Power system risk assessment method and system based on cross entropy convolutional neural network

Abstract

The invention discloses a power system risk assessment method and system based on a cross entropy 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, 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 + 1 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. By introducing a data construction algorithm based on a cross entropy technology, system states which are crucial to risk indexes are automatically identified and preferentially processed, and the proportion of effective samples in a training data set is improved, so that the training quality and prediction precision of the model are remarkably improved.

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 and system based on a cross entropy convolutional neural network. Background Art

[0002] In the field of power system risk assessment, there are two basic algorithms: analytical methods and simulation methods. The analytical method (also known as the state enumeration method) calculates the risk index of the system by enumerating all possible states of the power system and analyzing each state. However, as the number of system components increases, the number of states increases exponentially, making the analytical method impractical when dealing with large-scale systems. Therefore, it is usually applied to risk assessment of small-scale systems or screening of critical fault states. Simulation methods, such as Monte Carlo simulation, estimate risk indicators by randomly sampling possible states of the system. The accuracy of this method is highly dependent on the number of samples, so it faces high computational costs when pursuing high-precision assessments.

[0003] To address these challenges, advanced machine learning techniques have been introduced into power system risk assessment in recent years, aiming to improve the efficiency and accuracy of assessment through data-driven approaches. These methods attempt to find a balance between accuracy and computational speed by using machine learning algorithms to explore the complex nonlinear relationships between risk parameters and their influencing factors. In particular, deep learning techniques, such as convolutional neural networks (CNNs), have attracted attention due to their outstanding performance in areas such as image processing and time series data analysis. In power system risk assessment, CNNs can quickly assess system risk levels by learning patterns and features from historical data, thereby supporting system optimization and scheduling.

[0004] However, CNNs face some challenges when they are actually applied to large-scale power system risk assessment. Actual power systems usually exhibit high reliability, which leads to serious skewness in the training data obtained by random sampling methods and the presence of a large number of zero-valued labels, thus affecting the training effect of CNNs. In addition, traditional CNNs are prone to the vanishing gradient problem when dealing with deep networks, which limits the increase in network depth and the improvement of model learning ability, resulting in limited generalization ability, and the prediction accuracy may drop significantly when facing new or unseen system states. Summary of the invention

[0005] The purpose of the present invention is to provide a power system risk assessment method and system based on cross-entropy convolutional neural network, which automatically identifies and prioritizes system states that are critical to the risk index by introducing a data construction algorithm based on cross-entropy technology, thereby increasing the proportion of valid samples in the training data set, thereby significantly improving the training quality and prediction accuracy of the model.

[0006] The technical solution to achieve the purpose of the present invention is: A method for constructing neural network training data based on cross entropy, comprising the following steps: S01: Use forced outage rate Sample the status of generators and transmission lines, and combine the sampled component status with the system load level to obtain the system status ; S02: According to the current system status , obtain the possible load reduction ; S03: Calculate the risk indicator function and likelihood ratio, and update the corrected forced outage rate , get the fault samples in the cross entropy iteration, is the number of iterations; S04: For the current system status and possible load reduction , use the running parameter matrix to construct the input of the neural network training data and the corresponding labels.

[0007] In the preferred technical solution, the system status is obtained in step S01 Also includes: Randomly sample system load levels based on the load model used and then update , use power flow and optimal power flow tools to analyze the current system status .

[0008] In the preferred technical solution, the risk indicator function is calculated in step S03 Likelihood Ratio The formula is:

[0009]

[0010] in, is the total number of generating units and transmission lines, It is In the system status sample component states, where Indicates success status. Indicates a fault condition. is a vector whose elements are the forced outage rates of generating units and transmission lines, It is elements, The forced outage rate is elements.

[0011] In the preferred technical solution, in step S03, the forced outage rate is updated and corrected. Each element of :

[0012] in, is the update factor, is the number of cross entropy samples in each iteration.

[0013] In the preferred technical solution, the operating parameter matrix in step S04 includes a bus admittance matrix:

[0014] The available power matrix on the busbar is:

[0015] Bus load power matrix:

[0016] Inter-bus transmission capacity matrix:

[0017] Where n represents the number of system buses, Indicates The busbar and Mutual admittance between buses , Indicates The self-admittance of the busbar, Indicates The total available generating capacity on the busbars, Indicates The total load demand on each bus, Indicates The busbar and The maximum available transmission capacity between the buses.

