A method for ensuring the credibility of deep neural network results for large power grid reliability assessment

Abstract

The present invention belongs to the field of electric power systems, and specifically relates to a method for ensuring the credibility of deep neural network results for large power grid reliability assessment, including determining whether the parameters of the neural network model in the original scenario and the new scenario need to be updated according to data distribution; if the parameters need to be updated, the historical data of the new scenario is obtained to quickly update the deep neural network based on the transfer learning technology; for the updated neural network model, when a sample to be analyzed is input, the sample to be detected is compared with the historical data, and if the sample has similarities with the historical data, the prediction result of the neural network is used as the result of the flow analysis of the sample, otherwise the flow analysis of the sample is performed by a purely model-driven numerical method; the present invention realizes the rapid update of the deep neural network, saves training samples and training time, and can effectively judge the credibility of the classification of a single sample, and increase the robustness of the regression calculation of a single sample.

Description

Technical Field

[0001] The present invention belongs to the field of power systems, and in particular relates to a method for ensuring the credibility of deep neural network results for large power grid reliability assessment. Background Art

[0002] Reliability assessment of large power grids is an important tool to ensure the safe and reliable operation of power grids. With the proposal of the "dual carbon" goal, vigorously developing a high proportion of new energy is an important measure to support the "dual carbon" goal. However, a high proportion of new energy is bound to bring greater uncertainty to the power system, seriously threatening the safe and reliable operation of the power grid. Therefore, in order to timely and effectively assess the operation risks of the power grid, the assessment frequency will be increased from the previous hourly level to the minute level or even the second level. The essence of large power grid reliability assessment is to perform a large number of power flow and optimal power flow calculations, which has a heavy computational burden. The contradiction between high-frequency computing requirements and high computational burdens poses new challenges to the reliability assessment of large power grids. In this regard, scholars have proposed to use the high-precision approximation capability of deep neural networks to transfer the computational burden from online calculations to offline training, so as to achieve high-precision online calculations for large power grid reliability assessment. This method provides a new idea for large power grid reliability assessment: input the system operating conditions into the deep neural network flow module to directly obtain the flow results, and determine whether there is a flow over-limit situation. If so, it is a "problem" sample, otherwise it is a "successful" sample. Then, the system operating conditions of the problem sample are input into the deep neural network optimal flow module to determine whether load shedding will be performed. If so, it is a "fault" sample. For the fault sample, the interior point method is used to solve the minimum load shedding flow model. However, the high-precision calculation of the above-mentioned deep neural network usually assumes that the working sample and the training sample follow the same distribution. For complex and changeable power systems, this assumption is difficult to hold, and the credibility of the deep neural settlement results is difficult to guarantee. Therefore, how to judge the applicability of deep neural network reliability assessment models is one of the main bottlenecks in its promotion and application in power grids. Summary of the invention

[0003] In view of the problem that the credibility of large power grid reliability assessment results based on deep neural networks is difficult to ensure, the present invention proposes a method for ensuring the credibility of deep neural network results for large power grid reliability assessment, comprising the following steps:

[0004] Determine whether the neural network model needs to update parameters in the original scenario and the new scenario based on the data distribution;

[0005] If the parameters need to be updated, historical data of the new scene is obtained to quickly update the deep neural network based on transfer learning technology;

[0006] For the updated neural network model, when the sample to be analyzed is input, the sample to be tested is compared with the historical data. If there is similarity between the sample and the historical data, the prediction result of the neural network is used as the power flow analysis result of the sample. Otherwise, the power flow analysis of the sample is performed using a purely model-driven numerical method.

[0007] Furthermore, whether the deep neural network needs to be updated is determined based on the difference in data distribution between the original scene and the new scene. If the Jensen-Shannon value of the two scenes is 0, no update is required, otherwise it needs to be updated.

