Method for analyzing corona generation mechanism based on neural network algorithm
By constructing a neural network architecture based on CNN-LSTM-residual connections, the accuracy and efficiency problems of corona discharge in high-altitude environments were solved, and efficient and accurate analysis of the corona discharge mechanism and quantitative revelation of its dynamic evolution process were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-04-07
Smart Images
Figure CN121809232A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of conductor corona analysis technology, specifically a method for analyzing the corona generation mechanism based on a neural network algorithm. Background Technology
[0002] Ultra-high voltage (UHV) transmission lines, as the core carriers for inter-regional energy transmission, are seeing continuous expansion in their construction and operation at altitudes above 4000m. This region is characterized by unique low air pressure (30-60kPa), wide temperature range (-40℃~25℃), and strong ultraviolet radiation (20-80W / m²). 2 The large temperature difference between day and night leads to a significant "multi-field coupling" characteristic in the generation mechanism of corona discharge. Air pressure affects the air breakdown field strength, temperature regulates the particle motion rate, ultraviolet radiation triggers photoelectron emission, humidity changes the electron attachment probability, and the surface state of the conductor determines the initial electron source. The synergistic effect of multiple fields makes the entire process of corona discharge from triggering, development to stabilization exhibit extremely strong nonlinearity and dynamics.
[0003] Corona discharge directly affects the transmission efficiency, electromagnetic compatibility, and operational safety of power transmission lines. Accurately analyzing its generation mechanism is a core prerequisite for the optimized design, insulation configuration, and corona protection of high-altitude ultra-high voltage lines. However, existing methods for analyzing corona generation mechanisms still face many technical bottlenecks and are difficult to adapt to the complex scenarios of multi-field coupling at high altitudes. Traditional numerical methods suffer from three inherent drawbacks when analyzing corona phenomena in high-altitude environments: insufficient accuracy, low efficiency, and poor ability to fit nonlinear relationships. This leads to distorted boundary conditions, significant errors, and excessive computation time, making it difficult to meet real-time analysis requirements. Furthermore, existing models do not comprehensively consider multi-field coupling, generally neglecting key physical mechanisms such as photoelectron emission triggered by strong ultraviolet radiation and the evolution of wire surface roughness, resulting in fundamental limitations in revealing the essential laws governing corona.
[0004] Existing neural networks have significant limitations in analyzing corona characteristics. Their network architecture is not designed for the multi-field coupling characteristics at high altitudes, making it difficult to capture the spatiotemporal evolution of field quantities. At the same time, the lack of high-quality high-altitude-specific datasets leads to insufficient model generalization ability. In addition, existing studies mostly focus on predicting single parameters such as the initiation voltage, failing to achieve simultaneous analysis of key physical quantities in all stages of discharge, thus failing to support in-depth quantitative research on discharge mechanisms.
[0005] With the rapid advancement of high-altitude ultra-high voltage power transmission projects, there is an urgent need for a method to analyze the corona generation mechanism that can overcome the limitations of traditional methods, adapt to multi-field coupling characteristics, and combine high efficiency and high precision. This method can accurately reveal the dynamic evolution law of corona discharge in high-altitude environments and provide reliable technical support for line engineering applications. Summary of the Invention
[0006] In view of this, the present invention addresses the shortcomings of the prior art by providing a method for analyzing the corona generation mechanism based on a neural network algorithm.
[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for analyzing the corona generation mechanism based on a neural network algorithm, comprising the following steps: S1. Obtain high-altitude multi-field coupling feature data and perform data processing to obtain the input vector; S2. Establish a high-altitude multi-field coupling mathematical model and determine the output target; S3. Based on the high-altitude multi-field coupling mathematical model and output target, construct a high-altitude multi-field coupling adaptive neural network architecture; S4. Feed the input vector into the high-altitude multi-field coupled adaptive neural network for training; S5. Based on the output results of the trained high-altitude multi-field coupled adaptive neural network, the corona generation mechanism is analyzed.
[0008] Furthermore, in S1, the high-altitude multi-field coupling characteristic data include air pressure, thermodynamic temperature, relative humidity, ultraviolet radiation intensity, and conductor surface roughness.
