Power system risk assessment method based on dense connection convolutional neural network

Abstract

The invention discloses a power system risk assessment method based on a dense connection convolutional neural network, and the method comprises the steps: carrying out the sampling of the states of a generator set and a power transmission line through employing a forced outage rate vk-1, and combining the element state obtained through sampling with a system load level, so as to obtain a system state Si; according to the current system state Si, obtaining a possible load reduction amount LCi; a risk indicator function and a likelihood ratio are calculated, the forced outage rate vk is updated and corrected, a fault sample in cross entropy iteration is obtained, and k is the number of iterations; and for the current system state Si and the possible load reduction quantity LCi, constructing input and corresponding labels of neural network training data by using the operation parameter matrix. The invention provides an innovative architecture, namely a dense connection convolutional neural network, and aims to overcome the gradient disappearance problem and enhance generalization ability and prediction precision. In addition, a method based on cross entropy is used in DenseNet training data generation to improve the representativeness of the training data.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system risk assessment, and the present invention relates to a power system risk assessment method based on a densely connected convolutional neural network. Background Art

[0002] In the field of power system risk assessment, there are two main methods: analytical method and simulation method. The analytical method is more efficient, but the number of states increases rapidly as the number of system components increases. It is suitable for small systems or identifying critical faults. Simulation methods, especially Monte Carlo simulation, estimate risk indicators by randomly sampling possible system states, but its accuracy is directly related to the sample size and requires a lot of computing resources to obtain accurate assessments.

[0003] In recent years, advanced machine learning techniques have been introduced into power system risk assessment to balance computational efficiency and accuracy. These methods use machine learning to study the complex nonlinear relationship between system risk level and system operating parameters. Among them, convolutional neural networks (CNNs) have attracted widespread attention due to their excellent performance in image processing and time series analysis. However, standard CNN models often encounter the problem of gradient vanishing in deep network architectures, which limits the network depth and learning ability, and affects the generalization ability of the model. Summary of the invention

[0004] The purpose of the present invention is to provide a power system risk assessment method based on densely connected convolutional neural networks, and an innovative architecture is proposed: densely connected convolutional neural networks (DenseNet), which aims to overcome the gradient vanishing problem and enhance generalization ability and prediction accuracy. In addition, the present invention also uses a cross-entropy-based method in DenseNet training data generation to improve the representativeness of training data.

[0005] The technical solution to achieve the purpose of the present invention is:

[0006] A method for constructing neural network training data based on cross entropy, comprising the following steps:

[0007] S01: Use forced outage rate v k-1 The generator and transmission line states are sampled, and the sampled component states are combined with the system load level to obtain the system state S i ;

[0008] S02: According to the current system status S i , obtain the possible load reduction LC i ;

[0009] S03: Calculate the risk indicator function and likelihood ratio, and update the corrected forced outage rate v k, get the fault samples in the cross entropy iteration, k is the number of iterations;

[0010] S04: For the current system state S i and possible load reduction LC i , use the running parameter matrix to construct the input of the neural network training data and the corresponding labels.

[0011] In the preferred technical solution, the risk indicator function H(S) is calculated in step S03. i ) and the likelihood ratio W(S i ,u,v k ) is:

[0012]

[0013]

[0014] Among them, N GL is the total number of generators and transmission lines in the power system, S i,j is the jth component state in the i-th system state sample, where S i,j =1 indicates normal state, S i,j = 0 indicates a fault state, u is a vector whose elements are the forced outage rates of generators and transmission lines, u j is the jth element, v k-1,j is the jth element of the modified forced outage rate during the kth iteration.

[0015] In the preferred technical solution, in step S03, the modified forced outage rate v is updated. k For each element j of

[0016]

[0017] Among them, α is the update factor, N CE is the number of cross entropy samples in each iteration.