[0008] Furthermore, the Jensen-Shannon values ​​of the two scenarios are expressed as:

[0009]

[0010] Among them, D JS (D o ||D e ) represents the original scene data distribution D o With the new scene data distribution D e Jensen-Shannon value between; D KL (D o ||D e ) is the original scene data distribution D o With the new scene data distribution D e The divergence between Ω is the variable set of all input and output of the reliability assessment model.

[0011] Furthermore, the fast update of the deep neural network includes the following steps:

[0012] Preprocess the raw data;

[0013] Use the trained deep neural network parameters to initialize the new deep neural network in the evolution scenario;

[0014] Based on the input raw data, all parameters of the new deep neural network are fine-tuned using the Rmsprop algorithm;

[0015] Whether to end fine-tuning training is determined when the number of epoch iterations reaches the set value or the early stopping condition is met during training. If the training is ended, the update of the neural network is completed.

[0016] Furthermore, the process of preprocessing the raw data includes:

[0017]

[0018] Among them, v meanand v std are the mean and standard deviation of the original data vector V respectively.

[0019] Furthermore, the process of determining whether a sample adopts the prediction result of the neural network as its power flow analysis includes:

[0020] The historical data is divided into k groups based on the K-means clustering method, and the Euclidean distance between the sample to be analyzed and the center of each cluster is calculated;

[0021] The class with the smallest Euclidean distance to the sample to be analyzed is selected, and two samples with the smallest Euclidean distance to the sample to be analyzed are found as similar samples;

[0022] If the training error of the two most similar training samples exceeds the set threshold, the deep neural network is not used to analyze the test sample and a pure model-driven approach is adopted; otherwise, the deep neural network is used to analyze the sample.

[0023] Furthermore, in the process of deep neural network processing optimal power flow calculation, the classification prediction result of the sample output by the neural network model is a probability. The value of the probability is used to judge whether the result output by the neural network model is credible, and the power flow analysis is performed on the unreliable samples through the pure model-driven numerical method. The process of judging whether the sample is credible is as follows:

[0024]

[0025]

[0026] in, is the output of the deep neural network optimal power flow module.

[0027] Furthermore, in the process of processing power flow calculation by deep neural network, a relaxation factor ξ is introduced into the neural network to ensure that samples with a proportion of 1-λ will not be misjudged as problematic samples due to a small error e. The relationship between the relaxation factor ξ and the errors e and λ is expressed as:

[0028]

[0029] Where, M is the system power flow parameter; M and They correspond to the upper and lower limits of the system flow parameters respectively; the value of the smaller error e mentioned in the present invention is determined according to the experience of those skilled in the art. The field determines that the value of the error e in a certain range is smaller according to actual conditions, and the present invention does not limit the specific range of this value.

[0030] The present invention can be widely used in online calculation of large power grid reliability based on artificial intelligence, and is particularly suitable for power flow regression and optimization classification problems based on deep neural networks. Compared with the prior art, the present invention has the following advantages:

[0031] 1. The adaptability criterion of the deep neural network for reliability assessment proposed in the present invention can determine whether to continue to use the deep neural network for reliability assessment for new scenarios and single samples respectively. For new scenarios, if the adaptability criterion is not met, the present invention proposes a deep neural network reliability model based on transfer learning to achieve rapid updating of the deep neural network, saving training samples and training time. For a single sample, if the adaptability criterion is not met, a model-driven numerical solution method (Newton method, interior point method, etc.) can be used to solve the reliability assessment model.

[0032] 2. The deep neural network output correction strategy for large power grid reliability assessment proposed in the present invention starts from the two aspects of regression model and classification model, based on theoretical derivation and rule design, and predicts and corrects samples with unreasonable calculation results. The proposed method can effectively judge the credibility of single sample classification and increase the robustness of single sample regression calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a method for ensuring the credibility of deep neural network results for large power grid reliability assessment according to the present invention;

[0034] Figure 2 A schematic diagram for searching two historical samples that are most similar to the sample to be detected for the present invention;