[0009] Furthermore, in S1, the data processing methods include: Data standardization: The Z-score standardization method is used to normalize each feature data so that the mean is 0 and the standard deviation is 1. Where X is the original data, μ is the mean, and σ is the standard deviation; Feature dimensionality reduction and redundancy removal: Principal component analysis algorithm is used to perform linear transformation on the standardized five-dimensional feature data.
[0010] Furthermore, in S2, the high-altitude multi-field coupling mathematical model is a five-field coupling architecture consisting of air pressure field, temperature field, humidity field, ultraviolet radiation field, and conductor surface state field. The interaction paths of each field are as follows: air pressure affects air density and breakdown field strength; temperature regulates particle velocity; humidity affects electron attachment probability; ultraviolet radiation promotes photoelectron emission; and the conductor surface state determines the initial electron source. The model equations are constructed as follows: Electric field distortion equation: Where ε is the dielectric constant of air considering the effects of air pressure and temperature, φ is the electric potential, and ρ is the space charge density; Particle transport equations: Where n is the free electron concentration, vn Where S is the electron drift velocity, S is the electron generation rate, and R is the electron recombination rate; Ultraviolet radiation triggering equation: Where k is the light emission coefficient of the conductor surface, I is the ultraviolet radiation intensity, and σ is the absorption cross section of the conductor surface; Meteorological parameter regulation equations: Correlation function of air dielectric constant ε: Where P is the actual air pressure at high altitude, P0 is the standard atmospheric pressure, T is the actual thermodynamic temperature at high altitude, T0 is the standard temperature, and H is the relative humidity; the polarizability correction term is: Electron drift velocity v n Association function: Where E is the electric field strength; Correlation function of electron recombination rate R: Where α0 is the standard environmental recombination coefficient, n is the free electron concentration, and n ion Ion concentration; humidity correction term: Correlation function of corona initiation voltage U0: Solve the particle transport equation and Poisson's equation, with the boundary condition being the surface potential of the conductor and the ground potential being 0, to obtain the corona initiation voltage U0. Stream expansion rate V s Association function: Where A is the cross-sectional area of the stream channel; Correlation function of discharge channel current density J: Where e is the electron charge, approximately 1.602E-19C; The output targets are the corona initiation voltage U0, the free electron concentration n0, and the streamer propagation velocity V. s and discharge channel current density J.
[0011] Furthermore, in S3, the high-altitude multi-field coupling adaptive neural network architecture adopts a CNN-LSTM-residual connection hybrid architecture: the CNN module extracts the spatial coupling features of multiple field quantities, and automatically learns the local correlation and high-level abstract features between different field quantities by sliding the convolution kernel on the feature dimension; the LSTM module captures the temporal dynamic evolution law of the discharge process, and selectively memorizes and forgets historical information through the gating mechanism to accurately model the temporal causal relationship in the discharge process; Network parameter configuration for a high-altitude, multi-field coupled adaptive neural network architecture: 5-dimensional input layer, 3 CNN convolutional blocks + 2 LSTM layers in the feature extraction layer, 4 fully connected branches in the output layer, and ReLU + Sigmoid activation function.
[0012] Furthermore, in S5, based on the four parameters output by the neural network—corona initiation voltage, free electron concentration, streamer expansion velocity, and discharge channel current density—the regulatory weights of each field quantity on corona generation are inversely calculated. The critical concentration threshold of free electron avalanche, the negative correlation between streamer expansion velocity and gas pressure are analyzed, and an electric field distribution cloud map is output. Based on the sequence prediction capability of the LSTM network, the spatiotemporal evolution curve of electron concentration is reconstructed, the variation characteristics of electron concentration at the streamer head are recorded, the thickness parameter of the space charge layer is measured, and the distribution law of propagation velocity is analyzed.
[0013] Furthermore, the spatial coupling characteristics of multiple fields include the correlation between air pressure and electric field distribution, temperature and electron mobility, humidity and electron adhesion rate, ultraviolet radiation and photoelectric emission, and the correlation between conductor surface roughness and local field strength distortion; the temporal dynamic evolution of the discharge process includes the electron avalanche growth sequence, the streamer expansion sequence, the space charge layer evolution sequence, and the discharge current pulsation sequence.
