Evaluation method for photovoltaic maximum safety access capacity of medium-voltage power distribution network

Abstract

The invention discloses a method for evaluating the photovoltaic maximum safety access capacity of a medium-voltage power distribution network. The method comprises the following steps: calculating the real part and the imaginary part of the voltage of each node in an actual power distribution area; pre-training the two layers of neural networks and setting physical constraints to obtain a reconstructed power flow model; establishing a power flow solver model based on a multilayer perceptron (MLP); inputting a voltage real part and an imaginary part predicted by the power flow solver model into the reconstructed power flow model to obtain active power and reactive power, comparing the active power and the reactive power, and adding a comparison result as a regularization item into the power flow solver model to obtain a trained power flow solver model; and predicting the voltage of each node, and carrying out screening statistics to obtain the photovoltaic maximum penetration capacity in the safety domain. The method is high in precision, high in efficiency and low in operation and maintenance cost, dynamic visualization of result data is achieved, and the method is suitable for the condition that part of topological parameters are uncertain.

Description

Technical Field

[0001] The present invention relates to the field of distribution networks, and particularly to a method for evaluating the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network. Background Art

[0002] Currently, there are technical challenges in medium-voltage distribution networks such as incomplete topological information (e.g., incorrect line connection relationships, missing branch parameters) and dynamic changes in operating states (frequent reconstruction, insufficient measurement information), resulting in the failure of traditional capacity evaluation methods based on physical power flow equations. Existing solutions (such as genetic algorithms, clustering analysis) rely on complete topological parameters, are prone to introducing errors due to missing information, and are difficult to handle high-frequency reconstruction scenarios. Summary of the Invention

[0003] To solve the above technical problems, the present invention provides a method for evaluating the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network with high accuracy, high efficiency, and low operation and maintenance costs.

[0004] The technical solution of the present invention to solve the above technical problems is: a method for evaluating the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network, including the following steps:

[0005] S1: Collect the voltage, phase angle, actual active power, and actual reactive power of each node in the actual distribution substation area, and calculate the real part and imaginary part of the voltage of each node;

[0006] S2: Take the real part and imaginary part of the voltage of each node as the input of a two-layer neural network, pre-train the two-layer neural network, initialize the weight matrix using the sparsity and symmetry of the admittance matrix, and set physical constraints to obtain a reconstructed power flow model;

[0007] S3: Establish a power flow solver model based on a multi-layer perceptron MLP, and input the active power and reactive power of each node to predict the real part and imaginary part of the voltage;

[0008] S4: Input the real part and imaginary part of the voltage predicted by the power flow solver model into the reconstructed power flow model to obtain the active power and reactive power. Compare the obtained active power and reactive power with the actual active power and actual reactive power collected in step S1, and add the comparison result as a regularization term to the power flow solver model to obtain a trained power flow solver model;

[0009] S5: Generate multiple photovoltaic deployment scenarios through the Monte Carlo simulation method, input the actual active power and actual reactive power of each node under all photovoltaic deployment scenarios into the trained power flow solver model, predict the voltage of each node, and screen and statistically obtain the maximum access capacity of photovoltaic power within the safety domain.

[0010] In the above assessment method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks, in step S1, the real part and the imaginary part of the voltage at each node are expressed as:

[0011]

[0012] where i = 1, …, n, n is the number of nodes in the distribution network, |v i | is the voltage amplitude at the i-th node, θ i is the voltage phase angle difference between the i-th node and the reference node (the connection node of the distribution network bus), u i and ω i are respectively the real part and the imaginary part of the voltage phasor at the i-th node.

[0013] In the above assessment method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks, in step S2, the structure of the two-layer neural network includes an input layer, a multiplication layer, and a linear layer:

[0014] Input layer: Receives the real part and the imaginary part data of the voltage at each node;

[0015] Multiplication layer: Extracts the features of the input data;

[0016] Linear layer: Performs weighted summation on the output of the multiplication layer to predict the output of the power flow model.