[0018] The present invention also discloses a power system risk assessment method based on a cross entropy convolutional neural network, comprising the following steps: Construct a convolutional neural network based on cross entropy; Constructing input data for a neural network using any of the methods described above; The relationship between the load reduction amount and the system state variables of the power system is obtained through the trained cross-entropy-based convolutional neural network; Analyze system status and calculate risk indicator functions.

[0019] In the preferred technical solution, the cross entropy-based convolutional neural network includes Layers, each layer implements a transformation ,in Indicates the number of layers, and is a composite function of three consecutive operations: convolution, followed by batch normalization and rectified linear units. It is a flattening operation that transforms the two-dimensional feature map Reconstruct into one-dimensional data , and Both are composed of fully connected layers and rectified linear units. The output is the load reduction amount.

[0020] In the preferred technical solution, the batch normalization normalizes the input by zero centering and scaling, and each layer uses two parameters to adapt, and the input is standardized using the mean and standard deviation of the current mini-batch, as follows:

[0021]

[0022]

[0023]

[0024] in, is the input mean vector, evaluated over the entire mini-batch, is the number of instances in the mini-batch, is the input standard deviation vector, evaluated over the entire mini-batch, is the zero-centered and normalized input vector of the instance, is a decimal number that avoids division by zero, is the output scaling parameter vector of the layer, represents vector element-wise multiplication, is the output shift parameter vector of the layer, is the output of batch normalization, which is a rescaled and shifted version of the input.

[0025] For the inference phase, the mean and variance are no longer recalculated for each mini-batch. Instead, the values ​​estimated by moving average during training are used. The mathematical formula for the adjustment is:

[0026] in, is the mean vector used for prediction, with an initial value of 0. is the variance vector used for prediction, with an initial value of 1. is the momentum scalar, after each batch of training, and Will be updated when the model training is completed. and The value of is determined and used directly in the prediction phase.

[0027] The present invention further discloses a power system risk assessment system based on a cross entropy convolutional neural network, comprising: Convolutional neural network building module, building a convolutional neural network based on cross entropy; An input data acquisition module, which constructs input data of a neural network using any of the methods described above; The prediction module obtains the relationship between the load reduction of the power system and the system state variables through the trained cross-entropy-based convolutional neural network; The risk assessment module analyzes the system status and calculates the risk indicator function.

[0028] 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.

[0029] Compared with the prior art, the present invention has the following significant advantages: 1. Improve the validity of data samples: Traditional random sampling methods can cause serious deviations in the training data set of high-reliability power systems, including a large number of zero-value labels, which affects the training effect of the model. The present invention introduces a data construction algorithm based on cross-entropy technology to automatically identify and prioritize system states that are critical to the risk index, thereby increasing the proportion of valid samples in the training data set and significantly improving the training quality and prediction accuracy of the model.

[0030] 2. Reduce the gradient disappearance problem: Traditional convolutional neural networks are prone to the vanishing gradient problem when processing deep networks, which limits the increase in network depth and the improvement of model learning ability. The present invention introduces a batch normalization (BN) layer in the CNN architecture, which effectively alleviates the vanishing gradient problem and enhances the generalization ability of the model, significantly improving its prediction accuracy when facing new or unseen system states.

[0031] 3. Improve the accuracy and efficiency of risk assessment: The present invention significantly improves the accuracy of risk assessment through an improved CNN model and an optimized data sample construction method. Specifically, the accuracy of load reduction prediction has been significantly improved. In addition, since deep learning technology can quickly analyze and process a large amount of historical data, the method of the present invention also significantly improves the evaluation efficiency and can complete the risk assessment of large-scale power systems more quickly.

[0032] 4. Reduce computational costs: Compared with traditional simulation methods (such as Monte Carlo simulation), the present invention effectively balances evaluation accuracy and calculation speed through a data-driven approach. The improved CNN can achieve high-precision evaluation results with fewer samples, thereby significantly reducing the calculation cost, especially when performing high-precision evaluation.

[0033] 5. Enhanced system optimization and scheduling support: Since the method proposed in this invention can quickly and accurately assess the system risk level, it provides strong support for the optimization and dispatch of the power system. By timely and accurately identifying potential risks, power system operators can better formulate response strategies, optimize resource allocation, and improve the overall reliability and stability of the system.

