network is sensitive to initial weights and easily falls into local optimality; (2) utilize external validation methods to assess the actual generalization ability of the model in future applications and employ SHAP analysis to evaluate the importance of structural parameters to the flow index; (3) design multi-class model comparison experiments to verify the versatility of BP-PSO. This study provides data and theoretical support for the design and performance optimization of drip irrigation emitters. Materials and methods The flow index is one of the main indicators for measuring the hydraulic performance of drip irrigation emitters, and it is an important parameter that must be considered during the design and development stages of drip irrigation emitters. Predicting the flow index of drip irrigation emitters quickly and accurately is of great significance for optimizing the design of drip irrigation systems, improving irrigation uniformity and enhancing anti-clogging performance. In this study, Machine learning (ML) method techniques were adopted to establish the non-linear relationship between the flow channel structural parameters of pit emitters and its flow index. The research framework of the proposed model is shown in Fig. 1. The research framework is divided into two parts. The first part is to screen four key structural parameters that affect the performance of drip irrigation emitters and determine their range of values. Subsequently, a controllable hydraulic performance test was carried out in the hydraulic laboratory to measure the stable flow of the drip irrigation emitter under different working pressures. Finally, the flow index was calculated through mathematical model (Eq. 1). Based on a large amount of experimental data, a mapping relationship data set of “Drip emitter structure parameters-Flow index” is constructed, and is divided into a training-testing set for training and adjusting models and an external verification set for final inspection. The research goes from simple to complex, and BP neural networks with single hidden layers (BP1) and dual hidden layers (BP2) are constructed as the basic models. PSO is introduced to intelligently optimize the hyperparameters of BP network, and two optimization models, BP1-PSO and BP2-PSO, are obtained. Six evaluation indicators are also introduced to make the evaluation very systematic. The optimized BP-PSO model is compared with six classic machine learning models such as CatBoost and LSBoost. Fig. 1Technical roadmap diagram. Technical roadmap diagram. Structural design and physical model of drip irrigation emitter A new type of pit drip irrigation device with diversion, rapid rotation and con-vergence was constructed using plant biomimetic principles. Tracheids are one of the main structures for water transportation in the xylem of plants26. The transportation of water mainly relies on the pit structure between adjacent tracheids, and the pit structure has a stable pressure-drop effect. This study adopted the most representative torus margo structure, as shown in Fig. 2 (a), and designed a novel pit drip irrigation emitter (PDIE), as shown in Fig. 2 (b). The optimized design was achieved by reducing the flow area at the upper and lower ends of the PDIE and adding resistance-forming triangles on both sides of the torus margin (where the oblique edges of the inner ring of the flow channel were parallel to the adjacent inner tooth edges), as shown in Fig. 2(c)27. The pit flow channel consisted of four structural parameters. l represented the pit depth, h represented the inner and outer boundary spacing, θ represented the flow channel angle, and j represented the pit aperture. The flow channel depth D was set to 0.6 mm, the height of the inner ring of the flow channel was 1.70 mm, and the number of pit units was 10. The physical model is shown in Fig. 3. According to the actual design requirements of the drip irrigation emitter and reference28, the value ranges of the key geometric parameters were determined (Table 1). Fig. 2(a) Schematic of torus-margo border pit (b) PDIE model (c) Optimized PDIE model. (a) Schematic of torus-margo border pit (b) PDIE model (c) Optimized PDIE model. Fig. 3Schematic diagram o f O-PDIE physical model structure. Schematic diagram o f O-PDIE physical model structure. Table 1The value ranges of structural parameters of O-PDIE emitter.Parametersl(mm)h(mm)θ(°)j(mm)Ranges1.2 ~ 1.60.4 ~ 0.8100 ~ 1200.6 ~ 1.0 Table 1The value ranges of structural parameters of O-PDIE emitter.Parametersl(mm)h(mm)θ(°)j(mm)Ranges1.2 ~ 1.60.4 ~ 0.8100 ~ 1200.6 ~ 1.0 The value ranges of structural parameters of O-PDIE emitter. Experimental method This study used 3D software to design a experimental model, which was carved using an