A composite tower overvoltage simulation model design method

Abstract

This invention discloses a design method for an overvoltage simulation model of composite towers, belonging to the interdisciplinary field of electromagnetic transient simulation and artificial intelligence for transmission line towers. The method includes constructing a multi-wave impedance model and a grounding resistance model and determining the parameters to be optimized; generating training samples composed of lightning current characteristics and tower structural characteristics; constructing a deep neural network surrogate model based on dual-current multi-head self-attention and residual connections, extracting deep representations of the two types of heterogeneous features and capturing their interaction relationships; constructing a near-end policy optimization reinforcement learning framework, and training the surrogate model offline with the goal of minimizing the overvoltage peak deviation; and using the trained surrogate model for adaptive parameter prediction. This invention overcomes the shortcomings of traditional models in representing the frequency-varying characteristics of composite materials, significantly improving the accuracy and efficiency of overvoltage simulation.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic transient simulation and artificial intelligence interdisciplinary technology of transmission line towers, specifically to a design method for a composite tower overvoltage simulation model. Background Technology

[0002] Composite towers are increasingly used in power transmission line engineering due to their excellent external insulation performance, corrosion resistance, and lightweight properties. Accurate simulation calculation of lightning overvoltage is a key technical step in the design and operation and maintenance of composite towers, crucial for evaluating their lightning withstand level and optimizing insulation coordination schemes. However, the dielectric properties of composite materials cause significant frequency-varying effects and nonlinear characteristics in the propagation of lightning waves within the tower. When traditional multi-wave impedance models based on metal towers are directly applied to composite towers, significant problems arise, including insufficient simulation accuracy and difficulty in calibrating model parameters.

[0003] In existing technologies, the model design methods for overvoltage simulation of composite towers typically employ a fixed-parameter equivalent circuit modeling approach. This method establishes separate models for the tower's multi-wave impedance, grounding down conductor distributed parameters, transmission line, and grounding resistance based on the composite tower's structural dimensions and material properties. These sub-models are then combined according to actual electrical connections to form the final composite tower overvoltage simulation model. This approach suffers from the following technical drawbacks: First, the tower's wave impedance parameters heavily rely on empirical formulas, which fail to adequately consider the dynamic influence of the composite material's frequency-varying dielectric constant and permeability on the wave impedance parameters, leading to significant deviations between the wave impedance parameters under different lightning current waveforms and actual conditions. Second, the grounding resistance model generally uses a simplified impulse coefficient correction method, which fails to accurately characterize the complex coupling relationship between lightning current amplitude, waveform steepness, and the nonlinear ionization process of the soil, making it difficult to precisely describe the dynamic changes in grounding resistance. Third, the parameter interactions between the sub-models are fixed, making the entire simulation model unadjustable once established, lacking adaptive adjustment capabilities for different lightning current conditions and tower structural characteristics. The aforementioned defects together result in a large deviation between the simulation results and the measured values ​​of overvoltage in existing models when the amplitude of lightning current varies greatly or the tower structure differs significantly. This makes it impossible to provide a reliable technical basis for the refined lightning protection design of composite towers.

[0004] In recent years, deep learning technology has demonstrated powerful capabilities in complex nonlinear mapping in the field of power system transient simulation, particularly its attention mechanism, which shows significant advantages in capturing long-range dependencies between input features. Deep reinforcement learning, on the other hand, has shown great potential in adaptive parameter optimization. Currently, existing technologies have not systematically combined deep neural networks enhanced by multi-head self-attention mechanisms with deep reinforcement learning methods and have not yet provided publicly available technical solutions for adaptive parameter design of composite tower overvoltage simulation models.

[0005] To address the aforementioned issues, there is an urgent need for a design method for a composite tower overvoltage simulation model to solve the problems associated with traditional methods. Summary of the Invention