[0018] In the preferred technical solution, the operating parameter matrix in step S04 includes a bus admittance matrix:

[0019]

[0020] The available power matrix on the busbar is:

[0021]

[0022] Bus load power matrix:

[0023]

[0024] Inter-bus transmission capacity matrix:

[0025]

[0026] Where n represents the number of system buses, B mp represents the mutual admittance between the mth busbar and the pth busbar (m≠p), B mm represents the self-admittance of the mth busbar, GC m represents the total available generating capacity on the mth bus, L m represents the total load demand on the mth bus, TC mp Represents the maximum available transmission capacity between the mth bus and the pth bus.

[0027] The present invention also discloses a power system risk assessment method based on a densely connected convolutional neural network, comprising the following steps:

[0028] Build densely connected convolutional neural networks;

[0029] Constructing input data for a neural network using any of the methods described above;

[0030] The relationship between the load reduction amount and the system state variables of the power system is obtained through the trained densely connected convolutional neural network;

[0031] Analyze system status and calculate risk indicator functions.

[0032] In the preferred technical solution, the densely connected convolutional neural network includes l layers, each layer performs a transformation H l (·), where l represents the number of layers, H1 is a composite operation of three consecutive operations: a convolution operation followed by batch normalization and a leaky corrected linear activation function, H2 and H4 perform maximum pooling operations, H3 is a dense block, H5 is a flattening operation that flattens the feature map X4 into a one-dimensional vector data X5, and H6 includes a fully connected operation and a leaky corrected linear activation function, and finally obtains a scalar output X6, which represents the possible load reduction corresponding to the current system state.

[0033] In the preferred technical solution, the dense block includes 3 layers, each layer performs a composite operation: first batch normalization, then modified linear activation function and convolution operation. In the dense block, the input not only enters the first convolution operation, but also serves as the input of each subsequent convolution operation. The output of each convolution will also serve as the input of the next convolution operation and each subsequent convolution operation in the block.

[0034] In the preferred technical solution, the function of the leakage corrected linear activation function LReLU is:

[0035] LReLU(x)=max(αx,x)

[0036] Among them, α is a positive number less than 1. When α is equal to zero, LReLU degenerates into ReLU.

[0037] In the preferred technical solution, the load reduction amount of the system state predicted by the neural network will be used to calculate the expected shortage power EDNS and the load loss probability LOLP of the typical risk indicators of the power system; when calculating the LOLP indicator, the system state with a load reduction amount greater than 0 is regarded as a fault state, otherwise it is a normal state;

[0038] To evaluate the predictive performance of the model, the performance indicators of the neural network model are as follows:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] Wherein, MAE is the mean absolute error, MSE is the mean square error, ACC is the accuracy, PR is the precision, RC is the recall, F1 is the F1 score, N is the total number of predicted system state samples, x is the predicted load reduction, y is the actual load reduction, and the superscript i is the i-th system state sample sampled; N = N TP +N TN +N FP +N FN , N TP N represents the number of samples that are predicted to be in a fault state and are actually in a fault state. TN Represents the number of samples that are predicted to be normal and are actually normal, N FP N represents the number of samples that are predicted to be in a fault state but are actually in a normal state. FN Indicates the number of samples that were predicted to be in a normal state but were actually in a fault state.

[0046] The present invention further discloses a computer storage medium on which a computer program is stored. When the computer program is executed, any of the above-mentioned methods for constructing neural network training data based on cross entropy is implemented.

[0047] Compared with the prior art, the present invention has the following significant advantages:

[0048] 1. Improve prediction accuracy: By introducing the DenseNet architecture, the gradient vanishing problem is effectively alleviated, and the learning ability of the deep network is enhanced, thereby improving the prediction accuracy of power system risk assessment.

[0049] 2. Enhanced generalization ability: DenseNet improves the generalization ability of the model by reusing the feature maps of the early layers, so that it can still maintain high prediction performance when facing new or unseen system states.