EM-G32S-X32 high-precision carving machine (with a carving accuracy of 0.01 mm), and made of organic glass to form proportional pit drip irrigation emitters. Figure 4 showed the physical sample of the pit drip irrigation emitter and the experimental device. The measurement site is the School of Hydraulic and Electric Power of Heilongjiang University, the water temperature is 22 ~ 24℃, and the laboratory is a hydraulic branch room. The device mainly consisted of a water tank, a water pump, a pressure-reducing valve, a pressure gauge (0.1-grade precision), and several pipelines. The experimental device was divided into one layer, with five drip irrigation pipes installed on each layer. Each pipeline was equipped with 5 pit drip irrigation emitters, for a total of 25 pit drip irrigation emitters. Firstly, tested the flow quantity of the drip irrigation device with clean water under a working pressure of 50 kPa. Each test lasted for 5 min and was repeated 3 times. Recorded the flow quantity of each test with a measuring cylinder and took the average as the rated flow quantity. In the subsequent testing, the inlet pressure of the clean water experiment increased every 25 kPa and stopped at 250 kPa. The above operation was repeated to obtain the flow quantity at different working pressures. The relationship between the pressure and flow quantity of the irrigation flow channel can be obtained29:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q={k_{\mathrm{d}}}{h^x}$$\end{document} In the formula, q is the flow rate value (L/h) of the irrigation emitter under a given pressure, kd is the flow coefficient, h is the inlet pressure (kPa) and x is the flow index (whose value depends on the flow state inside the irrigation emitter). The pressure and flow curves of the irrigation emitter was fitted through regression analysis using IBM SPSS Statistics 25 software, and the value of the flow index was obtained. Fig. 4Physical and experimental setup of perforated drip irrigation emitter. Physical and experimental setup of perforated drip irrigation emitter. Establishment and division of datasets In this study, a total of two datasets were created. The first dataset contained 160 sets of experimental calculation values, excluding data points from orthogonal experiments. This dataset was only used for training and testing ML models and was named the training-testing dataset. The training-testing dataset was used to establish the ML model, as shown in Table 2. The second dataset was orthogonal experimental data points, which aims to analyze the performance of the trained ML model through external validation. The second dataset was named the external validation dataset, as shown in Table 3. Table 2Training-testing dataset.LevelGeometry parameters valuesLevelGeometry parameters valuesl(mm)h(mm)θ(°)j(mm)l(mm)h(mm)θ(°)j(mm)11.20.41050.8811.40.4110121.20.41101821.40.51050.731.20.41150.7831.40.51100.941.20.51001841.40.51150.651.20.51050.7851.40.61000.961.20.61000.9861.40.61050.671.20.61100.8871.40.61150.781.20.61151881.40.71000.891.20.71000.8891.40.71051101.20.71100.7901.40.71100.7111.20.71150.9911.40.71150.9121.20.81000.7921.40.81000.7131.20.81050.9931.40.81100.6141.20.81100.6941.40.81150.8151.20.81151951.40.41100.6161.20.61001961.40.41150.8171.20.51000.7971.40.41150.9181.20.61001981.40.41200.9191.20.61050.7991.40.51200.8201.20.41000.71001.40.61050.6211.20.41000.81011.40.61201221.20.41000.91021.40.71200.8231.20.410011031.40.81050.9241.20.41050.61041.40.81150.8251.20.41050.71051.40.81201261.20.41050.91061.40.41200.9271.20.410511071.40.61201281.20.41100.61081.50.41000.6291.20.41100.71091.50.41051301.20.41150.81101.50.41100.9311.20.41150.91111.50.41150.8321.20.411511121.50.51000.8331.20.51000.71131.50.51050.7341.20.51000.81141.50.51100.6351.20.61100.61151.50.51151361.20.61100.71161.50.61001371.20.61100.91171.50.61050.9381.20.611011181.50.61100.8391.20.61000.71191.50.61150.7401.20.61000.81201.50.71000.7411.20.610011211.50.71050.6421.20.51000.91221.50.71101431.20.61050.71231.50.81000.9441.20.61050.81241.50.81050.8451.20.61050.91251.50.81101461.20.51100.71261.50.81150.6471.20.51100.81271.50.41150.7481.20.51050.91281.50.51050.6491.20.51150.61291.50.51050.8501.20.61050.61301.50.51050.9511.20.710511311.50.51200.9521.30.41000.61321.50.71150.9531.30.41050.91331.50.71200.8541.30.41100.71341.50.81100.7551.30.411511351.50.41200.7561.30.51000.91361.60.71200.6571.30.511011371.60.41101581.30.51150.81381.60.41200.9591.30.61000.71391.60.51200.8601.30.610511401.60.61200.7611.30.61100.81411.60.81200.8621.30.61150.61421.60.41000.6631.30.710511431.60.41150.7641.30.71100.61441.60.51001651.30.71150.71451.60.51050.7661.30.81000.81461.60.51100.9671.30.81050.61471.60.61050.6681.30.81100.71481.60.61100.8691.30.811511491.60.61151701.30.611011501.60.71000.8711.30.71050.71511.60.71051721.30.51100.81521.60.71150.9731.30.51100.91531.60.81000.7741.30.51050.71541.60.81050.9751.30.41200.81551.60.81100.6761.30.71200.91561.60.81151771.30.61200.91571.60.81000.8781.30.71000.81581.60.61051791.40.41000.61591.60.51101801.40.41050.81601.60.51150.7 