Method and system for evaluating bearing capacity of power distribution network for large-scale automobile charging load

Abstract

The invention discloses a power distribution network bearing capacity evaluation method and system for large-scale automobile charging loads, and the method comprises the steps: firstly, achieving the efficient expansion of data through combining a variational auto-encoder and a polynomial fitting method, effectively increasing the number of training samples, reducing the risk of over-fitting, and improving the reliability of the power distribution network bearing capacity evaluation. Then, the stability in a complex training environment is improved by adopting a self-adaptive biological influence optimization method, the accuracy of feature extraction and the training efficiency are improved, then, data are carefully captured and reconstructed by adopting a self-encoding neural network with self-adaptive refined features, the information flow is optimized, the dimensionality is reduced, important information is reserved, and the accuracy of feature extraction is improved. Then, an adaptive error compensation mechanism of a kernel extreme learning machine is adopted to adjust an error learning rate and weight update in real time, and the accuracy and robustness of the model are improved; through the above optimization training, the model can fully master the characteristics of the charging load data, and achieves the precise evaluation of the bearing capacity of the power distribution network through the output final category. Therefore, the method is high in accuracy.

Description

Technical Field

[0001] The present invention relates to a means for evaluating the carrying capacity of a distribution network, and belongs to the technical field of electric power safety, and in particular to a method and system for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads. Background Art

[0002] Due to the uncertainty and time variability of electric vehicle charging behavior, the operation of distribution networks has increased in complexity. With the popularization of electric vehicles and the rapid development of charging infrastructure, distribution networks are facing unprecedented load pressure, which puts higher requirements on the stability and reliability of distribution networks. Therefore, how to accurately evaluate the carrying capacity of distribution networks under the influence of large-scale electric vehicle charging is of great significance to ensure the safe operation of the power grid and optimize the allocation of power grid resources.

[0003] When dealing with such problems, existing technologies usually rely on traditional load forecasting models and power grid simulation technologies. Although these methods can provide certain support, they often lack sufficient flexibility and accuracy and are difficult to effectively cope with highly dynamic charging loads. In addition, existing technologies still have certain limitations in data processing and model training. For example, when expanding, extracting, reducing dimensions and classifying data, there are problems such as incomplete data collection, inaccurate feature extraction, and unstable classification performance. These have limited the effect and application scope of distribution network carrying capacity assessment, resulting in low accuracy of the final assessment results. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects and problems existing in the prior art and to provide a method and system for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads with high accuracy.

[0005] To achieve the above objectives, the technical solution of the present invention is: a distribution network carrying capacity assessment method for large-scale automobile charging loads, comprising:

[0006] S1. Obtain the data generated when the electric vehicle interacts with the distribution network during the charging process of the electric vehicle, and annotate it to obtain the original distribution network data training set;

[0007] The data include grid load, voltage, frequency, frequency of electric vehicle use, mileage, charging habits, charging time, charging station location, charging type, and vehicle type;

[0008] The marked categories include normal charging state, high load state, low load state and fault state;

[0009] S2. Based on the composite data expansion method of variational sparse matrix decomposition and polynomial fitting, the original distribution network data training set is expanded, and the data expansion model is trained to obtain the expanded distribution network data training set;

[0010] S3, using a neural network algorithm based on adaptive biological influence optimization to extract features from the expanded distribution network data training set, and training a feature extraction model to obtain a feature distribution network data training set;

[0011] S4. Using an auto-encoding neural network algorithm based on feature adaptive refinement, perform feature dimensionality reduction on the feature distribution network data training set, and train a feature dimensionality reduction model to obtain a dimensionality reduction distribution network data training set;

[0012] S5. Using a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, classifying the reduced-dimensional distribution network data training set, and training the classifier model to obtain the distribution network state classification category;

[0013] S6. Input the new distribution network data set samples into the feature extraction model and the feature dimension reduction model in sequence for feature processing, and then input them into the classifier model for classification to obtain the classification results. According to the classification results, the carrying capacity of the distribution network can be obtained.

[0014] The step S2 specifically includes:

[0015] S21, initialize the parameters of the variational autoencoder; the parameters include the weight and bias of the variational autoencoder; the expression of the initialization is as follows:

[0016]

[0017] in: is the initial weight matrix of the variational autoencoder, is the initial bias vector of the variational autoencoder, is the variance of the initialization of the variational autoencoder;

[0018] S22, input the original distribution network data training set into the variational autoencoder, map the data into the potential feature space through nonlinear transformation, and use nonlinear activation function to extract key features; the expression of the transformation is as follows:

[0019]

[0020] Where: Z c is the encoded potential feature, Res() is based on the steep ReLU activation function, are the weights and biases of the variational autoencoder, X cis the vector of the input original distribution network data training set;

[0021] The steep ReLU activation function uses an adaptive parameter to control the activation threshold, and its calculation expression is as follows:

[0022]

[0023] Where: X is the input of the ReLU activation function, σ is the adaptive weight, and β is the parameter that controls the steepness of the Sigmoid curve;

[0024] S23, using L1 regularization to implement sparsity constraints in the latent feature space, optimizing the key feature representation of the latent feature space, and calculating the total loss function;

[0025] The expression of the sparsity constraint is as follows:

[0026]

[0027] in: is the sparsity loss function, λ c is the sparsity regularization coefficient, z ci is the i-th element in the encoded latent feature;

[0028] The expression of the total loss function is as follows:

[0029]

[0030] in: is the total loss of the variational autoencoder, is the reconstruction loss of the variational autoencoder;

[0031] The sparsity regularization coefficient is set in an adaptive adjustment manner, and its expression is as follows:

[0032]

[0033] Where: γ c is the adjustment coefficient, ρ c is the target sparsity, is the average sparsity of the current iteration;

[0034] S24, performing polynomial fitting on key features in the potential feature space, and dynamically adjusting coefficients of the polynomial according to distribution characteristics of the key features;

[0035] The expression of the polynomial fitting is as follows:

[0036]

[0037] in: is the key feature after polynomial fitting, a cn are the coefficients of the nth degree polynomial, is the power of the potential distribution network data feature, Nf is the total number of samples in the current input batch;

[0038] The ∩th order polynomial coefficient a cn The calculation expression of the optimization process is as follows:

[0039]

[0040] Where: α ck is the basic weight of the polynomial coefficient, Kf is the highest order of the fitting polynomial;

[0041] S25, mapping the key features after polynomial fitting back to the original data space through the decoder to generate new distribution network data points;

[0042] The expression of the new distribution network data point is as follows:

[0043]

[0044] in: is a new distribution network data point, Sig( ) is the Sigmoid activation function, are the weight and bias of the decoder respectively;

[0045] S26. Dynamically adjust the learning rate of the variational autoencoder according to the loss function of the variational autoencoder;

[0046] The expression of the learning rate is as follows:

[0047]

[0048] in: are the learning rates for iterations t+1 and t, respectively, c is the learning rate decay coefficient, is the gradient of total loss with respect to learning rate;

[0049] S27. Repeat steps S21-S26 until the preset stop iteration condition is met, to obtain the expanded distribution network data training set and complete the data expansion model training.