[0050] 3. Improve the quality of training data: The cross-entropy-based training data generation method is adopted to increase the proportion of valid samples in the training data and improve the representativeness and robustness of model training.

[0051] 4. Improve analysis efficiency: Using neural networks for state analysis improves the analysis efficiency of power system risk assessment and promotes system optimization and planning.

[0052] In general, the present invention significantly improves the accuracy and efficiency of power system risk assessment through the innovative DenseNet architecture and optimized training data generation method, providing strong support for the safe and reliable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A flowchart of a method for constructing neural network training data based on cross entropy according to a preferred embodiment;

[0054] Figure 2 This is a diagram of the traditional convolutional neural network (CNN) structure used for risk assessment status analysis;

[0055] Figure 3 This is a structural diagram of a densely connected convolutional neural network (DenseNet) for risk assessment status analysis in a preferred embodiment. DETAILED DESCRIPTION

[0056] The principle of the present invention is: an innovative architecture is proposed: densely connected convolutional neural network (DenseNet), which aims to overcome the gradient vanishing problem and enhance generalization ability and prediction accuracy. In addition, the present invention also uses a cross-entropy-based method for DenseNet training data generation to improve the representativeness and effectiveness of training data, thereby improving the accuracy of risk assessment.

[0057] Embodiment 1:

[0058] like Figure 1 As shown, a method for constructing neural network training data based on cross entropy includes the following steps:

[0059] S01: Use forced outage rate v k-1The generator and transmission line states are sampled, and the sampled component states are combined with the system load level to obtain the system state S i ;

[0060] S02: According to the current system status S i , obtain the possible load reduction LC i ;

[0061] S03: Calculate the risk indicator function and likelihood ratio, and update the corrected forced outage rate v k , get the fault samples in the cross entropy iteration, k is the number of iterations;

[0062] S04: For the current system state S i and possible load reduction LC i , use the running parameter matrix to construct the input of the neural network training data and the corresponding labels.

[0063] In a preferred embodiment, the risk indicator function H(S) is calculated in step S03. i ) and the likelihood ratio W(S i ,u,v k ) is:

[0064]

[0065]

[0066] Among them, N GL is the total number of generators and transmission lines in the power system, S i,j is the jth component state in the i-th system state sample, where S i,j =1 indicates normal state, S i,j = 0 indicates a fault state, u is a vector whose elements are the forced outage rates of generators and transmission lines, u j is the jth element, v k-1,j is the jth element of the modified forced outage rate during the kth iteration.

[0067] In a preferred embodiment, in step S03, the modified forced outage rate v is updated. k For each element j of

[0068]

[0069] Among them, α is the update factor, N CE is the number of cross entropy samples in each iteration.

[0070] In a preferred embodiment, the operating parameter matrix in step S04 includes a bus admittance matrix:

[0071]

[0072] The available power matrix on the busbar is:

[0073]

[0074] Bus load power matrix:

[0075]

[0076] Inter-bus transmission capacity matrix:

[0077]

[0078] Where n represents the number of system buses, B mp represents the mutual admittance between the mth busbar and the pth busbar (m≠p), B mm represents the self-admittance of the mth busbar, GC m represents the total available generating capacity on the mth bus, L m represents the total load demand on the mth bus, TC mp Represents the maximum available transmission capacity between the mth bus and the pth bus.

[0079] Another embodiment is a computer storage medium having a computer program stored thereon, wherein the computer program, when executed, implements any of the above-mentioned methods for constructing neural network training data based on cross entropy.

[0080] In another embodiment, a method for power system risk assessment based on a densely connected convolutional neural network comprises the following steps:

[0081] Build densely connected convolutional neural networks;

[0082] Constructing input data for a neural network using any of the methods described above;

[0083] The relationship between the load reduction amount and the system state variables of the power system is obtained through the trained densely connected convolutional neural network;

[0084] Analyze system status and calculate risk indicator functions.