Table 2Training-testing dataset.LevelGeometry parameters valuesLevelGeometry parameters valuesl(mm)h(mm)θ(°)j(mm)l(mm)h(mm)θ(°)j(mm)11.20.41050.8811.40.4110121.20.41101821.40.51050.731.20.41150.7831.40.51100.941.20.51001841.40.51150.651.20.51050.7851.40.61000.961.20.61000.9861.40.61050.671.20.61100.8871.40.61150.781.20.61151881.40.71000.891.20.71000.8891.40.71051101.20.71100.7901.40.71100.7111.20.71150.9911.40.71150.9121.20.81000.7921.40.81000.7131.20.81050.9931.40.81100.6141.20.81100.6941.40.81150.8151.20.81151951.40.41100.6161.20.61001961.40.41150.8171.20.51000.7971.40.41150.9181.20.61001981.40.41200.9191.20.61050.7991.40.51200.8201.20.41000.71001.40.61050.6211.20.41000.81011.40.61201221.20.41000.91021.40.71200.8231.20.410011031.40.81050.9241.20.41050.61041.40.81150.8251.20.41050.71051.40.81201261.20.41050.91061.40.41200.9271.20.410511071.40.61201281.20.41100.61081.50.41000.6291.20.41100.71091.50.41051301.20.41150.81101.50.41100.9311.20.41150.91111.50.41150.8321.20.411511121.50.51000.8331.20.51000.71131.50.51050.7341.20.51000.81141.50.51100.6351.20.61100.61151.50.51151361.20.61100.71161.50.61001371.20.61100.91171.50.61050.9381.20.611011181.50.61100.8391.20.61000.71191.50.61150.7401.20.61000.81201.50.71000.7411.20.610011211.50.71050.6421.20.51000.91221.50.71101431.20.61050.71231.50.81000.9441.20.61050.81241.50.81050.8451.20.61050.91251.50.81101461.20.51100.71261.50.81150.6471.20.51100.81271.50.41150.7481.20.51050.91281.50.51050.6491.20.51150.61291.50.51050.8501.20.61050.61301.50.51050.9511.20.710511311.50.51200.9521.30.41000.61321.50.71150.9531.30.41050.91331.50.71200.8541.30.41100.71341.50.81100.7551.30.411511351.50.41200.7561.30.51000.91361.60.71200.6571.30.511011371.60.41101581.30.51150.81381.60.41200.9591.30.61000.71391.60.51200.8601.30.610511401.60.61200.7611.30.61100.81411.60.81200.8621.30.61150.61421.60.41000.6631.30.710511431.60.41150.7641.30.71100.61441.60.51001651.30.71150.71451.60.51050.7661.30.81000.81461.60.51100.9671.30.81050.61471.60.61050.6681.30.81100.71481.60.61100.8691.30.811511491.60.61151701.30.611011501.60.71000.8711.30.71050.71511.60.71051721.30.51100.81521.60.71150.9731.30.51100.91531.60.81000.7741.30.51050.71541.60.81050.9751.30.41200.81551.60.81100.6761.30.71200.91561.60.81151771.30.61200.91571.60.81000.8781.30.71000.81581.60.61051791.40.41000.61591.60.51101801.40.41050.81601.60.51150.7 Training-testing dataset. Both datasets consisted of 4 inputs and 1 output. Determine key geometric parameters according to relevant requirements and selected variables for input and output. The pit depth l, the inner and outer boundary spacing h, the flow channel angle θ, and the pit aperture j were used as the input parameters of the prediction model, while the flow index x was the output parameter employed in this study. Table 3External validation dataset.LevelGeometry parameters valuesLevelGeometry parameters valuesl(mm)h(mm)θ(°)j(mm)l(mm)h(mm)θ(°)j(mm)11.20.41000.6141.40.71200.621.20.51100.9151.40.81050.931.20.61200.7161.50.4110141.20.71051171.50.51200.851.20.81150.8181.50.61050.661.30.41200.9191.50.71150.971.30.51050.7201.50.81000.781.30.61151211.60.41050.891.30.71000.8221.60.51150.6101.30.81100.6231.60.61000.9111.40.41150.7241.60.71100.7121.40.51001251.60.81201131.40.61100.8 Table 3External validation dataset.LevelGeometry parameters valuesLevelGeometry parameters valuesl(mm)h(mm)θ(°)j(mm)l(mm)h(mm)θ(°)j(mm)11.20.41000.6141.40.71200.621.20.51100.9151.40.81050.931.20.61200.7161.50.4110141.20.71051171.50.51200.851.20.81150.8181.50.61050.661.30.41200.9191.50.71150.971.30.51050.7201.50.81000.781.30.61151211.60.41050.891.30.71000.8221.60.51150.6101.30.81100.6231.60.61000.9111.40.41150.7241.60.71100.7121.40.51001251.60.81201131.40.61100.8 External validation dataset. The training-testing dataset was randomly shuffled. 