[0006] The purpose of this invention is to provide a design method for a composite tower overvoltage simulation model, which overcomes the shortcomings of traditional fixed parameter models in characterizing the frequency-varying characteristics of composite materials and significantly improves the accuracy and efficiency of composite tower overvoltage simulation.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for designing a simulation model of overvoltage in a composite tower includes: Step 1: Construct a multi-wave impedance model and a grounding resistance model for the composite tower, and determine the set of parameters to be optimized; Step 2: Generate a training sample dataset, where each training sample contains an input feature vector and a target output vector; Step 3: Construct a deep neural network proxy model based on dual-stream multi-head self-attention and residual connection. The input of the deep neural network proxy model is the input feature vector, and the output is the predicted value vector of the parameter to be optimized. The deep neural network proxy model consists of a lightning current condition feature extraction sub-network, a tower structure feature extraction sub-network, a multi-head self-attention fusion module, and a regression sub-network. Step 4: Construct a near-end policy optimization reinforcement learning framework, define the state space as the input feature vector, define the action space as the predicted value vector of the parameters to be optimized, define the reward function as the negative mean square error based on the difference between the simulated overvoltage peak and the target overvoltage peak, and use the near-end policy optimization algorithm to train the network weights of the deep neural network surrogate model offline. Step 5: Deploy the trained deep neural network surrogate model into the overvoltage simulation process. For any given lightning current condition and tower structure, the deep neural network surrogate model outputs the corresponding predicted values ​​of wave impedance parameters and grounding resistance parameters. Substitute the predicted values ​​into the composite tower multi-wave impedance model and grounding resistance model to complete the adaptive construction of the composite tower overvoltage simulation model.

[0008] Furthermore, in step 1, a multi-wave impedance model and a grounding resistance model for the composite tower are constructed, and the set of parameters to be optimized is determined, specifically as follows: The composite tower is divided into multiple segments along the height direction, and the wave impedance parameters of each segment and the wave impedance parameters of the tower crossarm are obtained. The two constitute the parameters to be optimized for the multi-wave impedance model of the composite tower. The grounding resistance model is a nonlinear impulse grounding resistance model that considers the soil ionization effect. The impulse grounding resistance parameters are obtained and used as the parameters to be optimized in the grounding resistance model. The parameters to be optimized in the composite tower multi-wave impedance model and the grounding resistance model are combined to obtain the set of parameters to be optimized.

[0009] Furthermore, in step 2, the input feature vector is formed by splicing the lightning current condition feature sub-vector and the tower structure feature sub-vector, and the target output vector is a vector composed of the optimized target values ​​of the segmented wave impedance parameters, the crossarm wave impedance parameters, and the impulse grounding resistance parameters in the grounding resistance model.

[0010] Furthermore, the lightning current condition characteristic sub-vector includes lightning current amplitude, lightning current wavefront time, lightning current half-peak time, lightning current steepness, and lightning strike point location parameters; the tower structure characteristic sub-vector includes tower nominal height, tower segment height, tower segment equivalent radius, composite material relative permittivity, composite material magnetic permeability, soil resistivity, and grounding device geometric dimensions.

[0011] Further, in step 3, the lightning current condition feature extraction sub-network is used to extract deep features of lightning current conditions. The lightning current condition feature extraction sub-network takes the lightning current condition feature sub-vector as input and outputs the deep feature vector of lightning current conditions. The tower structure feature extraction sub-network is used to extract deep features of the tower structure. The tower structure feature extraction sub-network takes the tower structure feature sub-vector as input and outputs the tower structure deep feature vector.

[0012] Further, in step 3, the multi-head self-attention fusion module stacks the deep feature vectors of lightning current conditions and tower structures into a joint feature matrix. Multiple parallel attention heads linearly project the joint feature matrix to obtain the query matrix, key matrix, and value matrix of each head. The output of each attention head is calculated using a scaling dot product attention mechanism. The outputs of all attention heads are concatenated along the feature dimension and linearly transformed to obtain the initial output of the multi-head self-attention. The initial output is residually connected to the joint feature matrix and layer normalization is applied to obtain the enhanced feature matrix. The row vectors corresponding to the lightning current conditions and the row vectors corresponding to the tower structures in the enhanced feature matrix are concatenated along the feature dimension to form a fused feature vector.

[0013] Furthermore, in step 3, the regression subnetwork takes the fused feature vector as input, transforms it layer by layer through multiple fully connected hidden layers, and outputs the predicted value vector of the parameter to be optimized, wherein the dimension of the predicted value vector of the parameter to be optimized is the same as the dimension of the vectorized representation of the set of parameters to be optimized.

[0014] Furthermore, in step 4, the near-end strategy optimization algorithm specifically includes: The policy network is based on a deep neural network agent model, and uses the output of the deep neural network agent model as the mean of the action distribution, superimposed with Gaussian noise with a preset variance to form a stochastic policy; its value network is a neural network that takes the state vector as input and outputs the estimated value of the expected cumulative reward; during training, the policy network weights are updated by maximizing the truncated objective function, and the value network weights are updated by minimizing the mean square error between the value network output and the discounted cumulative reward objective value.