[0085] In a preferred embodiment, the densely connected convolutional neural network comprises l layers, each layer performing a transformation H l(·), where l represents the number of layers, H1 is a composite operation of three consecutive operations: a convolution operation followed by batch normalization and a leaky corrected linear activation function, H2 and H4 perform maximum pooling operations, H3 is a dense block, H5 is a flattening operation that flattens the feature map X4 into a one-dimensional vector data X5, and H6 includes a fully connected operation and a leaky corrected linear activation function, and finally obtains a scalar output X6, which represents the possible load reduction corresponding to the current system state.

[0086] In a preferred embodiment, the dense block includes 3 layers, each layer performs a composite operation: first batch normalization, then a modified linear activation function and a convolution operation. In the dense block, the input not only enters the first convolution operation, but also serves as the input of each subsequent convolution operation. The output of each convolution will also serve as the input of the next convolution operation and each subsequent convolution operation in the block.

[0087] In a preferred embodiment, the function of the leakage corrected linear activation function LReLU is:

[0088] LReLU(x)=max(αx,x)

[0089] Among them, α is a positive number less than 1. When α is equal to zero, LReLU degenerates into ReLU.

[0090] In a preferred embodiment, the load reduction amount of the system state predicted by the neural network will be used to calculate the expected shortage power EDNS and the load loss probability LOLP of the typical risk indicators of the power system; when calculating the LOLP indicator, the system state predicted that the load reduction amount is greater than 0 is regarded as a fault state, otherwise it is a normal state;

[0091] To evaluate the predictive performance of the model, the performance indicators of the neural network model are as follows:

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098] Wherein, MAE is the mean absolute error, MSE is the mean square error, ACC is the accuracy, PR is the precision, RC is the recall, F1 is the F1 score, N is the total number of predicted system state samples, x is the predicted load reduction, y is the actual load reduction, and the superscript i is the i-th system state sample sampled; N = N TP +N TN +N FP +N FN , N TP N represents the number of samples that are predicted to be in a fault state and are actually in a fault state. TN Represents the number of samples that are predicted to be normal and are actually normal, N FP N represents the number of samples that are predicted to be in a fault state but are actually in a normal state. FN Indicates the number of samples that were predicted to be in a normal state but were actually in a fault state.

[0099] Specifically, the following describes a power system risk assessment method based on a densely connected convolutional neural network using a specific example, as follows:

[0100] A. Neural Network Architecture

[0101] The traditional convolutional neural network (CNN) and the densely connected convolutional neural network (DenseNet) proposed in this paper are used to analyze the risk status of the power system, respectively. Figure 2 and Figure 3 As shown in Figure 2, the two models process the same input data: node admittance matrix, node available generation power matrix, node load demand matrix and inter-node transmission capacity matrix, see formulas (8)-(11) for details. Their final output is the possible load reduction amount of the current system state.

[0102] For ease of understanding, define the system operating parameter input of the convolutional network processing as X0. Assume that the convolutional network consists of l layers, each layer performs a transformation H l (·), where l represents the index of the layer. l (·) can be a composite operation, including traditional convolution (Conv) and rectified linear activation function (ReLU) operations. The output of the lth layer is represented by X l .

[0103] exist Figure 2In Figure 1, H1 and H3 are composite operations consisting of two operations: a convolution operation followed by a ReLU activation. The convolution uses 64 kernels, each of size 3×3. H2 and H4 perform 2×2 max pooling operations to reduce the size of the feature map. H5 is a flattening operation that flattens the feature map X4 into a one-dimensional vector data X5. X5 is then processed by H6, which includes a fully connected operation and a ReLU activation, and finally obtains a scalar output X6, which represents the possible load reduction corresponding to the current system state.