[0035] Figure 3 Schematic diagram of the maximum value of Jensen-Shannon divergence in input variables between different examples of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] The present invention proposes a method for ensuring the credibility of deep neural network results for large power grid reliability assessment. Figure 1 , including the following steps:

[0038] Determine whether the neural network model needs to update parameters in the original scenario and the new scenario based on the data distribution;

[0039] If the parameters need to be updated, historical data of the new scene is obtained to quickly update the deep neural network based on transfer learning technology;

[0040] For the updated neural network model, when the sample to be analyzed is input, the sample to be tested is compared with the historical data. If there is similarity between the sample and the historical data, the prediction result of the neural network is used as the power flow analysis result of the sample. Otherwise, the power flow analysis of the sample is performed using a purely model-driven numerical method.

[0041] Example 1

[0042] The purpose of the method of the present invention is to address the problem that the credibility of large power grid reliability assessment results based on deep neural networks is difficult to ensure, and to propose a deep neural network security strategy for large power grid reliability assessment. The technical solution to implement the present invention includes:

[0043] First, this embodiment proposes a deep neural network adaptability criterion for reliability assessment and a fast update strategy thereof, and determines scenes or samples with unsatisfactory error levels (referred to as untrustworthy scenes or samples) based on a quantification method of sample similarity;

[0044] For untrustworthy scenarios, the present invention further proposes a fast update strategy for the deep neural network reliability model based on transfer learning to achieve fast update of the deep neural network;

[0045] For untrusted samples, alternative model-driven tools can be used.

[0046] Secondly, this embodiment proposes a deep neural network output result correction strategy for operational reliability assessment, theoretically derives and determines a relaxation factor to reduce the misjudgment of line overload / voltage over-limit caused by small regression calculation errors of the deep neural network, and proposes a credibility quantification method for the output results of the "fault" sample classifier.

[0047] Reliability assessment considers the impact of various types of uncertainty on the continuous power supply capability of the power system. The input of reliability assessment is usually the distribution of various types of uncertainty. For different scenarios, the distribution of input variables is different, and the adaptability of deep neural networks is different. Therefore, this embodiment first determines the adaptability of the neural network in the new scenario and updates the neural network; then determines the adaptability of a single sample sampled from the input distribution to determine whether the prediction result of the neural network can be used as the detection result. Specifically includes

[0048] 1. Rapid update of neural networks in new scenarios

[0049] The good effect of deep neural networks is usually based on the following assumption: the test samples and training samples follow the same distribution. However, in actual operation, new scenarios / evolved scenarios (hereinafter referred to as new scenarios) often appear, such as increased load demand, increased penetration of new energy, and expansion of transmission lines. The data distribution in the new scenario must be inconsistent with the data distribution in the original scenario. Depending on whether the dimension of the feature vector has changed, the new scenario can be divided into two categories, namely, the data distribution is the same as that of the original scenario and the data distribution is different from that of the original scenario. If the data distribution is different from the original scene, that is, the dimension of the feature vector is different, the deep neural network must be updated; if the input and output dimensions of the new scene are consistent with the original scene, the generalization ability of the deep neural network may make the deep neural network trained for the original scene still usable in the new scene.

[0050] In this embodiment, whether the deep neural network is updated can be determined based on the difference in data distribution between the original scene and the new scene. In this embodiment, the difference in data distribution between the two scenes is evaluated based on the Jensen-Shannon divergence. If the value D of Jensen-Shannon is JS =0, it means that the two distributions are consistent; if D JS =1, it means that the two distributions are completely different. The Jensen-Shannon values ​​of the two scenarios are expressed as:

[0051]

[0052] Among them, D JS (D o ||D e ) represents the original scene data distribution D o With the new scene data distribution D e The Jensen-Shannon value between KL (D o ||D e ) is the original scene data distribution D o With the new scene data distribution D e The divergence between D o (r), D e (r) refers to the original scene data distribution D corresponding to data r in the variable set of all input and output of the reliability assessment model o , New scene data distribution D e ; Ω is the variable set of all input and output of the reliability assessment model.