[0014] Furthermore, in S1, high-altitude multi-field coupling characteristic data are obtained through on-site measurements, laboratory simulations, and data simulation extensions.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the inherent limitations of traditional numerical methods in handling complex boundaries, nonlinear fitting, and computational efficiency by constructing a five-field coupling feature system of "air pressure-temperature-humidity-ultraviolet radiation-conductor surface state" and a "CNN-LSTM-residual connection" neural network architecture. At the same time, it makes up for the shortcomings of existing neural network models in terms of insufficient adaptability to high-altitude multi-field coupling scenarios, weak dataset targeting, and single analytical dimension. Attached Figure Description
[0016] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2This is a schematic diagram illustrating the principle of multi-field coupled feature extraction in the CNN module of this invention. Figure 3 This is a schematic diagram of the timing dynamic modeling principle of the LSTM module in this embodiment of the invention; Figure 4 This is a diagram of the overall architecture of the CNN-LSTM hybrid neural network in an embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.
[0018] Example: Refer to Figure 1 The method for analyzing the corona generation mechanism based on neural network algorithms has the following steps: S1. Obtain high-altitude multi-field coupling feature data and perform data processing to obtain the input vector; S2. Establish a high-altitude multi-field coupling mathematical model and determine the output target; S3. Based on the high-altitude multi-field coupling mathematical model and output target, construct a high-altitude multi-field coupling adaptive neural network architecture; S4. Feed the input vector into the high-altitude multi-field coupled adaptive neural network for training; S5. Based on the output results of the trained high-altitude multi-field coupled adaptive neural network, the corona generation mechanism is analyzed.
[0019] In step S1, the high-altitude multi-field coupling characteristic data includes air pressure, thermodynamic temperature, relative humidity, ultraviolet radiation intensity, and conductor surface roughness. This data is obtained through field measurements, laboratory simulations, and data simulation extensions. Core field parameters are selected for the 4000-5500m altitude range, including air pressure (30-60 kPa), thermodynamic temperature (233-298 K), relative humidity (10%-80%), and ultraviolet radiation intensity (20-80 W / m²). 2 The surface roughness of the conductor (0.1-0.5μm) is measured to identify and process missing and outlier values in field measurements, laboratory simulations, and modeling data to ensure that data from different sources are aligned in terms of timestamps and operating conditions.
[0020] In step S1, the data processing method includes: Data standardization: The Z-score standardization method is used to normalize each feature data so that the mean is 0 and the standard deviation is 1. Where X is the original data, μ is the mean, and σ is the standard deviation; Feature Dimensionality Reduction and Redundancy Removal: To avoid the curse of dimensionality and highlight principal components, Principal Component Analysis (PCA) is used to linearly transform the standardized five-dimensional feature vector. By retaining the top N principal components with a cumulative variance contribution rate ≥ 95%, the input vector is compressed from 5 dimensions to N dimensions (usually N ≥ 3). While retaining most of the original information, a low-redundancy, highly representative final input vector is formed for feeding into the neural network.
[0021] The system integrates 3,000 sets of field-measured data from four altitude gradients (4,000-5,500m), which are the most authentic and authoritative, and are used for final model validation and core training. It also includes 2,000 sets of controllable data from a low-pressure, low-temperature simulation laboratory. This data has the advantages of "controllable operating conditions, accurate parameters, and no noise interference," effectively supplementing the lack of field data under extreme boundary conditions (such as extremely low temperatures and extreme dryness), and improving the model's predictive reliability in sparse data regions. Finally, it includes 5,000 sets of extended COMSOL simulation data. This low-cost, wide-coverage simulation data efficiently fills the feature space and can realize dangerous or ideal operating conditions that are difficult to access in the field and experiments, providing a solid "data foundation" for the model and enhancing its generalization ability and robustness.
[0022] Three methods—field measurement, laboratory simulation, and simulation extension—were used to obtain data considering five input features, resulting in a high-quality labeled dataset of 10,000 sets representing the input features and the true output values.