[0015] In summary, the present invention has at least one of the following beneficial technical effects: First, this invention introduces a dual-flow multi-head self-attention deep neural network and a near-end policy optimization reinforcement learning framework into the parameter design process of a composite tower overvoltage simulation model. It constructs an end-to-end technical framework from heterogeneous feature extraction and feature interaction fusion to adaptive optimization of model parameters, overcoming the shortcomings of traditional empirical formulas in characterizing the frequency-varying characteristics of composite materials.

[0016] Second, by constructing a lightning current condition feature extraction subnetwork and a tower structure feature extraction subnetwork, the present invention forms a dual-current feature extraction structure, which can deeply mine and characterize two heterogeneous features with significantly different physical meanings: lightning current condition and tower structure, effectively preserving the physical attribute information of each feature.

[0017] Third, by introducing a multi-head self-attention fusion module, this invention can adaptively learn the interaction weights between deep features of lightning current conditions and deep features of tower structures, effectively capturing the fine-grained coupling relationships within and between the two types of features. Furthermore, by using residual connections and layer normalization operations, the stability and convergence speed of deep network training are improved, significantly enhancing the surrogate model's ability to fit the complex nonlinear mapping relationship between input features and parameters to be optimized.

[0018] Fourth, this invention employs a near-end strategy optimization algorithm to train a deep neural network surrogate model offline. The negative mean square error of the difference between the simulated overvoltage peak and the target overvoltage peak is used as the optimization objective. This enables the adaptive adjustment of model parameters according to lightning current conditions and tower structural characteristics, significantly improving the accuracy of overvoltage simulation for composite towers.

[0019] Fifth, the deep neural network proxy model of this invention can be directly deployed in the overvoltage simulation process after training. For any given lightning current condition and tower structure, it can quickly output the predicted values ​​of the corresponding wave impedance parameters and grounding resistance parameters through a single forward calculation, without having to repeatedly call the high-precision electromagnetic transient simulation model for parameter calibration, which greatly improves the design and construction efficiency of the composite tower overvoltage simulation model. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] like Figure 1 As shown, this invention provides a method for designing a simulation model of overvoltage for composite towers, including: Step 1: Construct a multi-wave impedance model and a grounding resistance model for the composite tower, and determine the set of parameters to be optimized; Step 2: Generate a training sample dataset, where each training sample contains an input feature vector and a target output vector; Step 3: Construct a deep neural network proxy model based on dual-stream multi-head self-attention and residual connection. The input of the deep neural network proxy model is the input feature vector, and the output is the predicted value vector of the parameter to be optimized. The deep neural network proxy model consists of a lightning current condition feature extraction sub-network, a tower structure feature extraction sub-network, a multi-head self-attention fusion module, and a regression sub-network. Step 4: Construct a near-end policy optimization reinforcement learning framework, define the state space as the input feature vector, define the action space as the predicted value vector of the parameters to be optimized, define the reward function as the negative mean square error based on the difference between the simulated overvoltage peak and the target overvoltage peak, and use the near-end policy optimization algorithm to train the network weights of the deep neural network surrogate model offline. Step 5: Deploy the trained deep neural network surrogate model into the overvoltage simulation process. For any given lightning current condition and tower structure, the deep neural network surrogate model outputs the corresponding predicted values ​​of wave impedance parameters and grounding resistance parameters. Substitute the predicted values ​​into the composite tower multi-wave impedance model and grounding resistance model to complete the adaptive construction of the composite tower overvoltage simulation model.

[0023] Next, the above steps will be explained in detail with reference to some specific implementation examples and parameter settings.

[0024] Step 1 involves constructing a multi-wave impedance model and a grounding resistance model for the composite tower, and determining the set of parameters to be optimized. This includes the following steps: Step 101: Establish the multi-wave impedance model of the composite tower Composite towers are divided along the height direction into: The segment, the first The wave impedance parameters of each segment are denoted as... In the multi-wave impedance model of a composite tower, the wave impedance parameters of each segment constitute the first part of the parameter vector to be optimized, denoted as... The wave impedance parameter of the tower crossarm is denoted as... The number of crossarms and the crossarm wave impedance parameters constitute the second part of the parameter vector to be optimized, denoted as . .