[0017] In the above assessment method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks, the specific process of step S2 is as follows:

[0018] Express the mapping fS1 of the two-layer neural network as fS1(x) = W1δ1(x), where δ1(x) represents the mapping from the input x to the multiplication term, and W1 is the weight matrix of the linear layer; for fS1, use the mean squared error MSE of the output result as the objective function, and add the l1 norm of W1 as the regularization term to enhance sparsity:

[0019]

[0020] where L MSE is the mean squared error loss function, which is used to measure the error between the model prediction value and the true value; represents the normalization factor for the number of samples N, which is used to calculate the average error; y n is the true value of the n-th sample; x n represents the n-th sample, that is, the real part and the imaginary part of the voltage at the n-th node; ||·|| p is the l p norm, and λ1 is a positive constant;

[0021] Pre-train the two-layer neural network, set the elements corresponding to the nodes without relationships in W1 to 0, and then constrain the elements in W1 according to the symmetry of the admittance matrix, so that u i u p is equal to the element corresponding to v i v p in W1, and u i u p is equal to the element corresponding to u p u i in W1.

[0022] In the above method for evaluating the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks, in step S3, the power flow solver model based on the multi-layer perceptron MLP is: a single-hidden-layer MLP for voltage prediction of an n-node distribution system. The input layer of the power flow solver model contains 2n linear neurons, and the hidden layer and output layer contain m and 2n neurons respectively; the indices r, j, and k are used to indicate the neurons in the input layer, hidden layer, and output layer respectively, where r = 1, 2,..., 2n, j = 1, 2,..., m, k = 1, 2,..., 2n; in forward propagation, the result of multiplying the weighted vector by the input vector is added to the bias vector as the input of the transfer function, and the output of the transfer function is the input of the next layer; each layer has a weighted vector and a bias vector. In forward propagation, for the input X = (x1, x2,..., x 2n ) T where x 2n is the characteristic element of the 2nth input sample, and the characteristic elements include the actual active power P and the actual reactive power Q. The superscript T represents the transpose. The input layer does not use an activation function, and the calculation function f1(x r ) of the rth neuron x r in the input layer is as follows:

[0023] f1(x r ) = x r , r = 1,..., 2n

[0024] The hidden layer uses the sigmoid activation function, and the calculation of the jth neuron z j in the hidden layer is as follows:

[0025]

[0026]

[0027] In the formula, is the weighted vector between the input layer and the hidden layer, is the bias vector between the input layer and the hidden layer, and sigmoid(z j ) is the activation of z j ;

[0028] The output o of the k-th neuron in the output layer k is as follows:

[0029] o k = f1(y k ) = y k , k = 1, 2,..., 2n

[0030]

[0031] where y k is the weighted sum of sigmoid(z j ), is the weighted vector between the hidden layer and the output layer, is the bias vector between the hidden layer and the output layer.

[0032] In the above method for evaluating the maximum safe access capacity of photovoltaic power in the medium-voltage distribution network, in step S3, the power flow solver model is trained using the standard backpropagation algorithm. The gradient of the loss function with respect to the weights and biases is calculated layer by layer through the chain rule. The core process includes: first, calculating the gradient of the loss with respect to the activation value from the output layer and obtaining the error term by combining the derivative of the activation function; then, propagating the error back through the weight matrix layer by layer to the hidden layer while calculating the gradient of the weights and biases for each layer; finally, updating the weight and bias parameters of the model using gradient descent according to the learning rate to minimize the loss function E:

[0033]

[0034] where E is the root mean square error corresponding to the input x, and d k is the expected output of the k-th neuron in the output layer corresponding to o k .

[0035] In the above method for evaluating the maximum safe access capacity of photovoltaic power in the medium-voltage distribution network, during the training process of the power flow solver model in step S3, Gaussian noise is added to the input data. The expression for the noise input is as follows:

[0036] s (n) = x (n) + d

[0037] where x (n) is the training input data without error, s (n) is the input data with error, and d is a random variable.