[0050] The step S3 specifically includes:

[0051] S31, initialize the parameters of the neural network, and define the weights and biases of the neural network; the expressions of the weights and biases are as follows:

[0052]

[0053] Where: WNN 、b NN are the weights and biases of the neural network, respectively, ~ means obeying a specific distribution, σ sn is the initial standard deviation, are the initial weights and initial biases of the neural network, is a normal distribution;

[0054] S32. At the beginning of each iteration, the loss function under the current parameter configuration is calculated, and the environmental fitness is estimated, and then resources are allocated based on the contribution of each parameter;

[0055] The expression of the environmental adaptability is as follows:

[0056] E NN =exp(-β NN ·(W NN , b NN ));

[0057] Where: E NN is the environmental fitness, β NN is the sensitivity parameter;

[0058] The loss function is calculated using cross entropy loss, and its calculation expression is as follows:

[0059]

[0060] Where: L NN (W NN , b NN ) is the loss function, C is the number of categories, y c is the one-hot encoding of the true label, is the predicted probability,

[0061] The predicted probability is calculated by the softmax function, and its calculation expression is as follows:

[0062]

[0063] Where: z c is the logistic regression probability output by the neural network, z k is the logistic regression probability of the kth category output by the network;

[0064] The resource calculation expression for each parameter competition allocation is as follows:

[0065]

[0066] Where: R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and 1 are both indexes, γ NN is a hyperparameter, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of

[0067] S33, based on the resource competition results, update the weights and biases, and use the momentum term to smooth the learning process of the weight parameters;

[0068] The updated expression is as follows:

[0069]

[0070] in: is the weight parameter update of the jth neuron in the i-th neural network layer, is the bias parameter update of the i-th layer of neurons, is the bias of the i-th layer neuron, R i The resources allocated for the parameters of the i-th neural network layer, α NN is the learning rate;

[0071] The expression of the momentum term to smooth the learning process of the weight parameters is as follows:

[0072]

[0073] in: is the momentum term of the jth neuron in the i-th neural network layer at the t-th iteration, is the momentum term of the jth neuron in the i-th neural network layer at the t-1th iteration, β xm is the momentum decay parameter, is the weight parameter of the jth neuron in the i-th neural network layer at the t+1th iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-th iteration,

[0074] S34, adjusting the sensitivity parameter and the learning rate parameter according to the result of the environmental perception evaluation; the result of the environmental perception evaluation refers to the difference between the environmental fitness of the current iteration and the previous iteration;

[0075] The expression for updating the learning rate is as follows:

[0076]

[0077] in: is the updated learning rate parameter, is the updated sensitivity parameter, η NN , k NN are adjustment factors, ΔE NN is the difference between the environmental fitness of this iteration and the previous iteration;

[0078] S35. Based on the result of the environment perception evaluation, dynamically adjust the hyperparameters controlling the resource allocation sensitivity between the global search and the local search; the expression of the adjustment is as follows:

[0079]

[0080] in: is the hyperparameter for the sensitivity of the adjusted control resource allocation, ρ NN is a balanced hyperparameter;

[0081] S36. Repeat steps S31-S35 until the preset iteration stop condition is reached, obtain the characteristic distribution network data training set and complete the feature extraction model training.

[0082] The step S4 specifically includes:

[0083] S41, initialize the network parameters of the autoencoder, the network includes the weight and bias of the autoencoder; the expression is as follows:

[0084]

[0085] in: is the initial weight of the autoencoder, W r is the weight parameter of the autoencoder, The mean of the table is 0 and the variance is The normal distribution of is the number of neurons in the input layer of the autoencoder, is the initial bias of the autoencoder, b r is the bias parameter of the autoencoder;

[0086] S42, propagating the distribution network data after feature extraction through the encoder layer sequence and converting it into a low-dimensional representation; the expression of the low-dimensional representation is as follows:

[0087]

[0088] Where: H r is the low-dimensional representation of the encoder output, Re() is the ReLU activation function, are the weight and bias of the encoder respectively, is the reconstructed output of the distribution network data, are the weight and bias of the decoder respectively;

[0089] S43. In the hidden layer of the autoencoder, the weights are adjusted by a dynamic weight adjustment layer, and the gradient of the loss function with respect to the weights is solved by the chain rule;

[0090] The expression of the adjustment weight is as follows:

[0091]

[0092] Where: ΔW r is the weight update amount, tanh() is the hyperbolic tangent function, λ r is the learning rate, γ r is the adjustment factor, is the loss function L r About weight W r The gradient of L r is the reconstruction error of the loss function,

[0093] The chain rule solves the gradient of the loss function with respect to the weight as follows:

[0094]

[0095] in: is the partial derivative of the loss function with respect to the reconstruction error output;

[0096] The calculation expression of the mean square error of the loss function is as follows:

[0097]

[0098] Among them: |||| is the L2 norm, α r is the regularization coefficient, is the L1 norm of the encoder weight;

[0099] S44. Based on the adjusted weights, the weight parameters and bias parameters of the autoencoder are updated using the gradient descent method; the expressions of the weight parameters and bias parameters are as follows:

[0100]

[0101] in: is the weight parameter of the autoencoder at the t+1th iteration, is the weight parameter of the autoencoder at the tth iteration, η r is the learning rate, is the bias parameter of the autoencoder at the t+1th iteration, is the bias parameter of the autoencoder at the tth iteration, t is the number of iterations, are the gradients of the loss function with respect to weights and biases respectively;

[0102] The learning rate is set by adaptive adjustment, and its expression is as follows:

[0103]

[0104] in: is the learning rate for the t+1th iteration, is the learning rate of the tth iteration, δ is the sensitivity parameter for adjusting the learning rate;

[0105] S45. Repeat steps S41-S44 until the iteration stop condition is met, obtain the reduced-dimensional distribution network data training set and complete the feature-dimensionality reduction model training.

[0106] The step S5 specifically includes:

[0107] S51, initializing the parameters of the kernel function and the weight and bias of the network, and dynamically adjusting the width parameter of the kernel function according to the error distribution of each iteration;

[0108] The initialization expression is as follows:

[0109] K u (x u , x′ u )=exp(-γ u ||x u -x′ u || 2 );

[0110]

[0111] Where: K u is the kernel function, x u is the distribution network data point, x′ u To distinguish from x u Another distribution network data point, γ u is the width parameter of the kernel function, are the initialized weights and biases, is the initialization variance;

[0112] The expression for the width parameter is as follows:

[0113]

[0114] Where: γ u (t+1) is the width parameter of the kernel function at the t+1th iteration, γ u (t) is the width parameter of the kernel function at the tth iteration, ||e u(t)||2 is the L2 norm of the error of the current iteration, ||x u (t)||2 is the L2 norm of the reduced-dimensional distribution network data training set, ∈ ug To prevent division by zero, η γ The learning rate adjusted for the kernel parameters;

[0115] S52. Before each training iteration begins, calculate the current error compensation factor according to the error of the previous iteration; the expression of the error compensation factor is as follows:

[0116]

[0117] Where: u (t) is the error compensation factor, || || is the L2 norm, λ u is the adjustment coefficient of error compensation, e u (t-1) is the error vector of the t-1th iteration, e u (t-2) is the error vector of the t-2th iteration, e u (t) is the error vector of the tth iteration, y u is the true label vector, is the prediction for the tth iteration;

[0118] S53, dynamically adjust the weight of the kernel extreme learning machine based on the error compensation factor and kernel technology;

[0119] The expression of the weight of the kernel extreme learning machine is as follows:

[0120]

[0121] in: is the weight of the kernel extreme learning machine at the t+1th iteration, is the weight of the kernel extreme learning machine at the tth iteration, η u is the learning rate of the kernel extreme learning machine, is the gradient of the loss function of the kernel extreme learning machine with respect to the weight, L u ( ) is the loss function of the kernel extreme learning machine, nE is the number of samples in the current batch, and I is the unit matrix;

[0122] The expression of the loss function of the kernel extreme learning machine is as follows:

[0123]

[0124] in: is the predicted value calculated by kernel technique and weights, is the true label, λ fg is the regularization parameter;

[0125] S54, after the weight adjustment is completed, check the gradient value and determine whether the gradient value exceeds the stage threshold;

[0126] If the gradient value exceeds the preset cutoff threshold, Then perform truncation processing; the expression of the truncation processing is as follows:

[0127]

[0128] Where: θ u is the gradient cutoff threshold;

[0129] S55. Repeat steps S51-S54 until the iteration stop condition is met, the distribution network status classification category is obtained and the classifier model training is completed.