[0034] In summary, the present invention not only significantly improves the accuracy and efficiency of power system risk assessment through innovative data construction algorithms and improved CNN architecture, but also effectively reduces the computing cost and enhances the support capabilities of system optimization and scheduling. It has broad application prospects and significant practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart of a method for constructing neural network training data based on cross entropy according to a preferred embodiment; Figure 2 This is a structural diagram of a cross-entropy convolutional neural network based on a preferred embodiment. DETAILED DESCRIPTION

[0036] The principle of the present invention is: a training sample construction method based on cross entropy technology is proposed to improve the quality of training data, and the CNN model architecture with batch normalization layer is combined to analyze the power system state to reduce the gradient vanishing problem, enhance the generalization ability of the model, and improve the prediction accuracy. Specifically, the present invention increases the proportion of valid samples in the training data set by automatically identifying and prioritizing the system states that are critical to the risk indicators, thereby significantly improving the prediction accuracy and precision of the neural network model for risk assessment.

[0037] Embodiment 1:

[0038] like Figure 1 As shown, a method for constructing neural network training data based on cross entropy includes the following steps: S01: Use forced outage rate Sample the status of generators and transmission lines, and combine the sampled component status with the system load level to obtain the system status ; S02: According to the current system status , obtain the possible load reduction ; S03: Calculate the risk indicator function and likelihood ratio, and update the corrected forced outage rate , get the fault samples in the cross entropy iteration, is the number of iterations; S04: For the current system status and possible load reduction , use the running parameter matrix to construct the input of the neural network training data and the corresponding labels.

[0039] In a preferred embodiment, the system status is obtained in step S01. Also includes: Randomly sample system load levels based on the load model used and then update , use power flow and optimal power flow tools to analyze the current system status .

[0040] In a preferred embodiment, the risk index function is calculated in step S03 Likelihood Ratio The formula is:

[0041]

[0042] in, is the total number of generating units and transmission lines, It is In the system status sample component states, where Indicates success status. Indicates a fault condition. is a vector whose elements are the forced outage rates of generating units and transmission lines, It is elements, The forced outage rate is elements.

[0043] In a preferred embodiment, in step S03, the forced outage rate is updated and corrected. Each element of :

[0044] in, is the update factor, is the number of cross entropy samples in each iteration.

[0045] In a preferred embodiment, the operating parameter matrix in step S04 includes a bus admittance matrix:

[0046] The available power matrix on the busbar is:

[0047] Bus load power matrix:

[0048] Inter-bus transmission capacity matrix:

[0049] Where n represents the number of system buses, Indicates The busbar and Mutual admittance between buses , Indicates The self-admittance of the busbar, Indicates The total available generating capacity on the busbars, Indicates The total load demand on each bus, Indicates The busbar and The maximum available transmission capacity between the buses.

[0050] 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.

[0051] Another embodiment, a method for power system risk assessment based on a cross entropy convolutional neural network, comprises the following steps: Construct a convolutional neural network based on cross entropy; Constructing input data for a neural network using any of the methods described above; The relationship between the load reduction amount and the system state variables of the power system is obtained through the trained cross-entropy-based convolutional neural network; Analyze system status and calculate risk indicator functions.

[0052] like Figure 2 As shown, the convolutional neural network based on cross entropy contains Layers, each layer implements a transformation ,in Indicates the number of layers, and is a composite function of three consecutive operations: convolution, followed by batch normalization and rectified linear units. It is a flattening operation that transforms the two-dimensional feature map Reconstruct into one-dimensional data , and Both are composed of fully connected layers and rectified linear units. The output is the load reduction amount.

[0053] In a preferred embodiment, batch normalization normalizes the input by zero centering and scaling, using two parameters per layer to adapt, and normalizes the input using the mean and standard deviation of the current mini-batch, as follows:

[0054]

[0055]

[0056]

[0057] in, is the input mean vector, evaluated over the entire mini-batch, is the number of instances in the mini-batch, is the input standard deviation vector, evaluated over the entire mini-batch, is the zero-centered and normalized input vector of the instance, is a decimal number that avoids division by zero, is the output scaling parameter vector of the layer, represents vector element-wise multiplication, is the output shift parameter vector of the layer, is the output of batch normalization, which is a rescaled and shifted version of the input.

[0058] For the inference phase, the mean and variance are no longer recalculated for each mini-batch. Instead, the values ​​estimated by moving average during training are used. The mathematical formula for the adjustment is:

[0059] in, is the mean vector used for prediction, with an initial value of 0. is the variance vector used for prediction, with an initial value of 1. is the momentum scalar, after each batch of training, and Will be updated when the model training is completed. and The value of is determined and can be used directly in the prediction phase.