[0038] In the above evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks, in step S4, a regularizer is introduced through multi-task learning. The regularizer is inspired by a supervised autoencoder and consists of two parts: an encoder and a decoder, which respectively focus on the inverse problem and the forward problem: power flow solution and power flow modeling. The encoder part is an MLP based on noise injection, which focuses on solving the power flow problem by learning the mapping from power flow input to bus voltage, and ensures the network's resistance to noise by introducing random noise.

[0039] The decoder part is a neural network that performs an auxiliary task, that is, reconstructing the power flow model, mapping the bus voltage to power injection, and the decoder output is y PQ :

[0040] y PQ = f de (y uω ; W de , b de )

[0041] In the formula, f de is the calculation function, y uω is the output of the MLP encoder, W de is the weighted vector between the encoder and the decoder in the calculation function, b de is the bias vector of the weighted vector between the encoder and the decoder;

[0042] The total loss function is the weighted sum of the voltage prediction loss generated by the encoder and the power reconstruction loss generated by the decoder. Let Δμω = [μ; ω] - y μω , ΔPQ = [P; Q] - y PQ , where Δμω is the voltage prediction loss, [μ; ω] is the actual voltage, y μω is the voltage generated by the encoder, ΔPQ is the power reconstruction loss, y PQ is the power generated by the decoder. The total loss E 总 of a single training sample is:

[0043] E 总 = α MLP ζ(Δuω) + α SNN ζ(ΔPQ)

[0044] In the formula, ξ() represents the mean square error function, α MLP is the weight of the supervised loss, α SNN is the weight of the unsupervised loss, α MLP and α SNN are respectively hyperparameters for balancing the encoder and decoder parts.

[0045] In the above method for evaluating the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network, in step S4, a power flow solver model is trained to find the parameters that minimize the total loss function; when training the power flow solver model, the error of the voltage prediction layer is:

[0046]

[0047] In the formula, the first term Δuω directs the learning process of the MLP encoder to the parameters effective for minimizing the bus voltage prediction error. Under the guidance of the second term Among the parameters effective in generating voltage estimates, the learning algorithm preferentially selects the parameters that generate small power mismatches.

[0048] In the above method for evaluating the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network, in step S5, the process of screening and statistically obtaining the maximum access capacity of photovoltaic power within the safety domain is as follows: Determine whether the predicted voltage of each node is within the set voltage safety domain. If it is within the safety domain, increase the photovoltaic power deployed in the randomly generated photovoltaic scenario. If it is not within the set voltage safety domain, save the total photovoltaic value of the previous scheme. When the total number of photovoltaic deployment schemes is reached, output the maximum result.

[0049] The beneficial effects of the present invention are as follows: A hybrid-regularized MLP power flow solver model based on physical consistency is constructed to replace the traditional physical model, handle distribution networks with uncertain topological parameters, and improve the generalization ability and adaptability of the evaluation system. The noise injection technology is adopted to enhance the model's resistance to noise and uncertainty, and improve the generalization ability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the overall flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0051] The present invention will be further described below with reference to the drawings and embodiments.

[0052] As Figure 1 shown, a method for evaluating the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network includes the following steps:

[0053] S1: Collect the voltage, phase angle, actual active power, and actual reactive power of each node in the actual distribution substation area, and calculate the real part and imaginary part of the voltage of each node.

[0054] The real part and imaginary part of the voltage of each node are expressed as:

[0055]

[0056] In the formula, i = 1,..., n, where n is the number of nodes in the distribution network, |v i | is the voltage amplitude at the i-th node, and θi The voltage phase angle difference between the i-th node and the reference node (the connection node of the distribution network bus), u i and ω i are the real part and the imaginary part of the voltage phasor at the i-th node respectively.

[0057] S2: Use the real part and the imaginary part of the voltage of each node as the input of a two-layer neural network. Pre-train the two-layer neural network, initialize the weight matrix by using the sparsity and symmetry of the admittance matrix, and set physical constraints, so as to obtain the reconstructed power flow model.