[0025] Step 102: Establish a grounding resistance model.

[0026] The grounding resistance model adopts a nonlinear impulse grounding resistance model that considers the soil ionization effect, and the impulse grounding resistance parameters are denoted as... , This constitutes the third part of the parameter vector to be optimized.

[0027] Step 103: Determine the set of parameters to be optimized.

[0028] The above parameters are combined to form a set of parameters to be optimized. Its vectorized representation is the parameter vector to be optimized. The total number of parameters to be optimized is .

[0029] Step 2 involves generating a training sample dataset, where each training sample contains an input feature vector and a target output vector. This process includes the following steps: Step 201: Construct the input feature vector.

[0030] Input feature vector Characteristic subvectors of lightning current operating conditions and tower structure feature vectors splicing together, that is Dimension of the input feature vector .

[0031] Lightning current condition characteristic subvector Including the following One characteristic component: lightning current amplitude The unit is Lightning current wavefront time The unit is Half-peak time of lightning current The unit is Lightning current steepness The unit is Lightning strike location parameters , dimensionless.

[0032] Tower structure feature sub-vectors Including the following One characteristic component: nominal tower height The unit is Height of each section of the tower , The unit is Equivalent radius of each segment of the tower The unit is Relative permittivity of composite materials Dimensionless; magnetic permeability of composite materials Dimensionless; soil resistivity The unit is Geometric dimensional parameters of grounding device The unit is .

[0033] Step 202: Generate sample data.

[0034] The Latin hypercube sampling method is used to generate features in the lightning current condition feature space and the tower structure feature space. For the sampling point, for the first sampling point The input feature vector corresponding to each sampling point The corresponding overvoltage waveforms are obtained using high-precision electromagnetic transient simulation software or field measured data. With the goal of minimizing the overvoltage peak deviation between the high-precision simulation model and the simplified multi-wave impedance model, the target values ​​of the corresponding parameter vector to be optimized are calibrated using numerical optimization methods. .

[0035] The first Each sample is denoted as ,all The training sample dataset consists of 10 samples. .

[0036] Step 3 involves constructing a deep neural network surrogate model based on two-stream multi-head self-attention and residual connections, specifically including the following steps: Step 301: Design a two-stream feature extraction subnetwork.

[0037] Considering the input feature vector Characteristic subvectors of lightning current conditions with significantly different physical meanings With tower structure feature vectors To enable the surrogate model to extract deep representations of the two types of features separately, this invention constructs a dual-stream feature extraction sub-network.

[0038] The lightning current condition feature extraction subnetwork is composed of It consists of several fully connected hidden layers, whose input is the input feature vector. Parsed from The output is a deep feature vector of lightning current conditions. Its recursive relation is: ; ; ; in, Here is the weight matrix of the first hidden layer. This is the bias vector of the first hidden layer. For the first The weight matrix of each hidden layer For the first The bias vectors of each hidden layer To modify the activation function of the linear unit, i.e. The number of neurons in the last hidden layer of the subnetwork is... .

[0039] The tower structure feature extraction subnetwork consists of It consists of several fully connected hidden layers, whose input is the input feature vector. Parsed from The output is the deep feature vector of the tower structure. Its recursive relation is: ; ; ; in, Here is the weight matrix of the first hidden layer. This is the bias vector of the first hidden layer. For the first The weight matrix of each hidden layer For the first The bias vectors of each hidden layer. The number of neurons in the last hidden layer of the subnetwork is also [number]. .

[0040] The deep feature vectors output by the two feature extraction sub-networks have the same dimension. This is to facilitate subsequent feature fusion.

[0041] Step 302: Construct a multi-head self-attention feature fusion module.

[0042] In order to effectively integrate the deep characteristics of lightning current conditions Deep features of tower structure Furthermore, this invention introduces a multi-head self-attention mechanism after the dual-stream feature extraction subnetwork to capture fine-grained interactions within and between two types of features.

[0043] First, and Stacking features along the feature dimension forms a joint feature matrix. ,Right now: ; Joint feature matrix Considered as including A sequence of tags, each tag having a dimension of . .

[0044] Multi-head self-attention module includes A parallel attention head. For the first... One's attention, Through linear transformation Mapped to query matrix Key matrix Sum matrix : ; in, For a trainable projection matrix, The feature dimensions for each attention head.