[0130] The classification result includes any one of a normal charging state, a high load state, a low load state and a fault state.

[0131] According to the classification results, the carrying capacity of the distribution network can be obtained as follows:

[0132] If the classification result is a normal charging state, it means that the current carrying capacity of the distribution network is normal, which means that the electric vehicle is charged under standard conditions and the impact on the distribution network is within an acceptable range;

[0133] If the classification result is a high load state, it means that the current carrying capacity of the distribution network is close to the upper limit, which means that during peak power consumption hours or when multiple electric vehicles are charging at the same time, the distribution network needs to handle a larger load;

[0134] If the classification result is a low load state, it means that the current carrying capacity of the distribution network still has redundant capacity, which means that there are fewer charging electric vehicles or the charging demand is low, and the distribution network can bear the additional load;

[0135] If the classification result is a fault state, it means that the distribution network does not have the carrying capacity, which means that the distribution network or the charging facility has a fault due to overload or other abnormal conditions.

[0136] A distribution network carrying capacity assessment system for large-scale automobile charging loads, the system is applied to the above method, and the system includes:

[0137] The training set acquisition module is used to acquire the data generated when the electric vehicle interacts with the distribution network during the charging process of the electric vehicle, and to annotate the data to obtain the original distribution network data training set;

[0138] A data expansion module is used to expand the original distribution network data training set based on a composite data expansion method of variational sparse matrix decomposition and polynomial fitting, and to train a data expansion model to obtain an expanded distribution network data training set;

[0139] A feature extraction module is used to extract features from the expanded distribution network data training set using a neural network algorithm based on adaptive biological influence optimization, and to train a feature extraction model to obtain a feature distribution network data training set;

[0140] A feature dimension reduction module is used to perform feature dimension reduction on a feature distribution network data training set by using an autoencoding neural network algorithm based on feature adaptive refinement, and to train a feature dimension reduction model to obtain a dimension-reduced distribution network data training set;

[0141] A data classification module is used to classify the reduced-dimensional distribution network data training set using a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, and to train a classifier model to obtain a distribution network state classification category;

[0142] The carrying capacity assessment module is used to input the new distribution network data set samples into the feature extraction model and the feature dimension reduction model in sequence for feature processing, and then input them into the classifier model for classification to obtain the classification results. According to the classification results, the carrying capacity of the distribution network can be obtained.

[0143] A distribution network carrying capacity assessment device for large-scale automobile charging loads, the device comprising a processor and a memory;

[0144] The memory is used to store computer program code and transmit the computer program code to the processor;

[0145] The processor is used to execute the above-mentioned distribution network carrying capacity assessment method for large-scale automobile charging loads according to the instructions in the computer program code.

[0146] A storage medium for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads stores a computer program, which, when executed by a processor, implements the above-mentioned method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads.

[0147] Compared with the prior art, the present invention has the following beneficial effects:

[0148] The present invention discloses a distribution network carrying capacity assessment method and system for large-scale automobile charging loads. The method first obtains and annotates interactive data of the charging process to obtain a training set; then, a composite data expansion method based on variational sparse matrix decomposition and polynomial fitting is used to expand the data; then, a neural network algorithm based on adaptive biological influence optimization is used to extract features; then, an autoencoder neural network algorithm based on feature adaptive refinement is used to reduce the feature dimension; then, a kernel extreme learning machine algorithm based on adaptive error compensation is used as a classifier for classification to obtain a distribution network state classification category; finally, a new distribution network data set sample is used as input for classification, and the carrying capacity of the distribution network can be obtained according to the classification result; in the application of the present invention, first, by combining variational autoencoders and polynomials, the distribution network state classification category is obtained. The term fitting method realizes the efficient expansion of data, effectively increases the number of training samples, and reduces the risk of overfitting. Then, the adaptive biological impact optimization method is used to improve the stability of the model in a complex training environment, improve the accuracy of feature extraction and training efficiency, and then the feature adaptive refinement autoencoder neural network is used to finely capture and reconstruct data, optimize information flow, reduce dimensions while retaining important information, and then the adaptive error compensation mechanism of the kernel extreme learning machine is used to adjust the error learning rate and weight update in real time, effectively improving the accuracy and robustness of the classification model; through the above optimization training of the model, the model can fully grasp the characteristics in the charging load data, and through the final category of the output, the carrying capacity of the distribution network is accurately evaluated. Therefore, the accuracy of the present invention is relatively high. BRIEF DESCRIPTION OF THE DRAWINGS

[0149] Figure 1 It is a flow chart of the method steps of the present invention.

[0150] Figure 2 It is a training flow chart of feature extraction of the present invention.

[0151] Figure 3 It is a training flow chart of feature dimensionality reduction of the present invention.

[0152] Figure 4 It is a schematic diagram of the system structure of the present invention.

[0153] Figure 5 It is a schematic diagram of the device structure of the present invention.

[0154] In the figure: training set acquisition module 1, data expansion module 2, feature extraction module 3, feature dimension reduction module 4, data classification module 5, carrying capacity assessment module 6, processor 7, memory 8, computer program code 81. DETAILED DESCRIPTION

[0155] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0156] Embodiment 1:

[0157] See also Figure 1-Figure 3 , a distribution network carrying capacity assessment method for large-scale vehicle charging loads, including:

[0158] S1. Obtain the data generated when the electric vehicle interacts with the distribution network during the charging process of the electric vehicle, and annotate it to obtain the original distribution network data training set;

[0159] Furthermore, in this embodiment, the multi-source heterogeneous data of the electric vehicle and the distribution network operation during the electric vehicle charging process are collected, and the data can be collected through the real-time monitoring system of the distribution network and the electric vehicle; the data generated when the electric vehicle and the distribution network interact in operation include but are not limited to: grid load, voltage, frequency, electric vehicle usage frequency, mileage, charging habits, charging time, charging station location, charging type, vehicle type, etc. The collected data are counted and stored in a structured JSON format, that is, formed into the original distribution network data training set. In the actual collection process, the types or attributes of the data generated during the interaction may reach dozens or even hundreds, and only a few representative ones are listed in this embodiment.

[0160] After the data is collected, it needs to be manually labeled. The labeled categories include normal charging state, high load state, low load state and fault state. These categories are the "real" labels before model training and are used to guide the model to learn how to identify and classify data.

[0161] During the charging process of electric vehicles, data exchange is achieved between electric vehicles, charging piles and distribution networks with the help of communication protocols. When an electric vehicle has a charging demand, it will send a charging request to the charging pile. After receiving the request, the charging pile will first perform a protocol matching operation to ensure the consistency of the communication protocol between the two parties so that data can be accurately exchanged. At the same time, the charging pile will also carry out authentication work to verify the identity of the electric vehicle and other related information to ensure the safety and compliance of the charging process. After completing these preliminary steps, the charging pile will send the charging request of the electric vehicle to the distribution network. After receiving the request, the distribution network will provide electric energy to the electric vehicle based on its current carrying capacity, including power capacity, line load and other factors; and the distribution network has a real-time coordination function, which will dynamically adjust the charging status of each electric vehicle according to the operating status of the entire distribution network system, such as the overall load situation, the stability of power supply, etc., and can coordinate and manage all charging electric vehicles within a large coverage area.