[0104] Figure 3 The architecture is roughly the same as Figure 2 Similar, but different on H1, H3 and H6. Figure 3 In the H3 dense block, the input not only enters the first convolution operation, but also serves as the input of each subsequent convolution operation. The output of each convolution is also used as the input of the next convolution operation and each subsequent convolution operation in the block. This reuse mechanism not only helps feature reuse and solves the problem of feature masking, but also alleviates the problem of gradient disappearance during back propagation. To ensure the comparability of the two models, the convolution kernel of H3 in the traditional CNN is set to 64, generating 64 new feature maps. In H3 of DenseNet, there are three convolution operations: the first and third convolutions use 16 kernels each, and the second uses 32 kernels, and the total number of new feature maps is also 64.

[0105] B. Activation Function

[0106] The traditional ReLU activation function has many advantages, including high computational efficiency and the ability to solve the gradient vanishing problem. However, it also has a significant disadvantage: the potential risk of neuron death. This phenomenon occurs during training when some neurons continuously output zero, making them inactive and unable to contribute to the output of the network. When excessive updates of weights push the input of neurons into the negative domain during backpropagation, it causes the output of these neurons to be zero forever. Since the gradient of ReLU is zero in the case of negative output, the weights associated with these neurons will no longer be updated, effectively silencing them permanently. In order to deal with neuron death caused by overly aggressive weight updates, the activation function can be adjusted to maintain a slight gradient at negative inputs, thereby achieving weight adjustment.

[0107] The activation function Leaky ReLU (LReLU) can solve this problem well. The function is defined as follows:

[0108] LReLU (x) = max(αx,x) (1)

[0109] where α is a positive number less than 1 (e.g., 10 -5 ), when α is equal to zero, LReLU degenerates into ReLU.

[0110] C. Model Indicators

[0111] The load reduction amount predicted by the neural network will be used to calculate the typical risk indicators of the power system, EDNS (Expected Shortage Power) and LOLP (Loss of Load Probability). When calculating the LOLP indicator, the system state with a load reduction greater than 0 is considered a fault state, otherwise it is a normal state.

[0112] Mean absolute error (MAE) and mean square error (MSE) can be used to evaluate the prediction performance of the model. Classification performance is evaluated using accuracy (ACC), precision (PR), recall (RC), and F1 score (F1). ACC is the ratio of the number of correctly predicted states to the total number of predicted states. PR indicates how many of the predicted fault states are actually fault states. RC measures the proportion of actual fault states that the model correctly identifies. The F1 score provides a single metric that balances PR and RC, which is particularly useful when the class distribution is unbalanced. The performance indicators of the neural network model are as follows:

[0113]

[0114] Where N represents the total number of predicted system state samples, x represents the predicted load reduction, y represents the actual load reduction, and the superscript i represents the i-th system state sample sampled; N = N TP +N TN +N FP +N FN , N TP N represents the number of samples that are predicted to be in a fault state and are actually in a fault state. TN Represents the number of samples that are predicted to be normal and are actually normal, N FP N represents the number of samples that are predicted to be in a fault state but are actually in a normal state. FN Indicates the number of samples that were predicted to be in a normal state but were actually in a fault state.

[0115] DenseNet training data construction method based on cross entropy:

[0116] When the state of the system components changes, the system topology and various operating constraints may change accordingly, thus affecting the system's ability to resist risks. This paper uses four matrices (8) to (11) as the input data of the neural network. These matrices are all n×n square matrices, where n represents the number of system buses. They represent the four different channels of the convolutional neural network input data.

[0117]

[0118]

[0119] Among them, B mp represents the mutual admittance between the mth bus and the pth bus (m≠p), B mm represents the self-admittance of the mth bus; GC m represents the total available power generation on the mth bus; L m represents the total load demand on the mth bus; TC mp It represents the maximum available transmission capacity between the mth bus and the pth bus, which depends on whether there are lines between the two busbars and the available capacity of these lines.