[0058] The structure of the two-layer neural network includes an input layer, a multiplication layer and a linear layer:

[0059] Input layer: Receive the real part and the imaginary part data of the voltage of each node;

[0060] Multiplication layer: Extract the features of the input data;

[0061] Linear layer: Perform weighted summation on the output of the multiplication layer to predict the output of the power flow model.

[0062] The specific process of step S2 is as follows:

[0063] Express the mapping fS1 of the two-layer neural network as fS1(x) = W1δ1(x), where δ1(x) represents the mapping from the input x to the multiplication term, and W1 is the weight matrix of the linear layer; for fS1, use the mean square error MSE of the output result as the objective function, and add the l1 norm of W1 as the regularization term to enhance sparsity:

[0064]

[0065] In the formula, L MSE is the mean square error loss function, which is used to measure the error between the model prediction value and the true value; represents the normalization factor for the number of samples N, which is used to calculate the average error; y n is the true value of the n-th sample; x n represents the n-th sample, that is, the real part and the imaginary part of the voltage of the n-th node; ||·|| p is the l p norm, and λ1 is a positive constant;

[0066] Pre-train the two-layer neural network, set the elements corresponding to the nodes that have no relationship in W1 to 0, and then constrain the elements in W1 according to the symmetry of the admittance matrix, so that u i u p and v i v p The corresponding elements in W1 are equal, u i u pcorresponding to u p u i The corresponding elements in W1 are equal.

[0067] S3: Establish a power flow solver model based on the multi-layer perceptron MLP, and input the active power and reactive power of each node to predict the real and imaginary parts of the voltage.

[0068] The power flow solver model based on the multi-layer perceptron MLP is as follows: a single-hidden-layer MLP for voltage prediction in an n-node distribution system. The input layer of the power flow solver model contains 2n linear neurons, and the hidden layer and output layer contain m and 2n neurons respectively; the indices r, j, and k are used to indicate the neurons in the input layer, hidden layer, and output layer respectively, where r = 1, 2,..., 2n, j = 1, 2,..., m, k = 1, 2,..., 2n; in the forward propagation, the result of multiplying the weighted vector by the input vector is added to the bias vector as the input of the transfer function, and the output of the transfer function is the input of the next layer; each layer has a weighted vector and a bias vector. In the forward propagation, for the input X = (x1, x2,..., x 2n ) T When, where x 2n is the characteristic element of the 2nth input sample, and the characteristic elements include the actual active power P and the actual reactive power Q. The superscript T represents the transpose. The input layer does not use an activation function, and the calculation function f1(x r ) of the rth neuron x r in the input layer is as follows:

[0069] f1(x r ) = x r , r = 1,..., 2n

[0070] The hidden layer uses the sigmoid activation function, and the calculation of the jth neuron z j in the hidden layer is as follows:

[0071]

[0072] In the formula, is the weighted vector between the input layer and the hidden layer, is the bias vector between the input layer and the hidden layer, and sigmoid(z j ) is the activation of z j ;

[0073] The output o k of the kth neuron in the output layer is:

[0074] o k = f1(y k ) = y k , k = 1, 2,..., 2n

[0075]

[0076] where y k is the weighted sum of sigmoid(z j ), is the weighted vector between the hidden layer and the output layer, is the bias vector between the hidden layer and the output layer.

[0077] The power flow solver model is trained using the standard backpropagation algorithm. By the chain rule, the gradients of the loss function with respect to the weights and biases are calculated layer by layer. The core process includes: first, calculate the gradient of the loss with respect to the activation values from the output layer, and combine it with the derivative of the activation function to obtain the error term; then, layer by layer, backpropagate the error through the weight matrix to the hidden layer, while calculating the gradients of the weights and biases of each layer; finally, use gradient descent to update the weight and bias parameters of the model according to the learning rate, so as to minimize the loss function E:

[0078]

[0079] where E is the root mean square error corresponding to the input x, and d k is the expected output of neuron k in the output layer corresponding to o k .

[0080] During the training process of the power flow solver model, Gaussian noise is added to the input data. The expression of the noise input is as follows:

[0081] s (n) = x (n) + d

[0082] where x (n) is the training input data without error, s (n) is the input data with error, and d is a random variable.