[0060] Another embodiment, a power system risk assessment system based on a cross entropy convolutional neural network, includes: Convolutional neural network building module, building a convolutional neural network based on cross entropy; An input data acquisition module, which constructs input data of a neural network using any of the methods described above; The prediction module obtains the relationship between the load reduction of the power system and the system state variables through the trained cross-entropy-based convolutional neural network; The risk assessment module analyzes the system status and calculates the risk indicator function.

[0061] Specifically, the following describes a power system risk assessment system based on a cross entropy convolutional neural network using a specific example, as follows: A. Neural Network Architecture Consider a single input passed through a convolutional network The network contains Layers, each layer implements a transformation ,in Indicates the number of layers. It can be a composite function of operations such as convolution (Conv), batch normalization (BN) or rectified linear unit (ReLU). The output of the layer is represented as .

[0062] The convolutional neural network structure proposed in this paper for power system state risk analysis is as follows: Figure 2 Its input data It includes the bus admittance matrix, the available power matrix on the bus, the load demand matrix on the bus, and the bus-to-bus transmission capacity matrix. For details, see Part C of this section. Output is the load reduction amount.

[0063] The proposed model and It is a composite function of three consecutive operations: a 4×4 convolution operation (Conv), followed by a batch normalization operation (BN) and a rectified linear unit (ReLU). and The convolution kernels of are all 64, and they extract relevant feature maps based on the system state variables. is a flattening operation, which converts the two-dimensional feature map Reconstruct into one-dimensional data . Transformation and The transformations are composed of fully connected (FC) layers and rectified linear units (ReLU). There are 128 outputs, There is 1 output, which is the load reduction amount.

[0064] B. Batch Normalization Layer Batch normalization (BN) is a key technology to solve the problem of gradient instability in neural network training and improve model accuracy. It normalizes the input by zero centering and scaling, using two parameters per layer to adapt. The essence of BN is to standardize the input using the mean and standard deviation of the current mini-batch, as described in formulas (6) to (9).

[0065] (6) (7) (8) (9) in, is the input mean vector, evaluated over the entire mini-batch; is the number of instances in the mini-batch; is the input standard deviation vector, also evaluated over the entire mini-batch. is the zero-centered and normalized input vector of the instance. is a decimal number that avoids division by zero; is the output scaling parameter vector of the layer; Represents vector element-wise multiplication; is the output shift parameter vector of the layer; is the output of the BN operation, which is a rescaled and shifted version of the input.

[0066] For the inference phase, the model usually predicts a single instance, and the BN operation is adjusted. It no longer recalculates the mean and variance for each mini-batch, but uses the values ​​estimated by moving average during training. This adjustment ensures consistent normalization based on the statistical properties of the training data, enhancing the stability and accuracy of single instance or mini-batch predictions.

[0067] The mathematical representation of this adjustment is contained in the following formula: (10) (11) in, is the mean vector used for prediction, with an initial value of 0. is the mean vector of each batch in training; is the variance vector used for prediction, with an initial value of 1. is the variance vector of each batch in training; is the momentum scalar (e.g., 0.99). After each batch of training, and will be updated. When the model training is completed, and The value of is determined and can be used directly in the prediction phase.

[0068] The addition of BN layers to traditional convolutional neural networks (CNNs) not only stabilizes the gradient flow and reduces the sensitivity to initial settings, but also accelerates training convergence and improves overall performance.

[0069] C. Data Acquisition Based on Cross Entropy In this paper, the following four operating parameter matrices (12)-(15) are used to construct the input data of the neural network. These matrices are square matrices of shape n times n, where n is the number of system buses. They can serve as different channels of input data for the convolutional neural network. The first matrix (12) is the bus admittance matrix, which represents the connectivity and power transmission capacity between each bus. The second matrix (13) is the available power matrix on the bus, which is a diagonal matrix that reflects the risk adjustment ability of the generator set. The third matrix (14) is the load power matrix on the bus, which is a diagonal matrix that represents the load size that the system needs to meet and reflects the upper limit of the load reduction. The fourth matrix (15) is the bus inter-transmission capacity matrix, which represents the upper limit of the transmission capacity between each bus. The definitions of these input data (system state variables) are as follows: (12) (13) (14) (15) Where n represents the number of system buses, Indicates The busbar and Mutual admittance between buses , Indicates The self-admittance of the busbar, Indicates The total available generating capacity on the busbars, Indicates The total load demand on each bus, Indicates The busbar and The maximum available transmission capacity between two busbars depends on whether there are lines between the two busbars and the available capacity of these lines.