[0120] This paper proposes a method for constructing DenseNet training data using cross entropy (CE) technology to enhance the training data set of the neural network model. By applying the CE method, we can preferentially sample component states that have a significant impact on risk indicators, thereby increasing the proportion of valuable system state samples. This method enables the neural network to more accurately learn the relationship between system operating parameters and load reduction. The specific steps are as follows:

[0121] Step 1) First, set the parameters, including the deterministic data and random parameters in the power system, as well as: (a) the current sample number i = 1, the number of cross entropy samples N for each iteration CE , and activate cross entropy update, set flag CE = True; (b) define the failure percentage ρ (e.g. 0.5); (c) set the update factor α (e.g. 0.5); (d) establish the maximum number of samples N max (e.g. 200,000) as the termination condition.

[0122] Step 2) starts from k=1, which indicates the first iteration of updating the forced outage rate of the component. Set v0=u, where v0 represents the initial forced outage rate and u is the forced outage rate vector of the generator set and the transmission line. Initially, the load of the power system is considered to be the peak load level.

[0123] Step 3) Use v k-1The generator and transmission line states are sampled and combined with the power system load level to obtain the system state S i , judge flag CE Is it True? If so, go to step 5, otherwise go to step 4.

[0124] Step 4) According to the system load continuity curve model, the system load level is sampled and the i-th system state sample S is updated accordingly. i .

[0125] Step 5) Use power flow and optimal power flow tools to evaluate the current system state S i , in order to accurately calculate whether load reduction will occur in the current system state and the possible load reduction amount LC i If flag CE If True, continue to step 6; if not, skip to step 9.

[0126] Step 6) Calculate the risk indicator function H(S) according to the following formulas (12)-(13): i ) and the likelihood ratio W(S i ,u,v k-1 ). Check whether the number of system state samples i is N CE If it is a multiple of , continue to step 7, otherwise go to step 9.

[0127]

[0128] Among them, N GL is the total number of generating units and transmission lines, S i,j is the jth component state in the i-th system state sample, where S i,j =1 means the component is in normal state, S i,j =0 indicates that the component is in a fault state.

[0129] Step 7) Update the corrected forced outage rate v using the following formula (14): k For each element in, j=1,2,...,N GL .

[0130]

[0131] Step 8) Update the cross entropy iteration counter k=k+1, and calculate the number of fault state samples in the current cross entropy iteration to the number of cross entropy samples N CE Does the ratio exceed ρ? If so, set the flag CE = False, and the forced outage rate of the component is no longer updated.

[0132] Step 9) For the current system state S iand possible load shedding LC i , combined with formulas (8)-(11), construct and save the input and corresponding output labels for neural network training data.

[0133] Step 10) Determine whether the number of samples i exceeds N max , if yes, stop; otherwise, update the sample number i=i+1 and return to step 3.

[0134] The proposed method was tested using the IEEE RTS-79 system. The method was implemented using Python, and all calculations were performed on an AMD Ryzen 9 5950X CPU and an NVIDIA GeForce RTX 3060 GPU. During the construction of training data, DC power flow and DC optimal power flow tools were used for system state fault analysis and load reduction calculations, and N max During the neural network training process, 10% of the samples (20,000 samples) were designated as the test data set, the training cycle and batch size were 200 and 50 respectively, the optimizer used Adam, and the learning rate was set to 10 -3 .

[0135] A. Different network architectures

[0136] Table 1. Comparison between different neural network architectures

[0137]

[0138]

[0139] Table 1 shows a comparative analysis of the performance metrics of the two models trained on the same dataset. Both models were validated using the same datasets, which were not used during the training phase. By examining regression error metrics such as MSE and MAE, we observe a significant reduction in error rate when transitioning from the traditional CNN architecture to DenseNet. Specifically, MSE and MAE are reduced by 72.8% and 54.5%, respectively.