[0045] No. The output of each attention head Calculated by scaled dot product attention: ; The outputs of all attention heads are concatenated along the feature dimension to obtain... And through a trainable projection matrix Perform a linear transformation to obtain the initial output of the multi-head self-attention. : ; To enhance the training stability and expressive power of the model, residual connections and layer normalization operations are introduced into the multi-head self-attention module. Let the feature matrix after residual connections and layer normalization be... The calculation formula is as follows: ; in, The representation layer normalization operation is applied to the feature dimension of each label.

[0046] Will Take out the row vectors respectively, and let the first row vector be... This corresponds to the enhanced lightning current operating condition characteristics; the second row vector is... This corresponds to the enhanced tower structure characteristics.

[0047] Will and The features are concatenated along the feature dimension to form a fused feature vector. : ; Step 303: Construct the regression subnetwork.

[0048] fuse feature vectors Input to In the regression subnetwork consisting of fully connected hidden layers, the final output is a vector of predicted values ​​for the parameters to be optimized. .

[0049] The recursive relationship of the regression subnetwork is as follows: ; ; ; in, Here is the weight matrix of the first hidden layer. This is the bias vector of the first hidden layer. For the first The weight matrix of each hidden layer For the first The bias vectors of each hidden layer This is the output layer weight matrix. This is the output layer bias vector. The output of the last hidden layer of the regression subnetwork. Without going through an activation function, it is directly linearly mapped to the final predicted value.

[0050] Step 304: Define the overall parameter set of the deep neural network surrogate model.

[0051] All trainable weights and biases in the above-mentioned lightning current condition feature extraction subnetwork, tower structure feature extraction subnetwork, multi-head self-attention module, and regression subnetwork are collectively denoted as the parameter set. Then, the deep neural network surrogate model can be represented as a function mapping. .

[0052] Step 4: Construct a near-end policy optimization reinforcement learning framework, defining the state space as the input feature vector, the action space as the predicted value vector of the parameters to be optimized, and the reward function as the negative mean square error based on the difference between the simulated overvoltage peak and the target overvoltage peak. The near-end policy optimization algorithm is then used to train the network weights of the deep neural network surrogate model offline, specifically including the following steps: Step 401: Define the state space for reinforcement learning.

[0053] State vectors in reinforcement learning It is directly taken from the input feature vector defined in step 2, that is State space Input feature vector The continuous space consisting of all possible values ​​of .

[0054] Step 402: Define the action space for reinforcement learning.

[0055] Action vectors in reinforcement learning Defined as the vector of predicted values ​​of the parameters to be optimized output by the deep neural network surrogate model in step 3, i.e. Action space Parameters to be optimized The continuous space consisting of all possible values ​​of .

[0056] Step 403: Define the reward function for reinforcement learning.

[0057] reward function Used to evaluate the state Next action The degree of quality is defined as the degree of superiority or inferiority of the action. The peak overvoltage was calculated after substituting the corresponding parameters to be optimized into the composite tower overvoltage simulation model. With the target overvoltage peak The negative mean square error between them, i.e.: ; in, To utilize actions The simulation model constructed with the corresponding parameters is in state The calculated overvoltage peak value under the corresponding operating conditions; The target value of overvoltage peak value is obtained using a high-precision electromagnetic transient simulation model or on-site measured data under the same operating conditions.

[0058] Overvoltage peak The calculation process is as follows: First, from the state The characteristic sub-vectors of lightning current operating conditions are extracted from the data. and tower structure feature vectors Then, from the action Wave impedance parameters were analyzed. , and grounding resistance parameters Substitute the multi-wave impedance model and grounding resistance model established in step 1 into the simulation model, and combine them with the transmission line model to form a complete composite tower overvoltage simulation model; finally, apply an voltage level of 400 kV to the simulation model. The corresponding lightning current excitation was used to obtain the overvoltage waveform at the top of the tower through electromagnetic transient simulation calculations, and its peak value was taken as... .

[0059] Step 404: Construct a training framework for the near-end policy optimization algorithm.

[0060] Define policy network This policy network is the deep neural network surrogate model constructed in step 3. To introduce an exploration mechanism during reinforcement learning training, the policy network is defined as outputting the surrogate model... As the mean, with The variance follows a Gaussian distribution, i.e.: ; in, Indicates a multivariate Gaussian distribution. It is the identity matrix. Explore the standard deviation for the preset action.