[0070] In order to obtain input data and labels for training the neural network, Monte Carlo methods, DC power flow, and DC optimal power flow tools are used. A simple method is simple random sampling, which samples component states according to the forced outage rate of the components and combines them to obtain the system state, then constructs the above four matrices as input data under the current system state, and uses the load reduction under the current system state as the label of the training data set.

[0071] However, for high-reliability systems, most of the samples obtained by this method are in normal state, and they will not have load reduction, so they will not contribute to the value of the risk index, which in turn makes it impossible for the neural network model trained using these data to effectively mine the relationship between the system load reduction and the operating parameters (system state variables). To this end, this paper proposes a method for constructing neural network training data based on cross entropy. This method is based on the idea of ​​cross entropy (CE) to find the component state that contributes more to the risk index. Then, the proportion of valid samples in the training data is increased, so that the trained neural network model can effectively discover the relationship between the load reduction and the operating parameters (system state variables). The specific process of the algorithm is as follows: Step 1) In addition to the system's deterministic data and random parameters, the following parameters need to be specified: (a) Sample number , the number of cross entropy samples per iteration , and the cross entropy sampling flag ; (b) Failure percentage (e.g., 0.5); (c) Update factor (e.g., 0.5); (d) Maximum number of samples as a stopping criterion.

[0072] Step 2) Setup (The number of iterations of the cross entropy method to update the forced outage rate of the component); define ,Right now is equal to the initial component forced outage rate, is a vector whose elements are the forced outage rates of generators and transmission lines. The initial system load level is the peak load.

[0073] Step 3) Use forced outage rate Non-sequential simulation sampling of the generator and transmission line states, i.e., non-sequential Monte Carlo simulation (NSMCS), can also be used to combine the sampled component states with the system load level to obtain the system state , and determine Is True, if so, go to step 5, otherwise go to step 4.

[0074] Step 4) Randomly sample the system load level based on the load model used and then update .

[0075] Step 5) Use power flow and optimal power flow tools to analyze the current system status (whether it is a fault state) to obtain the possible load reduction amount Then confirm Is True? If so, go to step 6, otherwise go to step 9.

[0076] Step 6) Calculate the risk indicator function according to formulas (16)-(17) Likelihood Ratio . Determine the sample number Is it If yes, go to step 7, otherwise go to step 9.

[0077] (16) (17) in is the total number of generating units and transmission lines, It is In the system status sample component states, where Indicates success status. Indicates a fault condition.

[0078] Step 7) Use formula (18) to update the corrected forced outage rate Each element of .

[0079] (18) Step 8) Update the cross entropy iteration counter , and calculate whether the proportion of faulty samples in the current cross entropy iteration exceeds If exceeded, set , and the forced outage rate of the component is no longer updated.

[0080] Step 9) For the current system status and possible load reduction , use formulas (12)-(15) to construct the input of the neural network training data and the corresponding labels.

[0081] Step 10) Determine the sample number Whether it exceeds If yes, stop; otherwise, update the sample number And return to step 3.

[0082] Risk indicators also include: System Status Defined as a vector containing the The component status and system load level in the sample. The estimated value of the LOLP (loss of load probability) risk index and the estimated value of the EPNS (expected power shortage) risk index can be written as follows: (1) (2) in N is the total number of system samples, H LOLP and H EPNS is a risk indicator function used to evaluate the current system state through power flow and optimal power flow tools , defined as follows: (3) (4) in is the total system state space which can be divided into two subspaces: the successful state and failed state The failure state means that the system must reduce load to meet the system operating constraints. is corresponding to The total load reduction power.

[0083] The estimated values ​​of the reliability indicators LOLP and EPNS are random variables whose uncertainty can be determined by the variance coefficient, which is defined as follows in the case of EPNS: (5) Where V(·) represents the variance operation of the indicator. It should also be noted that the variance coefficient of the EPNS risk indicator is usually larger than that of the LOLP risk indicator, which means that EPNS is more difficult to converge, so the variance coefficient of EPNS is usually used as the convergence criterion in risk assessment.