[0162] S2. Based on the composite data expansion method of variational sparse matrix decomposition and polynomial fitting, the original distribution network data training set is expanded, and the data expansion model is trained to obtain the expanded distribution network data training set;

[0163] In this embodiment, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect the accuracy of the model. Therefore, this solution adopts a composite data distribution network data expansion method based on variational sparse matrix decomposition and polynomial fitting, uses the framework of variational autoencoders, maps data distribution network data to latent space, and increases the expressiveness of data distribution network data by adopting sparsity constraints, thereby capturing the core features of data distribution network data.

[0164] Furthermore, the step S2 specifically includes:

[0165] S21, initialize the parameters of the variational autoencoder; the parameters include the weight and bias of the variational autoencoder; the expression of the initialization is as follows:

[0166]

[0167] in: is the initial weight matrix of the variational autoencoder, is the initial bias vector of the variational autoencoder, is the variance of the initialization of the variational autoencoder;

[0168] S22, input the original distribution network data training set into the variational autoencoder, map the data into the potential feature space through nonlinear transformation, and use nonlinear activation function to extract key features; the expression of the transformation is as follows:

[0169]

[0170] Where: z c is the encoded potential feature, Res( ) is based on the steep ReLU activation function, are the weights and biases of the variational autoencoder, X c is the vector of the input original distribution network data training set; the vector includes attributes such as grid load, voltage, frequency, electric vehicle usage frequency, mileage, charging habits, charging time, charging station location, charging type, and vehicle type.

[0171] The steep ReLU activation function uses an adaptive parameter to control the activation threshold, and its calculation expression is as follows:

[0172]

[0173] Where: X is the input of the ReLU activation function, σ is the adaptive weight (preferably set to 0.3), and β is the parameter that controls the steepness of the Sigmoid curve (preferably set to 3);

[0174] S23, using L1 regularization to implement sparsity constraints in the latent feature space, optimizing the key feature representation of the latent feature space, and calculating the total loss function;

[0175] In this scheme, sparsity constraints are used in the latent space to force the activity of some latent variables to be reduced, so as to achieve the purpose of extracting key features.

[0176] The expression of the sparsity constraint is as follows:

[0177]

[0178] in: is the sparsity loss function, λ c is the sparsity regularization coefficient, z ci is the i-th element in the encoded latent feature;

[0179] The expression of the total loss function is as follows:

[0180]

[0181] in: is the total loss of the variational autoencoder, is the reconstruction loss of the variational autoencoder;

[0182] The sparsity regularization coefficient is set in an adaptive adjustment manner, and changes dynamically according to the sparsity of the data distribution network data to optimize the training effect (preferably set to 0.3). Its expression is as follows:

[0183]

[0184] Where: γ c is the adjustment coefficient (preferably set to 2), ρ c is the target sparsity, is the average sparsity of the current iteration, where the sparsity is represented by the number of connection weights between neurons;

[0185] S24, performing polynomial fitting on key features in the latent feature space, and dynamically adjusting coefficients of the polynomial according to the distribution characteristics of the key features, so as to enhance the naturalness and coherence of data expansion of the data distribution network;

[0186] The expression of the polynomial fitting is as follows:

[0187]

[0188] in: is the key feature after polynomial fitting, a cn are the coefficients of the nth-order polynomial, is the nth power of the potential distribution network data feature, Nf is the total number of samples in the current input batch;

[0189] The nth order polynomial coefficient a cn The calculation expression of the optimization process is as follows:

[0190]

[0191] Where: α ck is the basic weight of the polynomial coefficient, Kf is the highest order of the fitting polynomial;

[0192] S25, mapping the key features after polynomial fitting back to the original data space through the decoder to generate new distribution network data points;

[0193] The expression of the new distribution network data point is as follows:

[0194]

[0195] in: is a new distribution network data point, Sig() is the Sigmoid activation function, are the weight and bias of the decoder respectively;

[0196] S26. Dynamically adjust the learning rate of the variational autoencoder according to the loss function of the variational autoencoder;

[0197] In this scheme, the learning rate of the variational autoencoder is dynamically adjusted according to the error feedback of the decoding result. The learning rate of the variational autoencoder is used to control the step size when adjusting the parameters of the variational autoencoder by gradient descent, so as to optimize the training process and improve the accuracy of data distribution network data generation; (preferably set to 0.3)

[0198] The expression of the learning rate is as follows:

[0199]

[0200] in: are the learning rates for iterations t+1 and t, respectively, c is the learning rate decay coefficient (preferably set to 0.95), is the gradient of total loss with respect to learning rate;

[0201] S27. Repeat steps S21-S26 until the preset stop iteration condition is met, to obtain the expanded distribution network data training set and complete the data expansion model training.

[0202] In this embodiment, the preset condition for stopping iteration is: reaching a preset maximum number of iterations; the preset maximum number of iterations is set to 1000 times.

[0203] S3, using a neural network algorithm based on adaptive biological influence optimization to extract features from the expanded distribution network data training set, and training a feature extraction model to obtain a feature distribution network data training set;

[0204] The expanded distribution network data training set is input into the distribution network data feature extraction model to train the distribution network data feature extraction model. In this scheme, a three-layer fully connected neural network is used to extract distribution network data features.

[0205] In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient exploding, or falling into local optimal solutions may occur, affecting the stability of training and the performance of the model. The present invention adopts a neural network algorithm based on adaptive biological impact optimization as a distribution network data feature extraction model, inspired by the biological adaptive behavior in nature, especially the ability of organisms to adjust strategies according to resource changes in the environment. The parameters of the neural network are optimized by simulating the survival competition and resource adaptation process of organisms, thereby avoiding the problems of gradient vanishing, gradient exploding, and falling into local optimal solutions, so as to improve the distribution network data feature extraction capability.

[0206] Furthermore, the step S3 specifically includes:

[0207] S31, initialize the parameters of the neural network, and define the weights and biases of the neural network; the expressions of the weights and biases are as follows:

[0208]

[0209] Where: W NN 、b NN are the weights and biases of the neural network, respectively, ~ means obeying a specific distribution, σ sn is the initialization standard deviation (preferably set to 0.01), are the initial weights and initial biases of the neural network, is a normal distribution;

[0210] S32. At the beginning of each iteration, evaluate the performance of the current network parameters in a specific training environment, calculate the loss function under the current parameter configuration, and estimate the environmental fitness as a basis for subsequent adjustments, and then allocate resources based on the contribution of each parameter;

[0211] The expression of the environmental adaptability is as follows:

[0212] E NN =exp(-β NN ·(W NN , b NN ));

[0213] Where: E NN is the environmental fitness, β NN is the sensitivity parameter;

[0214] The loss function is calculated using cross entropy loss, and its calculation expression is as follows:

[0215]

[0216] Where: L NN (W NN , b NN ) is the loss function, C is the number of categories, y c is the one-hot encoding of the true label, is the predicted probability,

[0217] The predicted probability is calculated by the softmax function, and its calculation expression is as follows:

[0218]

[0219] Where: z c is the logistic regression probability output by the neural network, z k is the logistic regression probability of the kth category output by the network;

[0220] In the resource competition behavior in nature, each parameter is regarded as an individual, competing to expand the resources in the distribution network dataset, and the resource allocation is determined by the contribution of the parameter;

[0221] The resource calculation expression for each parameter competition allocation is as follows:

[0222]

[0223] Where: R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the first neuron in the kth neural network layer, k and l are both indexes, γ NN is a hyperparameter, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of

[0224] S33, based on the resource competition results, update the weights and biases, and use the momentum term to smooth the learning process of the weight parameters;

[0225] The updated expression is as follows:

[0226]

[0227] in: is the weight parameter update of the jth neuron in the i-th neural network layer, is the bias parameter update of the i-th layer of neurons, is the bias of the i-th layer neuron, R i The resources allocated for the parameters of the i-th neural network layer, α NN is the learning rate;