[0061] Defining Value Network Value networks are used to estimate the value from the state. The initial expected cumulative reward. The value network employs a feedforward neural network structure similar to the regression subnetwork in step 3, with its input being the state vector. The output is a scalar value. Its trainable parameters are denoted as .

[0062] The training process of the proximal policy optimization algorithm includes the following steps: (1) Utilizing the current policy network It interacts with the environment to generate a batch of trajectory data. Each trajectory data consists of a series of states, actions, and rewards, denoted as... ,in For time step index, This refers to the trajectory length. The specific method of environmental interaction is as follows: from the training sample dataset... Randomly sample an input feature vector as the initial state. Policy network based on Sampling action The environment calculates rewards based on the reward function. Then the state transitions to the next sampling point. Repeat this process until a length of [length missing] is generated. The trajectory.

[0063] (2) For each state-action pair in the trajectory Calculate the advantage function estimate $\hat{A}{t}$, which is used to measure the action. The relative superiority or inferiority of actions compared to the average action. The advantage function estimate is calculated using the generalized advantage estimation method: ; in, As a discount factor, For the smoothing parameters of the generalized dominance estimate, The timing difference error is calculated using the following formula: ; (3) Using the collected trajectory data, the policy network weights are optimized by maximizing the objective function of the near-end policy. Update as needed. The objective function for near-end policy optimization is defined as: ; in, The importance sampling ratio is defined as the ratio of the probability of an action under the current policy to that under the old policy, i.e.: ; For the truncation function, Limited to the range Inside, This is the threshold parameter for truncation.

[0064] (4) Minimize the mean squared error loss function for the weights of the value network To update, the value network loss function is defined as: ; in, From state The initial target value for the accumulated discount reward is calculated using the following formula: ; (5) Repeat steps (1) to (4) until the policy network and value network converge or reach the preset maximum number of iterations.

[0065] Step 405: Complete the training of the deep neural network agent model.

[0066] After training, extract the policy network. Mean output As the final deep neural network proxy model, in which, To train the converged policy network weights.

[0067] In step 5, the trained deep neural network surrogate model is deployed into the overvoltage simulation process. For any given lightning current condition and tower structure, the deep neural network surrogate model outputs the corresponding predicted values ​​of wave impedance parameters and grounding resistance parameters. These predicted values ​​are then substituted into the composite tower multi-wave impedance model and grounding resistance model to complete the adaptive construction of the composite tower overvoltage simulation model. The specific steps include the following: Step 501: Input operating condition characteristics.

[0068] Obtain the lightning current characteristic subvector of the simulated operating condition. and tower structure feature vectors The input feature vector is formed by concatenating the features. .

[0069] Step 502: Predict parameters using the surrogate model.

[0070] Input feature vector Input the deep neural network surrogate model trained in step 4, and output the predicted value vector of the parameters to be optimized. .

[0071] from The wave impedance parameters of each segment of the tower are analyzed. Crossarm impedance parameters and impulse grounding resistance parameters .

[0072] Step 503: Construct a simulation model of overvoltage in composite towers.

[0073] Substitute the predicted parameters into the multi-wave impedance model and grounding resistance model established in step 1, and combine them with the transmission line model according to the actual electrical connection relationship to form a complete composite tower overvoltage simulation model.

[0074] Step 504: Perform overvoltage simulation calculation.

[0075] In the electromagnetic transient simulation environment, the input feature vector is applied to the overvoltage simulation model of the composite tower. The corresponding lightning current excitation is used to calculate the overvoltage waveform and peak value of each key node of the tower, and the overvoltage simulation results are output.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for designing a simulation model of overvoltage in a composite tower, characterized in that, include: Step 1: Construct a multi-wave impedance model and a grounding resistance model for the composite tower, and determine the set of parameters to be optimized; Step 2: Generate a training sample dataset, where each training sample contains an input feature vector and a target output vector; Step 3: Construct a deep neural network proxy model based on dual-stream multi-head self-attention and residual connection. The input of the deep neural network proxy model is the input feature vector, and the output is the predicted value vector of the parameter to be optimized. The deep neural network proxy model consists of a lightning current condition feature extraction sub-network, a tower structure feature extraction sub-network, a multi-head self-attention fusion module, and a regression sub-network. Step 4: Construct a near-end policy optimization reinforcement learning framework, define the state space as the input feature vector, define the action space as the predicted value vector of the parameters to be optimized, define the reward function as the negative mean square error based on the difference between the simulated overvoltage peak and the target overvoltage peak, and use the near-end policy optimization algorithm to train the network weights of the deep neural network surrogate model offline. Step 5: Deploy the trained deep neural network surrogate model into the overvoltage simulation process. For any given lightning current condition and tower structure, the deep neural network surrogate model outputs the corresponding predicted values ​​of wave impedance parameters and grounding resistance parameters. Substitute the predicted values ​​into the composite tower multi-wave impedance model and grounding resistance model to complete the adaptive construction of the composite tower overvoltage simulation model.