[0140] In terms of classification performance, both models achieved an accuracy of over 99% and an F1 score of over 97%. However, the DenseNet model showed further improvement over the traditional CNN. It should be emphasized that traditional CNNs often encounter problems such as premature death of neurons during training. In addition, setting an initial learning rate that is too high may lead to oscillations in model performance. Unlike traditional CNNs, the DenseNet model designed for power system risk analysis is not affected by these problems, demonstrating its robustness and efficiency.

[0141] B. IEEE RTS-79 Risk Assessment

[0142] The IEEE RTS-79 system includes 24 buses, 32 generators and 38 transmission lines, with an installed capacity of 3,405 MW and a peak load of 2,850 MW.

[0143] Table 2. IEEE RTS-79 Annual Risk Indicators

[0144]

[0145] Table 2 shows the IEEE RTS-79 annual risk index calculated using various methods. The annual risk index takes into account the impact of the system load curve. EDNS The traditional NSMCS (Non-Sequential Monte Carlo Simulation) method set to 2% is used as a benchmark for performance comparison of other methods.

[0146] The traditional NSMCS method uses 646,000 samples and takes more than 800 seconds to calculate the LOLP equal to 1.2229×10 -3 , the relative error is 2.13%, and the EDNS is 1.4947×10 -1 , with a relative error of 1.16%. The traditional CNN method uses 580,000 samples and takes 55 seconds to calculate the LOLP to be 1.2655×10 -3 , the relative error is 5.69% and the EDNS is 1.6122×10 -1 , with a relative error of 9.11%. In addition, the DenseNet method uses 620,000 samples and takes 115 seconds to calculate the LOLP to be 1.1984×10 -3 , the relative error is 0.08%, and the EDNS is 1.4307×10 -1 , the relative error is 3.17%.

[0147] In Beta EDNS At a threshold of 5%, both the traditional NSMCS method and the proposed DenseNet method produce reasonable risk indicators. However, the traditional CNN method shows a larger error. In addition, the proposed method shows a significant reduction in computation time when compared with the traditional NSMCS method. It should be noted that neural network-based state analysis can predict hundreds or thousands of system states simultaneously, limited only by GPU memory limitations.

[0148] Table 3. IEEE RTS-79 Annualized Risk Indicators

[0149]

[0150] Table 3 shows the EDNSThe IEEE RTS-79 annualized risk index calculated by three different methods is compared when the system load is kept at the peak load level. The risk index obtained by the traditional NSMCS method is used as a reference, and the relative errors of the indicators calculated by the other two neural network-based methods are compared.

[0151] The analysis shows that at a variance coefficient of 1%, the DenseNet method produces more accurate risk indicators than the traditional CNN method. In addition, in the case of the LOLP indicator, both methods show significant errors. However, the error of the DenseNet method is smaller than that of the traditional CNN method.

[0152] The accuracy of the EDNS metric is attributed to the fact that both models discussed in this study are small and use a regression loss function based on MSE, rather than a classification loss function. Despite these differences, the DenseNet architecture overall shows improvements in terms of error reduction.

[0153] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principles of the present invention shall be equivalent replacement modes and shall be included in the protection scope of the present invention.

Claims

1. A method for constructing neural network training data based on cross entropy, characterized in that: The following steps are involved: S01: Use forced outage rate v k-1 The generator and transmission line states are sampled, and the sampled component states are combined with the system load level to obtain the system state S i ; S02: According to the current system status S i , obtain the possible load reduction LC i ; S03: Calculate the risk indicator function and likelihood ratio, and update the corrected forced outage rate v k , get the fault samples in the cross entropy iteration, k is the number of iterations; S04: For the current system state S i and possible load reduction LC i , use the running parameter matrix to construct the input of the neural network training data and the corresponding labels.

2. The method for constructing neural network training data based on cross entropy according to claim 1, characterized in that: In step S03, the risk index function H(S i ) and the likelihood ratio W(S i ,u,v k ) is: Among them, N GL is the total number of generators and transmission lines in the power system, S i,j is the jth component state in the i-th system state sample, where S i,j =1 indicates normal state, S i,j = 0 indicates a fault state, u is a vector whose elements are the forced outage rates of generators and transmission lines, u j is the jth element, v k-1,j is the jth element of the modified forced outage rate during the kth iteration.