[0053] The maximum value of the Jensen-Shannon divergence between the input variables of the original scene and the new scene is used as the evaluation indicator to determine whether the deep neural network needs to be updated. For any input variable, if DJS The value is less than K limit , then the original deep neural network can be used; otherwise, the deep neural network needs to be updated. The process is expressed as:

[0054]

[0055] The value of Klimit is predetermined. The present invention sets the value of Klimit to 0.1 based on experience.

[0056] If the deep neural network needs to be updated, this embodiment provides a method for quickly updating the deep neural network reliability assessment model based on transfer learning. Transfer learning attempts to transfer the "knowledge" of some previous tasks to the target task. The definition of transfer learning is as follows: given a source domain D s and learning task T s , target domain D T and the corresponding learning task T T , aims to utilize D s and T s Knowledge in (where D s ≠D T , T s ≠T T ) helps to predict the target function f T (·) is learned by deep neural network on the source domain D s The learned knowledge is reflected in the update of its own parameters. Parameter transfer learning is the most common technique in transfer learning. It improves learning efficiency by providing better deep neural network parameter initialization (rather than random initialization). s and T s The parameters of the deep neural network are directly transferred to the D T and T T Deep neural networks, which have been successfully applied to many tasks.

[0057] The present invention takes into account that the well-trained deep neural network has extracted complex features useful for reliability assessment, so as to reduce the workload of re-collecting training data and rebuilding the deep neural network. In actual use, when the topological structure changes, such as adding nodes and branches, the number of neurons in the input and output layers of the deep neural network is no longer consistent. In this regard, the idea of ​​using parameter transfer learning for initialization in the present invention is: the parameters of the corresponding node positions of the input and output layers are directly migrated, and the parameters of the input or output neurons corresponding to the newly added nodes are randomly initialized. Generally, in order to avoid the overfitting problem, the existing transfer learning method fixes the parameters of the input layer and hidden layer of the original model, and only adjusts the corresponding weights and bias parameters of the output layer. However, the output of the neural network is the result of the interaction of all parameters in the model. Fixing certain parameters in the model may destroy the interaction between parameters. Therefore, this chapter chooses to adjust all parameters of the new model, that is, after the parameters are migrated, the learning algorithm Rmsprop or Adam is used to fine-tune all parameters of the new deep neural network. This embodiment uses the early stopping method to effectively avoid the overfitting problem. The fast update strategy of deep neural networks based on transfer learning is summarized as follows:

[0058] Step 1: Preprocess the raw data, including:

[0059]

[0060] Among them, v mean and v std are the mean and standard deviation of vector V respectively;

[0061] Step 2: Use the trained deep neural network parameters to initialize the new deep neural network in the evolution scenario;

[0062] Step 3: Fine-tune the parameters of the deep neural network under the evolution scenario, that is, use the Rmsprop algorithm to fine-tune all parameters of the new deep neural network according to the input feature vector X and the output feature vector Yn;

[0063] Step 4: The condition for stopping training in step 3 is determined based on whether the number of epoch iterations reaches the set value or the early stopping condition is met during training.

[0064] This method is simple and effective, and can effectively utilize the parameter information of well-trained deep neural networks.

[0065] 2. Adaptability of a single sample

[0066] Even if a deep neural network can be applied to a certain scenario, due to reasons such as hyperparameter settings and difficulty in ensuring global optimality of the training algorithm, the deep neural network cannot fully mine all the knowledge contained in the training data, resulting in a small number of samples that will inevitably have large calculation errors. Therefore, detecting such samples and avoiding using deep neural networks to analyze them is crucial to further improve the accuracy of data-driven reliability assessment. In essence, deep neural networks mine potential connections between data by learning the sensitivity of outputs to inputs during training. Therefore, very similar inputs usually correspond to similar outputs of deep neural networks.