[0083] S4: Input the real part and imaginary part of the voltage predicted by the power flow solver model into the reconstructed power flow model to obtain the active power and reactive power. Compare the obtained active power and reactive power with the actual active power and actual reactive power collected in step S1, and use the comparison result as a regularization term to add to the power flow solver model to obtain the trained power flow solver model.

[0084] In step S4, a regularizer is introduced through multi-task learning. The regularizer is inspired by a supervised autoencoder and consists of two parts: an encoder and a decoder, which focus on the inverse problem and the forward problem respectively: power flow solution and power flow modeling. The encoder part is an MLP based on noise injection, which focuses on solving the power flow problem by learning the mapping from power flow input to bus voltage, and ensures the network's resistance to noise by introducing random noise.

[0085] The decoder part is a neural network that performs an auxiliary task, i.e., reconstructing the power flow model, mapping the bus voltage to power injection, and the decoder output is y PQ :

[0086] y PQ = f de (y uω ; W de , b de )

[0087] where f de is the calculation function, y uω is the output of the MLP encoder, W de is the weighted vector between the encoder and the decoder in the calculation function, and b de is the bias vector of the weighted vector between the encoder and the decoder;

[0088] The total loss function is the weighted sum of the voltage prediction loss generated by the encoder and the power reconstruction loss generated by the decoder. Let Δμω = [μ; ω] - y μω , ΔPQ = [P; Q] - y PQ , where Δμω is the voltage prediction loss, [μ; ω] is the actual voltage, y μω is the voltage generated by the encoder, ΔPQ is the power reconstruction loss, and y PQ is the power generated by the decoder. The total loss E 总 of a single training sample is:

[0089] E 总 = α MLP ζ(Δuω) + α SNN ζ(ΔPQ)

[0090] where ξ() represents the mean square error function, and α MLP is the weight of the supervised loss (voltage prediction error), and α SNN is the weight of the unsupervised loss (power reconstruction error). α MLP and α SNN are hyperparameters that balance the encoder and decoder parts respectively.

[0091] Train a power flow solver model to find the parameters that minimize the total loss function; when training the power flow solver model, the error of the voltage prediction layer is:

[0092]

[0093] In the first term Δuω in the formula, the learning process of the MLP encoder is directed to the parameters that are effective for minimizing the bus voltage prediction error. This is the main training objective. However, training the MLP only based on the supervised loss is an underconstrained problem and will find parameters that overfit the data and cannot generalize well. Under the guidance of the second term Among the effective parameters that generate voltage estimates, the learning algorithm preferentially selects the parameters that generate small power mismatches. In this way, the combination of these two types of losses generates a power flow solver with physical consistency.

[0094] S5: Generate multiple photovoltaic deployment scenarios through the Monte Carlo simulation method. Input the actual active power and actual reactive power of each node under all photovoltaic deployment scenarios into the trained power flow solver model to predict the voltage of each node, and screen and statistically obtain the maximum admissible capacity of the photovoltaic in the safety domain.

[0095] The process of screening and statistically obtaining the maximum admissible capacity of the photovoltaic in the safety domain is as follows: judge whether the predicted voltage of each node is within the set voltage safety domain range. If it is within the safety domain range, increase the photovoltaic quantity of the randomly generated photovoltaic scenario deployment. If it is not within the set voltage safety domain, save the total photovoltaic value of the previous scheme. When the total number of photovoltaic deployment schemes is reached, output the maximum result.