[0228] The expression of the momentum term to smooth the learning process of the weight parameters is as follows:

[0229]

[0230] in: is the momentum term of the jth neuron in the i-th neural network layer at the t-th iteration, is the momentum term of the jth neuron in the i-th neural network layer at the t-1th iteration, β xm is the momentum decay parameter, is the weight parameter of the jth neuron in the i-th neural network layer at the t+1th iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-th iteration,

[0231] S34, adjusting the sensitivity parameter and the learning rate parameter according to the result of the environmental perception evaluation; the result of the environmental perception evaluation refers to using the difference between the environmental fitness of the current iteration and the previous iteration as the change of the environmental fitness, so as to achieve the perception ability of environmental changes in the process of parameter tuning and improve the accuracy of parameter tuning;

[0232] The expression for updating the learning rate is as follows:

[0233]

[0234] in: is the updated learning rate parameter, is the updated sensitivity parameter, η NN , K NNare adjustment factors, ΔE NN is the difference between the environmental fitness of this iteration and the previous iteration;

[0235] S35. Based on the result of the environmental perception evaluation, dynamically adjust the hyperparameters controlling the resource allocation sensitivity between the global search and the local search to avoid falling into the local optimal solution; the expression of the adjustment is as follows:

[0236]

[0237] in: is the hyperparameter for the sensitivity of the adjusted control resource allocation, ρ NN is a balanced hyperparameter (preferably set to 0.1), | | is the absolute value of compliance;

[0238] S36. Repeat steps S31-S35 until the preset iteration stop condition is reached, obtain the characteristic distribution network data training set and complete the feature extraction model training.

[0239] In this embodiment, the preset condition for stopping iteration is: reaching a preset maximum number of iterations; the preset maximum number of iterations is set to 1000 times.

[0240] S4. Using an auto-encoding neural network algorithm based on feature adaptive refinement, perform feature dimensionality reduction on the feature distribution network data training set, and train a feature dimensionality reduction model to obtain a dimensionality reduction distribution network data training set;

[0241] In this embodiment, the feature distribution network data set after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. This solution adopts an autoencoder neural network algorithm based on feature adaptive refinement for feature dimensionality reduction; the autoencoder neural network algorithm includes two parts: an encoder and a decoder; the encoder is used to convert the features extracted from the high-dimensional input into a low-dimensional hidden representation, and the decoder is used to reconstruct the low-dimensional hidden representation back to the original high-dimensional space to form a reconstructed output. This solution adopts a feature coupling mechanism so that the autoencoder can not only reduce the dimension of the data distribution network data, but also more finely capture and reconstruct the intrinsic correlation and structural characteristics in the input data distribution network data, and optimize the information flow of the network.

[0242] Further, S41, initialize the network parameters of the autoencoder, the network including the weight and bias of the autoencoder; the expression thereof is as follows:

[0243]

[0244] in: is the initial weight of the autoencoder, W r is the weight parameter of the autoencoder, The mean of the table is 0 and the variance is The normal distribution of is the number of neurons in the input layer of the autoencoder, is the initial bias of the autoencoder, b r is the bias parameter of the autoencoder;

[0245] S42, propagating the distribution network data after feature extraction through the encoder layer sequence and converting it into a low-dimensional representation; the expression of the low-dimensional representation is as follows:

[0246]

[0247] Where: H r is the low-dimensional representation of the encoder output, Re( ) is the ReLU activation function, are the weight and bias of the encoder respectively, is the reconstructed output of the distribution network data, are the weight and bias of the decoder respectively;

[0248] S43. In the hidden layer of the autoencoder, the weights are adjusted by a dynamic weight adjustment layer, and the gradient of the loss function with respect to the weights is solved by the chain rule;

[0249] The expression of the adjustment weight is as follows:

[0250]

[0251] Where: ΔW r is the weight update amount, tanh( ) is the hyperbolic tangent function, λ r is the learning rate (preferably set to 0.01), γ r is the adjustment factor (preferably set to 3.14), is the loss function L r About weight W r The gradient of L r is the reconstruction error of the loss function,

[0252] The chain rule solves the gradient of the loss function with respect to the weight as follows:

[0253]

[0254] in: is the partial derivative of the loss function with respect to the reconstruction error output;

[0255] The calculation expression of the mean square error of the loss function is as follows:

[0256]

[0257] Among them: |||| is the L2 norm, αr is the regularization coefficient (preferably set to 0.3), is the L1 norm of the encoder weight;

[0258] S44. Based on the adjusted weights, the weight parameters and bias parameters of the autoencoder are updated using the gradient descent method; the expressions of the weight parameters and bias parameters are as follows:

[0259]

[0260] in: is the weight parameter of the autoencoder at the t+1th iteration, is the weight parameter of the autoencoder at the tth iteration, η r is the learning rate, is the bias parameter of the autoencoder at the t+1th iteration, is the bias parameter of the autoencoder at the tth iteration, t is the number of iterations, are the gradients of the loss function with respect to weights and biases respectively;

[0261] The learning rate is set by adaptive adjustment to dynamically adjust according to different stages of network training, and its expression is as follows:

[0262]

[0263] in: is the learning rate for the t+1th iteration, is the learning rate of the tth iteration, δ is the sensitivity parameter for adjusting the learning rate (preferably set to 0.5);

[0264] S45. Repeat steps S41-S44 until the iteration stop condition is met, obtain the reduced-dimensional distribution network data training set and complete the feature-dimensionality reduction model training.

[0265] In this embodiment, the preset condition for stopping iteration is: reaching a preset maximum number of iterations; the preset maximum number of iterations is set to 1000 times.

[0266] S5. Using a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, classifying the reduced-dimensional distribution network data training set, and training the classifier model to obtain the distribution network state classification category;

[0267] In this embodiment, the reduced-dimensional distribution network data set is input into the classifier to train the classifier model; this scheme adopts a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, and adopts a dynamic error compensation factor, which is automatically adjusted according to the classification error in each training iteration, thereby optimizing the weight update process of subsequent iterations and improving the classification accuracy.

[0268] Furthermore, the step S5 specifically includes:

[0269] S51, initializing the parameters of the kernel function and the weight and bias of the network, and dynamically adjusting the width parameter of the kernel function according to the error distribution of each iteration;

[0270] The initialization expression is as follows:

[0271] K u (x u , x′ u )=exp(-γ u ||x u -x′ u || 2 );

[0272]

[0273] Where: K u is the kernel function, x u is the distribution network data point, x′ u To distinguish from x u Another distribution network data point, γ u is the width parameter of the kernel function, are the initialized weights and biases, is the initialization variance (preferably set to 0.01);

[0274] The expression for the width parameter is as follows:

[0275]

[0276] Where: γ u (t+1) is the width parameter of the kernel function at the t+1th iteration, γ u (t) is the width parameter of the kernel function at the tth iteration, ||e u (t)||2 is the L2 norm of the error of the current iteration, ||x u (t)||2 is the L2 norm of the reduced-dimensional distribution network data training set, ∈ ug A small constant (preferably set to 0.01) to prevent division by zero, η γ The learning rate adjusted for the kernel parameters;

[0277] S52. Before each training iteration, the current error compensation factor is calculated according to the error of the previous iteration to adjust the learning rate and weight update rule of this iteration to reduce the overall training error. The expression of the error compensation factor is as follows:

[0278]

[0279] Where: u (t) is the error compensation factor, || || is the L2 norm, λ u is the adjustment coefficient of error compensation (preferably set to 3), e u (t-1) is the error vector of the t-1th iteration, e u (t-2) is the error vector of the t-2th iteration, eu ( t) is the error vector of the tth iteration, y u is the true label vector, is the prediction for the tth iteration;

[0280] S53, dynamically adjust the weight of the kernel extreme learning machine based on the error compensation factor and kernel technology;

[0281] In this scheme, kernel technology is used to map the input distribution network data to a high-dimensional space, linear regression is performed in the high-dimensional space, the weight update formula is adjusted according to the calculated error compensation factor, and the rate and direction of change of the weight are dynamically adjusted. Taking the weight parameters of the kernel extreme learning machine as an example, the update rule adopts kernel technology and error compensation factor adjustment.