[0084] The proposed method was tested using the IEEE RTS-79 system. The method was implemented using Python and all computations were performed on an AMD Ryzen 9 5950X CPU and an NVIDIA GeForce RTX 3060 GPU. During data acquisition, linear DC power flow and DC optimal power flow tools were used for emergency analysis and load reduction calculations. During neural network training, the neural network training cycle and batch size were 200 and 64, respectively, and the optimizer used Adam with a learning rate set to 10 -5 .

[0085] A. Data Collection Methods and Neural Network Architecture In the cross entropy sampling process, we used α=0.5, ρ=0.5, N CE =10,000 and N max = 100,000. For the simple random sampling process, considering the high reliability of the IEEE RTS-79 system, we increase the effective number of samples by setting the number of samples to 10 times that of the cross entropy sampling, resulting in N max =1,000,000.

[0086] Table 1 Comparison of models trained using different datasets

[0087] Table 1 shows the comparison of the training results of our proposed CNN models using different datasets. Model 1 is trained using a dataset generated by simple random sampling, while Model 2 is trained using a dataset generated by cross entropy sampling. It should be noted that both Model 1 and Model 2 use the proposed CNN architecture with BN, and the only difference is the dataset used for training. For Model 1, 90% of the simple random sampling samples, i.e., 900,000 samples, are used for the training dataset, and the remaining 100,000 samples are used as the validation set. If the mean squared error (MSE) of the validation set increases three times in a row, the training of Model 1 is stopped. For cross entropy sampling, 10% of the samples (10,000 samples) are designated as the test dataset for Model 1 and Model 2, and of the remaining 90,000 samples, 90% (81,000 samples) are used to train Model 2, and the remaining 9,000 samples are used for validation. If the MSE of the validation set increases three times in a row, the training of Model 2 is stopped. The evaluation results show that the model trained with the dataset generated by simple random sampling has higher error than the model trained with the cross entropy sampling data, whether measured by MSE or mean absolute error (MAE).

[0088] Table 2 Evaluation of different neural network architectures

[0089] Table 2 compares two different model architectures, both trained using cross entropy sampled data. The evaluation uses the same test data, which is not included in the training process of either model. The results show that the CNN without any BN layers has higher evaluation errors, both MSE and MAE, than the proposed network with BN layers.

[0090] B. IEEE RTS-79 Risk Assessment The IEEE RTS-79 system consists of 24 buses, 32 generators and 38 transmission lines. The system has an installed capacity of 3,405 MW and a peak load of 2,850 MW.

[0091] Table 3. Annual risk indicators of IEEE RTS-79

[0092] Table 3 shows the results of the annual risk assessment of IEEE RTS-79 using different methods. The system load level is randomly selected from 8736 load levels. The risk index obtained using the traditional NSMCS (Non-Sequential Monte Carlo Simulation) method is used as the reference value, where the variance coefficient β EPNS is set to 2%. For the other three methods, the convergence condition is β EPNS It can be seen that the traditional NSMCS method takes more than 800 seconds to obtain a LOLP of 1.2229 × 10 -3 , the relative error is 2.13%, and the EPNS is 1.4947 × 10 -1 , with a relative error of 1.16%. The CNN-based NSMCS method took 272 seconds to obtain a LOLP of 1.3421 × 10 -3 , the relative error is 12.08%, and the EPNS is 1.7021 × 10 -1 , with a relative error of 15.19%. The proposed method (NSMCS based on BN-CNN) takes 318 seconds using 620,000 samples and obtains a LOLP of 1.1839 × 10 -3 , the relative error is 1.13%, and the EPNS is 1.4148 × 10 -1 , the relative error is 4.25%. EPNS When the risk index is 5%, both the traditional NSMCS method and the proposed method can reasonably obtain the risk index, but the risk index obtained by the CNN-based NSMCS method has a large error. In addition, compared with the traditional NSMCS method, it can be seen that the calculation time is significantly reduced after using CNN with BN layer for state analysis. It should be noted that the state analysis based on neural network can predict hundreds or thousands of system states at a time, and the specific limit depends on the GPU memory. Its main time consumption lies in the construction of neural network input data.