Claims

1. An evaluation method for the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network, characterized in that It includes the following steps: S1: Collect the voltage, phase angle, actual active power, and actual reactive power of each node in the actual distribution substation area, and calculate the real part and imaginary part of the voltage of each node; S2: Use the real part and imaginary part of the voltage of each node as the input of a two-layer neural network, pre-train the two-layer neural network, initialize the weight matrix using the sparsity and symmetry of the admittance matrix, and set physical constraints to obtain a reconstructed power flow model; S3: Establish a power flow solver model based on the multi-layer perceptron MLP, and input the active power and reactive power of each node to predict the real part and imaginary part of the voltage; S4: Input the real part and imaginary part of the voltage predicted by the power flow solver model into the reconstructed power flow model to obtain the active power and reactive power. Compare the obtained active power and reactive power with the actual active power and actual reactive power collected in step S1, and add the comparison result as a regularization term to the power flow solver model to obtain a trained power flow solver model; S5: Generate multiple photovoltaic deployment scenarios through the Monte Carlo simulation method, input the actual active power and actual reactive power of each node under all photovoltaic deployment scenarios into the trained power flow solver model, predict the voltage of each node, and screen and statistically obtain the maximum admissible capacity of photovoltaic in the safety domain.

2. The evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks according to claim 1, wherein In step S1, the real part and imaginary part of the voltage of each node are expressed as: where \(i = 1,\ldots,n\), \(n\) is the number of nodes in the distribution network, \(|v i |\) is the voltage amplitude at the \(i\)-th node, \(\theta i \) is the voltage phase angle difference between the \(i\)-th node and the reference node, \(u i \) and \(\omega i \) are the real and imaginary parts of the voltage phasor at the \(i\)-th node, respectively.

3. The evaluation method for the maximum safe access capacity of photovoltaic power in the medium-voltage distribution network according to claim 2, characterized in that, In step S2, the structure of the two-layer neural network includes an input layer, a multiplication layer, and a linear layer: Input layer: Receive the real part and imaginary part data of the voltage of each node; Multiplication layer: Extract the features of the input data; Linear layer: Perform weighted summation on the output of the multiplication layer to predict the output of the power flow model.

4. The evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks according to claim 2, characterized in that The specific process of step S2 is as follows: Express the mapping fS1 of the two-layer neural network as fS1(x) = W1δ1(x), where δ1(x) represents the mapping from the input x to the multiplication term, and W1 is the weight matrix of the linear layer; for fS1, use the mean square error MSE of the output result as the objective function, and add the l1 norm of W1 as a regularization term to enhance sparsity: where L MSE is the mean squared error loss function, which is used to measure the error between the predicted value and the true value of the model; represents the normalization factor for the number of samples N and is used to calculate the average error; y n is the true value of the nth sample; x n represents the nth sample, that is, the real and imaginary parts of the nth node voltage; ||·|| p is the l p norm, and λ1 is a positive constant; Pre-train the two-layer neural network, set the elements corresponding to the nodes without relationships in W1 to 0, and then constrain the elements in W1 according to the symmetry of the admittance matrix, so that u i u p is equal to the corresponding element of v i v p in W1, and u i u p is equal to the corresponding element of u p u i in W1.

5. The evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks according to claim 4, wherein In step S3, the power flow solver model based on the multi-layer perceptron MLP is: a single-hidden-layer MLP for voltage prediction in an n-node distribution system. The input layer of the power flow solver model contains 2n linear neurons, and the hidden layer and output layer contain m and 2n neurons respectively; the indices r, j, and k are used to indicate the neurons in the input layer, hidden layer, and output layer respectively, where r = 1, 2, …, 2n, j = 1, 2, …, m, k = 1, 2, …, 2n; in forward propagation, the result of multiplying the weighted vector by the input vector is added to the bias vector as the input to the transfer function, and the output of the transfer function is the input to the next layer; each layer has a weighted vector and a bias vector. In forward propagation, for the input X = (x1, x2, …, x 2n ) T when, where x 2n is the feature element of the 2nth input sample, and the feature element Including the actual active power P and the actual reactive power Q, where the superscript T represents transpose. The input layer does not use an activation function, and the r-th neuron x in the input layer r The calculation function f1(x r ) is as follows: f1(x r ) = x r , r = 1, ..., 2n The hidden layer uses the sigmoid activation function. The calculation of the \(z\) value of the \(j\)-th neuron in the hidden layer is as follows: j is as follows: Wherein, is the weighted vector between the input layer and the hidden layer, is the bias vector between the input layer and the hidden layer, and sigmoid(z j ) is the activation of z j . The output o of the k-th neuron in the output layer k is as follows: o k = f1(y k ) = y k , k = 1, 2, ..., 2n where y k is the weighted sum of sigmoid(z j ), is the weighted vector between the hidden layer and the output layer, is the bias vector between the hidden layer and the output layer.