[0282] The expression of the weight of the kernel extreme learning machine is as follows:

[0283]

[0284] in: is the weight of the kernel extreme learning machine at the t+1th iteration, is the weight of the kernel extreme learning machine at the tth iteration, η u is the learning rate of the kernel extreme learning machine, is the gradient of the loss function of the kernel extreme learning machine with respect to the weight, L u ( ) is the loss function of the kernel extreme learning machine, nE is the number of samples in the current batch, and I is the unit matrix;

[0285] The expression of the loss function of the kernel extreme learning machine is as follows:

[0286]

[0287] in: is the predicted value calculated by kernel technique and weights, is the true label, λ fg is the regularization parameter (preferably set to 0.3);

[0288] S54, after the weight adjustment is completed, check the gradient value and determine whether the gradient value exceeds the stage threshold;

[0289] If the gradient value exceeds the preset cutoff threshold, Then truncation is performed to prevent the gradient explosion problem during training and ensure the stability of the training process; the expression of the truncation is as follows:

[0290]

[0291] Where: θ u is the gradient cutoff threshold;

[0292] S55. Repeat steps S51-S54 until the iteration stop condition is met, the distribution network status classification category is obtained and the classifier model training is completed.

[0293] In this embodiment, the preset condition for stopping iteration is: reaching a preset maximum number of iterations; the preset maximum number of iterations is set to 1000 times.

[0294] S6. Input the new distribution network data set samples into the feature extraction model and the feature dimension reduction model in sequence for feature processing, and then input them into the classifier model for classification to obtain the classification results. According to the classification results, the carrying capacity of the distribution network can be obtained.

[0295] Furthermore, the classification result includes any one of a normal charging state, a high load state, a low load state and a fault state.

[0296] If the classification result is a normal charging state, it means that the current carrying capacity of the distribution network is normal. This state means that the electric vehicle is charged under standard conditions and the impact on the distribution network is within an acceptable range. This state can be used as a criterion for judging the current carrying capacity of the distribution network.

[0297] If the classification result is a high load state, it means that the current carrying capacity of the distribution network is close to the upper limit. This state means that during peak power consumption periods or when multiple electric vehicles are charging at the same time, the distribution network needs to handle a larger load; if the high load state occurs frequently and lasts for a long time, it means that the current carrying capacity of the distribution network is close to its upper limit, and it may be necessary to strengthen the infrastructure or conduct load management;

[0298] If the classification result is a low load state, it means that the current carrying capacity of the distribution network still has redundant capacity. This state means that there are fewer charging electric vehicles or the charging demand is low, and the distribution network can bear additional loads. The redundancy in this state can be used as a basis for optimizing the dispatching strategy.

[0299] If the classification result is a fault state, it means that the distribution network does not have the carrying capacity. This state represents that the distribution network or the charging facility has a fault due to load overload or other abnormal conditions. This state directly reflects the weakness of the distribution network in dealing with load fluctuations. By analyzing the frequency and cause of the fault state, the weak links in the distribution network can be identified and improved.

[0300] Furthermore, based on the classification results, the load management strategy of the distribution network can be dynamically adjusted. For example, after detecting multiple high loads or fault conditions, the operating parameters of the distribution equipment can be automatically adjusted to reduce the load pressure; or part of the load can be transferred to the low load period through the peak shaving and valley filling strategy.

[0301] In addition, the load status of the distribution network can be monitored and predicted in real time, and the stability of the distribution network can be maintained through intelligent control measures (such as scheduling electric vehicle charging time, enabling backup power supplies, etc.).

[0302] Embodiment 2:

[0303] See also Figure 4 , a distribution network carrying capacity assessment system for large-scale automobile charging loads, the system is applied to the method described in Example 1, the system comprising:

[0304] The training set acquisition module 1 is used to acquire the data generated when the electric vehicle interacts with the distribution network during the charging process of the electric vehicle, and to annotate the data to obtain the original distribution network data training set;

[0305] Furthermore, the data collected by the training set acquisition module 1 include: grid load, voltage, frequency, frequency of electric vehicle use, mileage, charging habits, charging time, charging station location, charging type, and vehicle type;

[0306] The marked categories include: normal charging state, high load state, low load state and fault state.

[0307] Data expansion module 2 is used to expand the original distribution network data training set based on a composite data expansion method of variational sparse matrix decomposition and polynomial fitting, and train a data expansion model to obtain an expanded distribution network data training set;

[0308] Further, the data expansion module 2 performs data expansion according to steps S21-S27 in Example 1;

[0309] The feature extraction module 3 is used to extract features from the expanded distribution network data training set by using a neural network algorithm based on adaptive biological influence optimization, and to train a feature extraction model to obtain a feature distribution network data training set;

[0310] Further, the feature extraction module 3 performs feature extraction according to steps S31-S36 in Example 1;

[0311] The feature dimension reduction module 4 is used to adopt an auto-encoding neural network algorithm based on feature adaptive refinement to perform feature dimension reduction on the feature distribution network data training set, and train the feature dimension reduction model to obtain the reduced dimension distribution network data training set;

[0312] Further, the feature dimension reduction module 4 performs feature dimension reduction according to steps S41-S45 in Example 1;

[0313] The data classification module 5 is used to classify the reduced-dimensional distribution network data training set by using a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, and train a classifier model to obtain a distribution network state classification category;

[0314] Further, the data classification module 5 performs classification according to steps S51-S55 in Example 1;

[0315] The carrying capacity assessment module 6 is used to input the new distribution network data set samples into the feature extraction model and the feature dimensionality reduction model in sequence for feature processing, and then input them into the classifier model for classification to obtain the classification results. According to the classification results, the carrying capacity of the distribution network can be obtained.

[0316] Furthermore, the classification result of the carrying capacity evaluation module 6 includes: any one of a normal charging state, a high load state, a low load state and a fault state.

[0317] The specific content of the carrying capacity of the distribution network can be obtained according to the classification result, please refer to step S6 in Example 1.

[0318] Embodiment 3:

[0319] See also Figure 5 , a distribution network carrying capacity assessment device for large-scale automobile charging loads, the device comprising a processor 7 and a memory 8;

[0320] The memory 8 is used to store the computer program code 81 and transmit the computer program code 81 to the processor 7;

[0321] The processor 7 is used to execute the distribution network carrying capacity assessment method for large-scale automobile charging loads described in Example 1 according to the instructions in the computer program code 81.

[0322] In this embodiment, a distribution network carrying capacity assessment storage medium for large-scale automobile charging loads is also included, on which a computer program is stored. When the computer program is executed by a processor, the distribution network carrying capacity assessment method for large-scale automobile charging loads described in Example 1 is implemented.

[0323] Generally speaking, the computer instructions for implementing the method of the present invention may be carried in any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media may include any computer-readable media, except for the signal itself that is temporarily propagating.

[0324] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EKROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, device, or device.

[0325] Computer program code for performing the operation of the present invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, SMalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages, in particular, Python suitable for neural network computing and platform frameworks based on TensorFlow, PyTorch, etc. can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer or to an external computer (for example, using an Internet service provider to connect via the Internet) through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0326] The above-mentioned device and non-temporary computer-readable storage medium can be found in the detailed description of a distribution network carrying capacity assessment method for large-scale automobile charging loads and its beneficial effects, which will not be repeated here.