3. The method for constructing neural network training data based on cross entropy according to claim 1, characterized in that: In step S03, the modified forced outage rate v is updated k For each element j of Among them, α is the update factor, N CE is the number of cross entropy samples in each iteration.

4. The method for constructing neural network training data based on cross entropy according to claim 1, characterized in that: In step S04, the operating parameter matrix includes the bus admittance matrix: The available power matrix on the busbar is: Bus load power matrix: Inter-bus transmission capacity matrix: Where n represents the number of system buses, B mp represents the mutual admittance between the mth busbar and the pth busbar (m≠p), B mm represents the self-admittance of the mth bus, GC m represents the total available generating capacity on the mth bus, L m represents the total load demand on the mth bus, TC mp Represents the maximum available transmission capacity between the mth bus and the pth bus.

5. A power system risk assessment method based on densely connected convolutional neural network, characterized in that: The following steps are involved: Build a densely connected convolutional neural network; Constructing input data of a neural network using the method described in any one of claims 1 to 4; The relationship between the load reduction amount and the system state variables of the power system is obtained through the trained densely connected convolutional neural network; Analyze system status and calculate risk indicator functions.

6. The power system risk assessment method based on densely connected convolutional neural network according to claim 5 is characterized in that: The densely connected convolutional neural network contains l layers, each of which performs a transformation H l (·), where l represents the number of layers, H1 is a composite operation of three consecutive operations: a convolution operation followed by batch normalization and a leaky corrected linear activation function, H2 and H4 perform maximum pooling operations, H3 is a dense block, H5 is a flattening operation that flattens the feature map X4 into a one-dimensional vector data X5, and H6 includes a fully connected operation and a leaky corrected linear activation function, and finally obtains a scalar output X6, which represents the possible load reduction corresponding to the current system state.

7. The power system risk assessment method based on densely connected convolutional neural network according to claim 5 is characterized in that: The dense block includes 3 layers, each of which performs a composite operation: batch normalization, then a modified linear activation function and a convolution operation. In a dense block, the input not only enters the first convolution operation, but also serves as the input of each subsequent convolution operation. The output of each convolution will also serve as the input of the next convolution operation and each subsequent convolution operation in the block.

8. The power system risk assessment method based on densely connected convolutional neural network according to claim 6 is characterized in that: The function of the leaky corrected linear activation function LReLU is: LReLU(x)=max(αx,x) Among them, α is a positive number less than 1. When α is equal to zero, LReLU degenerates into ReLU.

9. The power system risk assessment method based on densely connected convolutional neural network according to claim 5 is characterized in that: The load reduction of the system state predicted by the neural network will be used to calculate the expected shortage power EDNS and load loss probability LOLP, typical risk indicators of the power system; When calculating the LOLP index, the system state where the predicted load reduction is greater than 0 is considered as a fault state, otherwise it is a normal state; To evaluate the predictive performance of the model, the performance indicators of the neural network model are as follows: Wherein, MAE is the mean absolute error, MSE is the mean square error, ACC is the accuracy, PR is the precision, RC is the recall, F1 is the F1 score, N is the total number of predicted system state samples, x is the predicted load reduction, y is the actual load reduction, and the superscript i is the i-th system state sample sampled; N = N TP +N TN +N FP +N FN , N TP N represents the number of samples that are predicted to be in a fault state and are actually in a fault state. TN Represents the number of samples that are predicted to be normal and are actually normal, N FP N represents the number of samples that are predicted to be in a fault state but are actually in a normal state. FN Indicates the number of samples that were predicted to be in a normal state but were actually in a fault state.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the method for constructing neural network training data based on cross entropy according to any one of claims 1 to 4 is implemented.

Images