2. The design method for a composite tower overvoltage simulation model according to claim 1, characterized in that, In step 1, a multi-wave impedance model and a grounding resistance model for the composite tower are constructed, and the set of parameters to be optimized is determined, specifically: The composite tower is divided into multiple segments along the height direction, and the wave impedance parameters of each segment and the wave impedance parameters of the tower crossarm are obtained. The two constitute the parameters to be optimized for the multi-wave impedance model of the composite tower. The grounding resistance model is a nonlinear impulse grounding resistance model that considers the soil ionization effect. The impulse grounding resistance parameters are obtained and used as the parameters to be optimized in the grounding resistance model. The parameters to be optimized in the composite tower multi-wave impedance model and the grounding resistance model are combined to obtain the set of parameters to be optimized.

3. The design method for a composite tower overvoltage simulation model according to claim 2, characterized in that, In step 2, the input feature vector is formed by splicing the lightning current condition feature sub-vector and the tower structure feature sub-vector, and the target output vector is a vector composed of the optimized target values ​​of the segment wave impedance parameters, the crossarm wave impedance parameters and the impulse grounding resistance parameters in the grounding resistance model.

4. The design method for a composite tower overvoltage simulation model according to claim 3, characterized in that, The lightning current operating condition characteristic sub-vector includes lightning current amplitude, lightning current wavefront time, lightning current half-peak time, lightning current steepness, and lightning strike point location parameters; the tower structure characteristic sub-vector includes tower nominal height, tower segment height, tower segment equivalent radius, composite material relative permittivity, composite material permeability, soil resistivity, and grounding device geometric dimensions.

5. The design method for a composite tower overvoltage simulation model according to claim 4, characterized in that, In step 3, the lightning current condition feature extraction sub-network is used to extract deep features of lightning current condition. The lightning current condition feature extraction sub-network takes the lightning current condition feature sub-vector as input and outputs the deep feature vector of lightning current condition. The tower structure feature extraction sub-network is used to extract deep features of the tower structure. The tower structure feature extraction sub-network takes the tower structure feature sub-vector as input and outputs the tower structure deep feature vector.

6. The design method for a composite tower overvoltage simulation model according to claim 5, characterized in that, In step 3, the multi-head self-attention fusion module stacks the deep feature vectors of lightning current conditions and tower structures into a joint feature matrix. Multiple parallel attention heads linearly project the joint feature matrix to obtain the query matrix, key matrix, and value matrix of each head. The output of each attention head is calculated using a scaling dot product attention mechanism. The outputs of all attention heads are concatenated along the feature dimension and linearly transformed to obtain the initial output of the multi-head self-attention. The initial output is residually connected to the joint feature matrix and layer normalization is applied to obtain the enhanced feature matrix. The row vectors corresponding to the lightning current conditions and the row vectors corresponding to the tower structures in the enhanced feature matrix are concatenated along the feature dimension to form a fused feature vector.

7. The design method for a composite tower overvoltage simulation model according to claim 6, characterized in that, In step 3, the regression subnetwork takes the fused feature vector as input, transforms it layer by layer through multiple fully connected hidden layers, and outputs the predicted value vector of the parameter to be optimized. The dimension of the predicted value vector of the parameter to be optimized is the same as the dimension of the vectorized representation of the set of parameters to be optimized.

8. The design method for a composite tower overvoltage simulation model according to claim 7, characterized in that, In step 4, the near-end strategy optimization algorithm is specifically as follows: The policy network is based on a deep neural network agent model, and uses the output of the deep neural network agent model as the mean of the action distribution, superimposed with Gaussian noise with a preset variance to form a stochastic policy; its value network is a neural network that takes the state vector as input and outputs the estimated value of the expected cumulative reward; during training, the policy network weights are updated by maximizing the truncated objective function, and the value network weights are updated by minimizing the mean square error between the value network output and the discounted cumulative reward objective value.

Images