[0327] Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A distribution network carrying capacity assessment method for large-scale automobile charging loads, characterized in that: include: S1. Obtain the data generated when the electric vehicle interacts with the distribution network during the charging process of the electric vehicle, and annotate it to obtain the original distribution network data training set; The data include grid load, voltage, frequency, frequency of electric vehicle use, mileage, charging habits, charging time, charging station location, charging type, and vehicle type; The marked categories include normal charging state, high load state, low load state and fault state; S2. Based on the composite data expansion method of variational sparse matrix decomposition and polynomial fitting, the original distribution network data training set is expanded, and the data expansion model is trained to obtain the expanded distribution network data training set; S3, using a neural network algorithm based on adaptive biological influence optimization to extract features from the expanded distribution network data training set, and training a feature extraction model to obtain a feature distribution network data training set; S4. Using an auto-encoding neural network algorithm based on feature adaptive refinement, perform feature dimensionality reduction on the feature distribution network data training set, and train a feature dimensionality reduction model to obtain a dimensionality reduction distribution network data training set; S5. Using a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, classifying the reduced-dimensional distribution network data training set, and training the classifier model to obtain the distribution network state classification category; S6. Input the new distribution network data set samples into the feature extraction model and the feature dimension reduction model in sequence for feature processing, and then input them into the classifier model for classification to obtain the classification results. According to the classification results, the carrying capacity of the distribution network can be obtained.

2. The method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to claim 1 is characterized in that: The step S2 specifically includes: S21, initialize the parameters of the variational autoencoder; the parameters include the weight and bias of the variational autoencoder; the expression of the initialization is as follows: in: is the initial weight matrix of the variational autoencoder, is the initial bias vector of the variational autoencoder, is the variance of the initialization of the variational autoencoder; S22, input the original distribution network data training set into the variational autoencoder, map the data into the potential feature space through nonlinear transformation, and use nonlinear activation function to extract key features; the expression of the transformation is as follows: Where: z c is the encoded potential feature, Res( ) is based on the steep ReLU activation function, are the weights and biases of the variational autoencoder, X c is the vector of the input original distribution network data training set; The steep ReLU activation function uses an adaptive parameter to control the activation threshold, and its calculation expression is as follows: Where: X is the input of the ReLU activation function, σ is the adaptive weight, and β is the parameter that controls the steepness of the Sigmoid curve; S23, based on the encoded latent features, L1 regularization is used to implement sparsity constraints in the latent feature space, the key feature representation of the latent feature space is optimized, and the total loss function is calculated; The expression of the sparsity constraint is as follows: in: is the sparsity loss function, λ c is the sparsity regularization coefficient, z ci is the i-th element in the encoded latent feature; The expression of the total loss function is as follows: in: is the total loss of the variational autoencoder, is the reconstruction loss of the variational autoencoder; The sparsity regularization coefficient is set in an adaptive adjustment manner, and its expression is as follows: Where: γ c is the adjustment coefficient, ρ c is the target sparsity, is the average sparsity of the current iteration; S24, performing polynomial fitting on key features in the potential feature space, and dynamically adjusting coefficients of the polynomial according to distribution characteristics of the key features; The expression of the polynomial fitting is as follows: in: is the key feature after polynomial fitting, a cn are the coefficients of the nth degree polynomial, is the nth power of the potential distribution network data feature, Nf is the total number of samples in the current input batch; The nth order polynomial coefficient a cn The calculation expression of the optimization process is as follows: Among them: a ck is the basic weight of the polynomial coefficient, Kf is the highest order of the fitting polynomial; S25, mapping the key features after polynomial fitting back to the original data space through the decoder to generate new distribution network data points; The expression of the new distribution network data point is as follows: in: is a new distribution network data point, Sig() is the Sigmoid activation function, are the weight and bias of the decoder respectively; S26. Dynamically adjust the learning rate of the variational autoencoder according to the loss function of the variational autoencoder; The expression of the learning rate is as follows: in: are the learning rates for iterations t+1 and t, respectively, c is the learning rate decay coefficient, is the gradient of total loss with respect to learning rate; S27. Repeat steps S21-S26 until the preset stop iteration condition is met, to obtain the expanded distribution network data training set and complete the data expansion model training.

3. The method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to claim 1 is characterized in that: The step S3 specifically includes: S31, initialize the parameters of the neural network, and define the weights and biases of the neural network; the expressions of the weights and biases are as follows: Where: W NN 、b NN are the weights and biases of the neural network, respectively, ~ means obeying a specific distribution, σ sn is the initial standard deviation, are the initial weights and initial biases of the neural network, is a normal distribution; S32. At the beginning of each iteration, the loss function under the current parameter configuration is calculated, and the environmental fitness is estimated, and then resources are allocated based on the contribution of each parameter; The expression of the environmental adaptability is as follows: E NN sexp(-β NN ·(W NN ,b NN )) Where: E NN is the environmental fitness, β NN is the sensitivity parameter; The loss function is calculated using cross entropy loss, and its calculation expression is as follows: Where: L NN (W NN , b NN ) is the loss function, C is the number of categories, y c is the one-hot encoding of the true label, is the predicted probability, The predicted probability is calculated by the softmax function, and its calculation expression is as follows: Where: z c is the logistic regression probability output by the neural network, z k is the logistic regression probability of the kth category output by the network; The resource calculation expression for each parameter competition allocation is as follows: Where: R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and l are both indexes, γ NN is a hyperparameter, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of S33, based on the resource competition results, update the weights and biases, and use the momentum term to smooth the learning process of the weight parameters; The updated expression is as follows: in: is the weight parameter update of the jth neuron in the i-th neural network layer, is the bias parameter update of the i-th layer of neurons, is the bias of the i-th layer neuron, R i The resources allocated for the parameters of the i-th neural network layer, α NN is the learning rate; The expression of the momentum term to smooth the learning process of the weight parameters is as follows: in: is the momentum term of the jth neuron in the first neural network layer at the tth iteration, is the momentum term of the jth neuron in the i-th neural network layer at the t-1th iteration, β xm is the momentum decay parameter, is the weight parameter of the jth neuron in the i-th neural network layer at the t+1th iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-th iteration, S34, adjusting the sensitivity parameter and the learning rate parameter according to the result of the environmental perception evaluation; the result of the environmental perception evaluation refers to the difference between the environmental fitness of the current iteration and the previous iteration; The expression for updating the learning rate is as follows: in: is the updated learning rate parameter, is the updated sensitivity parameter, η NN , K NN are adjustment factors, ΔE NN is the difference between the environmental fitness of this iteration and the previous iteration; S35. Based on the result of the environment perception evaluation, dynamically adjust the hyperparameters controlling the resource allocation sensitivity between the global search and the local search; the expression of the adjustment is as follows: in: is the hyperparameter for the sensitivity of the adjusted control resource allocation, ρ NN is a balanced hyperparameter; S36. Repeat steps S31-S35 until the preset iteration stop condition is reached, obtain the characteristic distribution network data training set and complete the feature extraction model training.