[0023] In step S2, the high-altitude multi-field coupling mathematical model is a five-field coupling architecture consisting of air pressure field, temperature field, humidity field, ultraviolet radiation field, and conductor surface state field. The interaction paths of each field are as follows: air pressure affects air density and breakdown field strength; temperature regulates particle velocity; humidity affects electron attachment probability; ultraviolet radiation promotes photoelectron emission; and the conductor surface state determines the initial electron source. The model equations are constructed as follows: Electric field distortion equation: Where ε is the dielectric constant of air considering the effects of air pressure and temperature, φ is the electric potential, and ρ is the space charge density; Particle transport equations: Where n is the free electron concentration, v nS is the electron drift velocity (controlled by temperature and electric field), S is the electron generation rate (including ultraviolet radiation triggering term), and R is the electron recombination rate. Ultraviolet radiation triggering equation: Where k is the light emission coefficient of the conductor surface, I is the ultraviolet radiation intensity, and σ is the absorption cross section of the conductor surface; Meteorological parameter regulation equations: Correlation function of air dielectric constant ε: Where P is the actual air pressure at high altitude, P0 is the standard atmospheric pressure, T is the actual thermodynamic temperature at high altitude, T0 is the standard temperature, and H is the relative humidity; the polarizability correction term is: Electron drift velocity v n Association function: Where E is the electric field strength; Correlation function of electron recombination rate R: Where α0 is the standard environmental recombination coefficient, n is the free electron concentration, and n ion Ion concentration; humidity correction term: Correlation function of corona initiation voltage U0: Solve the particle transport equation and Poisson's equation, with the boundary condition being the surface potential of the conductor and the ground potential being 0, to obtain the corona initiation voltage U0. Stream expansion rate V s Association function: Where A is the cross-sectional area of the stream expansion; Correlation function of discharge channel current density J: Where e is the electron charge, approximately 1.602E-19C; Key parameters of the corona generation mechanism for focused output targets: corona initiation voltage U0, free electron concentration n0, and streamer propagation velocity V. s And the discharge channel current density J. These four factors are intrinsically related and collectively characterize the discharge mechanism. Establish the relationship between air pressure, temperature, humidity, and ε, v. nThe correlation function of parameters such as R is used to quantify the dynamic impact of environmental factors. The neural network learns four tasks simultaneously to capture the inherent constraints and correlations between different physical processes.
[0024] In step S3, the high-altitude multi-field coupling adaptive neural network architecture adopts a CNN-LSTM-residual connection hybrid architecture: the CNN module extracts the spatial coupling features of multiple field quantities (such as the correlation between air pressure and electric field distribution, temperature and electron mobility, humidity and electron adhesion, ultraviolet radiation and photoelectric emission, and the correlation between conductor surface roughness and local field strength distortion), the LSTM module captures the temporal dynamic evolution of the discharge process (such as the electron avalanche growth time series, the streamer expansion time series, the space charge layer evolution time series, and the discharge current pulsation time series), and the residual connection solves the gradient vanishing problem of deep networks and improves the fitting ability of complex coupling relationships.
[0025] Reference Figure 2 The CNN module extracts the spatial coupling features of multiple field quantities, and automatically learns the local correlation and high-level abstract features between different field quantities by "sliding" the convolution kernel along the feature dimension.
[0026] Reference Figure 3 The LSTM module captures the temporal dynamic evolution of the discharge process and uses a gating mechanism to selectively memorize and forget historical information, accurately modeling the temporal causal relationships in the discharge process.
[0027] During the computation of an LSTM unit from time step (t-1) to t, the flow of information is jointly regulated by three gating mechanisms: (1) Gate of Oblivion The information to be discarded from long-term memory is denoted by t, where t represents the current time and x represents the information to be discarded. t The inputs at the current moment are: air pressure P, temperature T, humidity H, ultraviolet radiation intensity UV, and wire roughness R. t-1 It is the hidden state from the previous time step, containing all historical sequence information up to the previous time step; it is the short-term memory of the LSTM. t This is the output of the forget gate, with a value between 0 and 1.
[0028] W and b are the weights and biases, which are parameters that need to be learned in the model. σ is the sigmoid activation function, which compresses the output to between 0 and 1.
[0029] (2) Input gate Store new information in long-term memory, i tThe output of the input gate, which is between 0 and 1, determines how much new candidate information is updated into long-term memory.
[0030] (3) Output gate Outputting information from long-term memory, o t It is the output of the output gate, which is between 0 and 1.
[0031] The output information f of the forget gate and the input gate t and i t It participates in cell state renewal and updates the state of long-term memory cells.
[0032] C represents the cell state. Based on the regulation of the output gate and the updated cell state, the hidden state output at the current time step is calculated.
[0033] h t It is the updated short-term memory, which is the final output of the LSTM at the current moment.