[0093] Table 4. Annualized risk indicators of IEEE RTS-79

[0094] Table 4 shows the results of annualized risk assessment of IEEE RTS-79 at peak load level using the above four methods. The risk index obtained using the traditional NSMCS method is used as the reference value, where β EPNS is set to 2%. For the other three methods, the convergence condition is β EPNS It can be seen that the traditional NSMCS method takes 12 seconds to obtain a LOLP of 8.10×10 -2 , the relative error is 4.91%, and the EPNS is 1.4034×10 1 , with a relative error of 4.45%. The CNN-based NSMCS method takes 4 seconds to obtain a LOLP of 1.005× 10 -1 , the relative error is 17.98%, and the EPNS is 1.5494×10 1 , with a relative error of 5.49%. The proposed method (NSMCS based on BN-CNN) takes 4.9 seconds to obtain a LOLP of 8.52×10 -2 , the relative error is 0.02%, and the EPNS is 1.4066×10 1 , the relative error is 4.23%. EPNS When the risk index is 5%, both the traditional NSMCS method and the proposed method can obtain reasonable risk indicators. Although the CNN-based NSMCS method obtains a more reasonable EPNS index, its error in calculating the LOLP index is large. In addition, compared with the traditional NSMCS method, it can be seen that the calculation time is significantly reduced after using the neural network for state analysis.

[0095] 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 Sample the status of generators and transmission lines, and combine the sampled component status with the system load level to obtain the system status ; S02: According to the current system status , obtain the possible load reduction ; S03: Calculate the risk indicator function and likelihood ratio, and update the corrected forced outage rate , get the fault samples in the cross entropy iteration, is the number of iterations; S04: For the current system status and possible load reduction , 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 S01, the system status is obtained. Also includes: Randomly sample system load levels based on the load model used and then update , use power flow and optimal power flow tools to analyze the current system status .

3. The method for constructing neural network training data based on cross entropy according to claim 1, characterized in that: Step S03: Calculate the risk index function Likelihood Ratio The formula is: ， ， in, is the total number of generating units and transmission lines, It is In the system status sample component states, where Indicates success status. Indicates a fault condition. is a vector whose elements are the forced outage rates of generating units and transmission lines, It is elements, The forced outage rate is elements.

4. The method for constructing neural network training data based on cross entropy according to claim 1, characterized in that: In step S03, the forced outage rate is updated and corrected. Each element of : ， in, is the update factor, is the number of cross entropy samples in each iteration.

5. 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, Indicates The busbar and Mutual admittance between buses , Indicates The self-admittance of the busbar, Indicates The total available generating capacity on the busbars, Indicates The total load demand on each bus, Indicates The busbar and The maximum available transmission capacity between the buses.

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

7. The power system risk assessment method based on cross entropy convolutional neural network according to claim 6 is characterized in that: The cross entropy based convolutional neural network includes Layers, each layer implements a transformation ,in Indicates the number of layers, and is a composite function of three consecutive operations: convolution, followed by batch normalization and rectified linear units. It is a flattening operation that transforms the two-dimensional feature map Reconstruct into one-dimensional data , and Both are composed of fully connected layers and rectified linear units. The output is the load reduction amount.

8. The power system risk assessment method based on cross entropy convolutional neural network according to claim 6 is characterized in that: The batch normalization normalizes the input by zero centering and scaling, using two parameters per layer to adapt, and standardizes the input using the mean and standard deviation of the current mini-batch, as follows: ， ， ， ， in, is the input mean vector, evaluated over the entire mini-batch, is the number of instances in the mini-batch, is the input standard deviation vector, evaluated over the entire mini-batch, is the zero-centered and normalized input vector of the instance, is a decimal number that avoids division by zero, is the output scaling parameter vector of the layer, represents vector element-wise multiplication, is the output shift parameter vector of the layer, is the output of batch normalization, which is a rescaled and shifted version of the input; For the inference phase, the mean and variance are no longer recalculated for each mini-batch. Instead, the values ​​estimated by moving average during training are used. The mathematical formula for the adjustment is: ； in, is the mean vector used for prediction, with an initial value of 0. is the variance vector used for prediction, with an initial value of 1. is the momentum scalar, after each batch of training, and Will be updated when the model training is completed. and The value of is determined and used directly in the prediction phase.

9. A power system risk assessment system based on cross entropy convolutional neural network, characterized in that: include: Convolutional neural network building module, building a convolutional neural network based on cross entropy; An input data acquisition module, which constructs input data of a neural network using the method described in any one of claims 1 to 5; The prediction module obtains the relationship between the load reduction of the power system and the system state variables through the trained cross-entropy-based convolutional neural network; The risk assessment module analyzes the system status and calculates the risk indicator function.

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 5 is implemented.

Images