4. The method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to claim 1 is characterized in that: The step S4 specifically includes: S41, initialize the network parameters of the autoencoder, the network includes the weight and bias of the autoencoder; the expression is as follows: in: is the initial weight of the autoencoder, Wr, is the weight parameter of the autoencoder, The mean of the table is 0 and the variance is The normal distribution of is the number of neurons in the input layer of the autoencoder, is the initial bias of the autoencoder, b r is the bias parameter of the autoencoder; S42, propagating the characteristic distribution network data training set through the encoder layer sequence and converting it into a low-dimensional representation; the expression of the low-dimensional representation is as follows: Where: H r is the low-dimensional representation of the encoder output, Re( ) is the ReLU activation function, are the weight and bias of the encoder respectively, To reconstruct the output, are the weight and bias of the decoder respectively; S43. In the hidden layer of the autoencoder, the weights are adjusted by a dynamic weight adjustment layer, and the gradient of the loss function with respect to the weights is solved by the chain rule; The expression of the adjustment weight is as follows: Where: ΔW r is the weight update amount, tanh( ) is the hyperbolic tangent function, λ r is the learning rate, γ r is the adjustment factor, is the loss function L r About weight W r The gradient of L r is the reconstruction error of the loss function, The chain rule solves the gradient of the loss function with respect to the weight as follows: in: is the partial derivative of the loss function with respect to the reconstruction error output; The calculation expression of the mean square error of the loss function is as follows: Among them: || || is the L2 norm, α r is the regularization coefficient, is the L1 norm of the encoder weight, X r is the data input to the autoencoder; S44. Based on the adjusted weights, the weight parameters and bias parameters of the autoencoder are updated using the gradient descent method; the expressions of the weight parameters and bias parameters are as follows: in: is the weight parameter of the autoencoder at the t+1th iteration, is the parameter of the autoencoder weight at the tth iteration, η r is the learning rate, is the parameter of the autoencoder bias at the t+1th iteration, is the parameter of the autoencoder bias at the tth iteration, t is the number of iterations, are the gradients of the loss function with respect to weights and biases respectively; The learning rate is set by adaptive adjustment, and its expression is as follows: in: is the learning rate for the t+1th iteration, is the learning rate of the tth iteration, δ is the sensitivity parameter for adjusting the learning rate; S45. Repeat steps S41-S44 until the iteration stop condition is met, obtain the reduced-dimensional distribution network data training set and complete the feature-dimensionality reduction model training.

5. The method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to claim 1 is characterized in that: The step S5 specifically includes: S51, initializing the parameters of the kernel function and the weight and bias of the neural network, and dynamically adjusting the width parameter of the kernel function according to the error distribution of each iteration; The initialization expression is as follows: K u (x u ,x′ u )=exp(-γ u ||x u -x′ u || 2 ); Where: K u is the kernel function, x u is the distribution network data point, x′ u To distinguish from X u Another distribution network data point, γ u is the width parameter of the kernel function, To initialize the weights and biases of the neural network, is the initialization variance; The expression for the width parameter is as follows: Where: γ u (t+1) is the width parameter of the kernel function at the t+1th iteration, γ u (t) is the width parameter of the kernel function at the tth iteration, ||e u (t)||2 is the L2 norm of the error of the current iteration, ||x u (t)||2 is the L2 norm of the reduced-dimensional distribution network data training set, ∈ ug To prevent division by zero, η γ The learning rate adjusted for the kernel parameters; S52. Before each training iteration begins, calculate the current error compensation factor according to the error of the previous iteration; the expression of the error compensation factor is as follows: Where: u (t) is the error compensation factor, || || is the L2 norm, λ u is the adjustment coefficient of error compensation, e u (t-1) is the error vector of the t-1th iteration, e u (t-2) is the error vector of the t-2th iteration, e u (t) is the error vector of the tth iteration, y u is the true label vector, is the prediction for the tth iteration; S53, dynamically adjust the weight of the kernel extreme learning machine based on the error compensation factor and kernel technology; The expression of the weight of the kernel extreme learning machine is as follows: in: is the weight of the kernel extreme learning machine at the t+1th iteration, is the weight of the kernel extreme learning machine at the tth iteration, η u is the learning rate of the kernel extreme learning machine, is the gradient of the loss function of the kernel extreme learning machine with respect to the weight, L u ( ) is the loss function of the kernel extreme learning machine, nE is the number of samples in the current batch, and I is the unit matrix; The expression of the loss function of the kernel extreme learning machine is as follows: in: is the predicted value calculated by kernel technique and weights, is the true label, λ fg is the regularization parameter; S54, after the weight adjustment is completed, check the gradient value and determine whether the gradient value exceeds the stage threshold; If the gradient value exceeds the preset cutoff threshold, Then perform truncation processing; the expression of the truncation processing is as follows: Where: θ u is the gradient cutoff threshold; S55. Repeat steps S51-S54 until the iteration stop condition is met, the distribution network status classification category is obtained and the classifier model training is completed.

6. The method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to claim 1 is characterized in that: The classification result includes any one of a normal charging state, a high load state, a low load state and a fault state.

7. The method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to claim 6 is characterized in that: According to the classification results, the carrying capacity of the distribution network can be obtained as follows: If the classification result is a normal charging state, it means that the current carrying capacity of the distribution network is normal, which means that the electric vehicle is charged under standard conditions and the impact on the distribution network is within an acceptable range; If the classification result is a high load state, it means that the current carrying capacity of the distribution network is close to the upper limit, which means that during peak power consumption hours or when multiple electric vehicles are charging at the same time, the distribution network needs to handle a larger load; If the classification result is a low load state, it means that the current carrying capacity of the distribution network still has redundant capacity, which means that there are fewer charging electric vehicles or the charging demand is low, and the distribution network can bear the additional load; If the classification result is a fault state, it means that the distribution network does not have the carrying capacity, which means that the distribution network or the charging facility has a fault due to overload or other abnormal conditions.

8. A distribution network carrying capacity assessment system for large-scale automobile charging loads, characterized by: The system is applied to the method described in any one of claims 1 to 7, and the system comprises: The training set acquisition module (1) is used to acquire data generated when the electric vehicle interacts with the distribution network during the charging process of the electric vehicle, and to annotate the data to obtain the original distribution network data training set; The data include grid load, voltage, frequency, frequency of electric vehicle use, mileage, charging habits, charging time, charging station location, charging type, and vehicle type; The marked categories include normal charging state, high load state, low load state and fault state; A data expansion module (2) is used to expand the original distribution network data training set based on a composite data expansion method of variational sparse matrix decomposition and polynomial fitting, and train a data expansion model to obtain an expanded distribution network data training set; A feature extraction module (3) is used to extract features from the expanded distribution network data training set using a neural network algorithm based on adaptive biological impact optimization, and to train a feature extraction model to obtain a feature distribution network data training set; A feature dimension reduction module (4) is used to use an autoencoder neural network algorithm based on feature adaptive refinement to perform feature dimension reduction on a feature distribution network data training set, and train a feature dimension reduction model to obtain a dimension-reduced distribution network data training set; A data classification module (5) is used to classify the reduced-dimensional distribution network data training set using a kernel extreme learning machine algorithm based on adaptive error compensation as a classifier, and to train a classifier model to obtain a distribution network state classification category; The carrying capacity assessment module (6) is used to input the new distribution network data set samples into the feature extraction model and the feature dimension reduction model in sequence for feature processing, and then input them into the classifier model for classification to obtain the classification results. According to the classification results, the carrying capacity of the distribution network can be obtained.

9. A distribution network carrying capacity assessment device for large-scale automobile charging loads, characterized by: The device comprises a processor (7) and a memory (8); The memory (8) is used to store computer program code (81) and transmit the computer program code (81) to the processor (7); The processor (7) is used to execute the distribution network carrying capacity assessment method for large-scale automobile charging loads according to any one of claims 1 to 7 according to the instructions in the computer program code (81).

10. A storage medium for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads, on which a computer program is stored, characterized in that: When the computer program is executed by a processor, the method for evaluating the carrying capacity of a distribution network for large-scale automobile charging loads according to any one of claims 1 to 7 is implemented.

Images