[0034] Reference Figure 4 The parameter configuration of the high-altitude multi-field coupling adaptive neural network is as follows: the input layer has a 5-dimensional dimension, the feature extraction layer contains 3 CNN convolutional blocks (3×3 kernel size) + 2 LSTM layers (128 hidden units), the output layer has 4 fully connected branches (corresponding to 4 types of target parameters), and the activation function adopts the ReLU+Sigmoid combination (adapting to different parameter value ranges).
[0035] Step S4 also includes training strategy optimization: network weights are initialized using transfer learning (pre-trained on low-altitude data, then fine-tuned on high-altitude data), the AdamW optimizer is selected, and a cosine annealing scheduler is employed. The learning rate gradually decays from its initial value of 0.001 to 0 according to a cosine function as the training progresses.
[0036] The weighted mean squared error (Weighted-MSE) is used as the loss function. Weights were assigned to each output parameter based on their engineering importance: initial voltage 30%, electron concentration 25%, streamer velocity 25%, and current density 20%. The model was trained iteratively for 1000 epochs using an early stopping method. Training was automatically stopped and the model weights with the lowest validation set loss were rolled back after 50 consecutive epochs to prevent overfitting. During training, performance was evaluated on the validation set after each epoch, and the model status was monitored in real time.
[0037] In step S5, the method for analyzing the corona generation mechanism is as follows: Quantification and visualization of corona generation mechanism: Based on four parameters output by neural network—corona initiation voltage, free electron concentration, streamer propagation velocity, and discharge channel current density—the regulatory weights of each field on corona generation are inversely calculated. The critical concentration threshold of free electron avalanche, the negative correlation between streamer propagation velocity and gas pressure are analyzed, and an electric field distribution cloud map is output. Based on the sequence prediction capability of LSTM network, the spatiotemporal evolution curve of electron concentration is reconstructed, the variation characteristics of electron concentration at the streamer head are recorded, the thickness parameter of space charge layer is measured, and the distribution law of propagation velocity is analyzed.
[0038] The standardized, preprocessed five-dimensional environmental feature vector is input into a trained CNN-LSTM hybrid neural network. The model can simultaneously output four key parameters—corona initiation voltage, free electron concentration, streamer propagation velocity, and discharge channel current density—within milliseconds through a single forward propagation. This end-to-end computational approach avoids the complex partial differential equation solving and iterative convergence processes of traditional numerical methods, improving computational efficiency from minutes to milliseconds in traditional finite element methods.
[0039] This invention provides a method for analyzing the corona generation mechanism using a neural network algorithm, overcoming the limitations of existing methods such as low analytical accuracy, poor computational efficiency, incomplete consideration of multi-field coupling, and insufficient neural network adaptability. It constructs a five-field coupling feature system encompassing air pressure, temperature, humidity, ultraviolet radiation, and conductor surface state, along with a dedicated neural network architecture of "CNN-LSTM-residual connection." This enables accurate model training and efficient fitting of nonlinear correlations among multiple fields supported by high-altitude, high-quality datasets, improving the analytical accuracy of key parameters throughout the corona generation process. It quantitatively reveals the complete dynamic evolution of corona discharge from free electron avalanche initiation and streamer development to stable discharge, clarifying the regulatory weights and mechanisms of each field. Finally, it establishes a three-in-one analytical result verification system of "on-site measurement - laboratory simulation - simulation verification" to ensure the reliability, generalization ability, and engineering applicability of the method.
[0040] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still make modifications or equivalent substitutions to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are within the protection scope of the claims of the present invention.
Claims
1. A method for analyzing the corona generation mechanism based on a neural network algorithm, characterized in that, Includes the following steps: S1. Obtain high-altitude multi-field coupling feature data and perform data processing to obtain the input vector; S2. Establish a high-altitude multi-field coupling mathematical model and determine the output target; S3. Based on the high-altitude multi-field coupling mathematical model and output target, construct a high-altitude multi-field coupling adaptive neural network architecture; S4. Feed the input vector into the high-altitude multi-field coupled adaptive neural network for training; S5. Based on the output results of the trained high-altitude multi-field coupled adaptive neural network, the corona generation mechanism is analyzed.
2. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 1, characterized in that: In S1, the high-altitude multi-field coupling characteristic data include air pressure, thermodynamic temperature, relative humidity, ultraviolet radiation intensity, and conductor surface roughness.
3. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 2, characterized in that, In S1, the data processing methods include: Data standardization: The Z-score standardization method is used to normalize each feature data so that the mean is 0 and the standard deviation is 1. Where X is the original data, μ is the mean, and σ is the standard deviation; Feature dimensionality reduction and redundancy removal: Principal component analysis algorithm is used to perform linear transformation on the standardized five-dimensional feature data.
4. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 3, characterized in that, In S2, the high-altitude multi-field coupling mathematical model is a five-field coupling architecture consisting of air pressure field, temperature field, humidity field, ultraviolet radiation field, and conductor surface state field. The interaction paths of each field are as follows: air pressure affects air density and breakdown field strength; temperature regulates particle velocity; humidity affects electron attachment probability; ultraviolet radiation promotes photoelectron emission; and the conductor surface state determines the initial electron source. The model equations are constructed as follows: Electric field distortion equation: Where ε is the dielectric constant of air considering the effects of air pressure and temperature, φ is the electric potential, and ρ is the space charge density; Particle transport equations: Where n is the free electron concentration, v n Where S is the electron drift velocity, S is the electron generation rate, and R is the electron recombination rate; Ultraviolet radiation triggering equation: Where k is the light emission coefficient of the conductor surface, I is the ultraviolet radiation intensity, and σ is the absorption cross section of the conductor surface; Meteorological parameter regulation equations: Correlation function of air dielectric constant ε: Where P is the actual air pressure at high altitude, P0 is the standard atmospheric pressure, T is the actual thermodynamic temperature at high altitude, T0 is the standard temperature, and H is the relative humidity; the polarizability correction term is: Electron drift velocity v n Association function: Where E is the electric field strength; Correlation function of electron recombination rate R: Where α0 is the standard environmental recombination coefficient, n is the free electron concentration, and n ion Ion concentration; humidity correction term: Correlation function of corona initiation voltage U0: Solve the particle transport equation and Poisson's equation, with the boundary condition being the surface potential of the conductor and the ground potential being 0, to obtain the corona initiation voltage U0. Stream expansion rate V s Association function: Where A is the cross-sectional area of the stream channel; Correlation function of discharge channel current density J: Where e is the electron charge; The output targets are the corona initiation voltage U0, the free electron concentration n0, and the streamer propagation velocity V. s and discharge channel current density J.
5. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 4, characterized in that: In S3, the high-altitude multi-field coupling adaptive neural network architecture adopts a CNN-LSTM-residual connection hybrid architecture: the CNN module extracts the spatial coupling features of multiple field quantities, and automatically learns the local correlation and high-level abstract features between different field quantities by sliding the convolution kernel on the feature dimension; the LSTM module captures the temporal dynamic evolution law of the discharge process, and selectively memorizes and forgets historical information through the gating mechanism to accurately model the temporal causal relationship in the discharge process. Network parameter configuration for a high-altitude, multi-field coupled adaptive neural network architecture: 5-dimensional input layer, feature extraction layer with 3 CNN convolutional blocks + 2 LSTM layers, output layer with 4 fully connected branches, and activation function using a combination of ReLU and Sigmoid.
6. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 5, characterized in that: In S5, based on the four parameters output by the neural network—corona initiation voltage, free electron concentration, streamer propagation velocity, and discharge channel current density—the regulatory weights of each field quantity on corona generation are inversely calculated. The critical concentration threshold of free electron avalanche and the negative correlation between streamer propagation velocity and gas pressure are analyzed. An electric field distribution cloud map is output. Based on the sequence prediction capability of the LSTM network, the spatiotemporal evolution curve of electron concentration is reconstructed, the variation characteristics of electron concentration at the streamer head are recorded, the thickness parameter of the space charge layer is measured, and the distribution law of propagation velocity is analyzed.
7. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 5, characterized in that: The spatial coupling characteristics of multiple field quantities include the correlation between air pressure and electric field distribution, temperature and electron mobility, humidity and electron adhesion rate, ultraviolet radiation and photoelectric emission, and the correlation between conductor surface roughness and local field strength distortion. The dynamic evolution of the discharge process over time includes the electron avalanche growth sequence, the streamer expansion sequence, the space charge layer evolution sequence, and the discharge current pulsation sequence.
8. The method for analyzing the corona generation mechanism based on a neural network algorithm according to claim 1, characterized in that: In S1, high-altitude multi-field coupling characteristic data are obtained through on-site measurements, laboratory simulations, and data simulation extensions.