6. The evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks according to claim 5, characterized in that, In step S3, the power flow solver model is trained using the standard backpropagation algorithm. Calculate the gradients of the loss function with respect to the weights and biases layer by layer through the chain rule. The core process includes: first calculate the gradient of the loss with respect to the activation value from the output layer, and combine the derivative of the activation function to obtain the error term; then layer by layer backpropagate the error through the weight matrix to the hidden layer, and calculate the gradients of the weights and biases of each layer at the same time; finally, use gradient descent to update the weight and bias parameters of the model according to the learning rate to minimize the loss function E: where E is the root mean square error corresponding to the input x, and d k is the expected output of neuron k in the output layer corresponding to o k .

7. The evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks according to claim 6, characterized in that, In step S3, during the training process of the power flow solver model, Gaussian noise is added to the input data, and the expression of the noise input is as follows: s (n) = x (n) + d where x (n) is the training input data without error, s (n) is the input data with error, and d is a random variable.

8. The evaluation method for the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network according to claim 7, characterized in that In step S4, a regularizer is introduced through multi-task learning. The regularizer is inspired by a supervised autoencoder and consists of two parts: an encoder and a decoder, which focus on the inverse problem and the forward problem respectively: power flow solution and power flow modeling. The encoder part is an MLP based on noise injection, which focuses on solving the power flow problem by learning the mapping from power flow input to bus voltage, and ensures the network's resistance to noise by introducing random noise. The decoder part is a neural network that performs an auxiliary task, namely reconstructing the power flow model, mapping the bus voltages to power injections, and the decoder output is y PQ : y PQ = f de (y uω ; W de , b de ) where f de is a calculation function, y uω is the output of the MLP encoder, W de is the weighted vector between the encoder and the decoder in the calculation function, b de is the bias vector of the weighted vector between the encoder and the decoder; The total loss function is a weighted sum of the voltage prediction loss generated by the encoder and the power reconstruction loss generated by the decoder. Let Δμω = [μ; ω] - y μω , ΔPQ = [P; Q] - y PQ , where Δμω is the voltage prediction loss, [μ; ω] is the actual voltage, and y μω is the voltage generated by the encoder, ΔPQ is the power reconstruction loss, and y PQ is the power generated by the decoder. The total loss E of a single training sample 总 is as follows: E 总 = α MLP ζ(Δuω) + α SNN ζ(ΔPQ) where ξ() represents the mean square error function, and α MLP is the weight of the supervised loss, α SNN is the weight of the unsupervised loss, α MLP and α SNN are hyperparameters for balancing the encoder and decoder parts respectively.

9. The evaluation method for the maximum safe access capacity of photovoltaic power in medium-voltage distribution networks according to claim 8, wherein, In step S4, the power flow solver model is trained to find the parameters that minimize the total loss function. When training the power flow solver model, the error of the voltage prediction layer is: In the formula, the first term Δuω directs the learning process of the MLP encoder to the parameters effective for minimizing the bus voltage prediction error. Under the guidance of the second term , among the parameters effective in generating the voltage estimate, the learning algorithm preferentially selects the parameters that generate small power mismatches.

10. The evaluation method for the maximum safe access capacity of photovoltaic power in a medium-voltage distribution network according to claim 9, characterized in that, In step S5, the process of screening and statistically obtaining the maximum admissible capacity of PV in the security region is as follows: determine whether the predicted node voltages are within the set voltage security region. If they are within the security region, increase the PV amount randomly generated for the PV scenario deployment. If they are not within the set voltage security region, save the total PV value of the previous scheme. When the total number of PV deployment schemes is reached, output the maximum result.

Images