85% of the data was used as the training subset, while the remaining 15% served as the testing subset. Determine the validation set in the training subset using the k-fold cross validation method (k = 5). During the training process, the particle swarm optimization algorithm and k-fold CV were utilized to adjust the hyperparameters. After hyperparameter optimization, the testing subset was used to evaluate the model. K-fold CV randomly divided the original training subset into k sub-datasets. Among them, k-1 sub-datasets were used as the training set, and the remaining dataset was used as the new testing set. This process was repeated 5 times, and the performance of the model was evaluated by calculating the average mean squared error (MSE) of the testing sets in the 5 sub-datasets. In this way, all available data participated in the training and testing phases. The specific process was shown in Fig. 5. Fig. 5(a). experimental data division and (b). K-fold cross-validation. (a). experimental data division and (b). K-fold cross-validation. Establishment of flow index prediction model The BP is an artificial neural network based on the error back-propagation algorithm. It consists of three parts: the input layer, the hidden layer, and the output layer. The hidden layer can contain multiple neurons. The pit depth l, the inner and outer boundary spacing h, the flow channel angle θ, and the pit aperture j were the input values, and the flow index x was the predicted output value. The structure of the pit drip irrigation emitter was determined, the network was trained using the input and output datasets. the network parameters (weights and thresholds) were continuously adjusted to enable the network to achieve or approximate the desired input-output mapping relationship by back-propagating the error and simultaneously correcting the error. The specific steps are as follows: Step 1: Initialization of the network. Assume that the number of nodes in the input layer was n, the number of nodes in the hidden layer was n1, and the number of nodes in the output layer was m. The weight from the input layer to the hidden layer was wij, the weight from the hidden layer to the output layer was wjk, the bias from the input layer to the hidden layer was aj, the bias from the hidden layer to the output layer was bk, and the learning rate was η. Step 2: Output of the hidden layer. The output Hj of the hidden layer was:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_j}=g\left( {\sum\limits_{{i=1}}^{n} {{w_{ij}}{x_i} - {a_j}} } \right)$$\end{document} Step 3: The output of the output layer. The output Ok of the output layer was:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O_k}=\sum\limits_{{j=1}}^{{{n_1}}} {({H_j}} {w_{jk}}) - {b_k}$$\end{document} Step 4: The calculation of the error. The error value E was:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E = {Y_k} - {O_k}$$\end{document} In the formula, Yk represented the actual value, and Ok represented the predicted value. Step 5: Weight update. In the process of error back-propagation, the gradient descent method was used to update the weights. The gradient direction was determined by calculating the partial derivatives of the loss function with respect to the weights. The weights from the hidden layer to the output layer wij were updated, and the weights from the input layer to the hidden layer wjk were updated. The weight update formulas were as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {\begin{array}{*{20}{c}} {{w_{ij}} = {w_{ij}} + \eta {H_j}\left( {1 - {H_j}} \right){x_i}\sum\limits_{k = 1}^m {{w_{jk}}{e_k}} }\\ {{w_{jk}} = {w_{jk}} + \eta {H_j}{e_k}} \end{array}} \right.$$\end{document} Step 6: Bias update. The bias from the input layer to the hidden layer aj was updated, and the bias from the hidden layer to the output layer bk was updated. The bias update formulas were as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {\begin{array}{*{20}{c}} {{a_j}={a_j}+\eta {H_j}\left( {1 - {H_j}} \right)\sum\limits_{{k=1}}^{m} {{\omega _{jk}}{e_k}} } \\ {{b_k}={b_k}+{e_k}} \end{array}} \right.$$\end{document} Step 7: Determine whether the algorithm iteration had ended. This can be judged by checking if the error had reached the required precision. If not, repeat from Step 2. The structure of BP ensures efficient and accurate propagation of errors, achieving strong nonlinear mapping capability. It can prevent overfitting and underfitting, boost computational efficiency, and meet the demands of diverse application scenarios. One hidden layer and two hidden layers were set respectively for the BP, denoted as BP1 and BP2. The number of nodes corresponding to the minimum MSE in each case was selected to ensure obtaining the optimal number of nodes for the performance indicators in each case. Generally, the number of neurons in the input layer was determined by influencing factors. Each influencing factor represented a neuron in the input layer. In this study, the influencing factors were determined by four structural parameters. The number of nodes in the hidden layer was determined using empirical Eq. (7). In this BP model, the range of the number of hidden-layer neurons was (3 ~ 12). A schematic diagram of the BP1 model structure was shown in Fig. 6.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} {n_1}<\sqrt {(m+n)} +c \hfill \\ {n_1}={\log _2}n \hfill \\ {n_1}=\sqrt {0.43mn+0.12{m^2}+2.54n+0.77m+0.86} \hfill \\ {n_1}