[0067] The present invention proposes a sample-oriented deep neural network adaptability criterion to determine whether to use a deep neural network to analyze a certain sample. Specifically, for a test sample to be calculated, if the training errors of the two most similar training samples are not satisfactory, the deep neural network is not used to analyze the test sample and a pure model-driven method is adopted; otherwise, the deep neural network is used to analyze the sample. The method of finding the two closest training samples is as follows: Figure 2 , specifically including the following steps:

[0068] In order to avoid comparison with all training samples and improve search efficiency, the present invention divides the training samples into K groups based on the K-means clustering method. Figure 2 The historical data is divided into four groups. The five-pointed stars in each group represent samples with high accuracy in each group, and × represents samples with low accuracy in each group. The square symbols in the figure represent samples to be tested.

[0069] Calculate the Euclidean distance between the test sample to be analyzed and each cluster center, and determine the group with the smallest Euclidean distance, for example, Figure 2 In the case shown, the distance between the sample to be tested and the center of the third group of data is closest.

[0070] Compare in the nearest group until the two most similar training samples are found. This chapter uses the Euclidean distance to quantify the distance between the training samples in the nearest group and the samples to be analyzed, expressed as:

[0071] D Euclidean (X train ,X cal )=||X train,i -X cal ||

[0072] Among them, X train With X calis the input feature vector after data preprocessing, which belongs to the training sample and the working sample to be analyzed. The training error of the two selected historical data samples is compared with the set error threshold. If the training error is smaller than the set error threshold, it is considered that the deep neural network can be used to calculate the relevant power flow of the sample to be tested.

[0073] The output of the reliability model based on the deep neural network includes regression results and classification results. For the classification results, when the deep neural network predicts and classifies a certain sample, that is, the optimal power flow calculation, the output is a numerical value, which must have a certain deviation from the true label (0 or 1). Therefore, in order to determine which category of samples the sample belongs to in the reliability assessment, it is usually necessary to give a threshold to determine, such as exceeding the threshold for one category, and not exceeding the threshold for another category. Deviation (the deviation between the output of the deep neural network and the true label) is a simple method to quantify the degree to which the deep neural network approximates the true label. In this regard, the present invention uses deviation to estimate the credibility of the output of the deep neural network, expressed as:

[0074]

[0075]

[0076] The above formula divides the output of the deep neural network into two levels, namely credible and unreliable. For credible results, the judgment results of the deep neural network are directly used. For samples with unreliable results, the power flow analysis (such as power flow calculation and minimum load shedding calculation) should be performed through pure model-driven numerical methods.

[0077] When the output of the reliability model based on a deep neural network is a regression result, that is, a power flow regression calculation, for credible samples of the input model, in the reliability assessment based on a deep neural network, a deep neural network will be used to approximate the power system power flow calculation, and its regression results will be used to determine whether there are system problems such as voltage over-limit and line overload, and then identify the problem samples. The present invention introduces a very small relaxation factor ξ at the system operating limit to make the deep neural network regression result more robust, thereby avoiding the misjudgment of problem samples due to a small regression error e. The following theory establishes the relationship between the relaxation factor ξ value, the regression error e and λ. This relationship can help determine the value of ξ to ensure that samples with a proportion of 1-λ will not be misjudged as problem samples due to a small error e, expressed as:

[0078]

[0079] Wherein, M is the system power flow parameter, including branch power, voltage amplitude, etc. These parameters are the samples of the present invention, and M is related to They correspond to the upper and lower limits of the system power flow respectively.

[0080] After the above parameters are adjusted, when E(e 2 ) is small, the deep neural network regression error is less than The probability is greater than 1-λ, that is, if

[0081]

[0082] Therefore, if you need to:

[0083]

[0084] Then there may be

[0085] P{|e|<ξ}=1-P{|e|≥ξ}≥1-λ

[0086] From the above formula, we can see that P{|e|≥ξ} cannot be greater than 1, so the ξ value will be greater than E(e 2 ) value, when E(e 2 ) is large, ξ must be large, which will make it meaningless to avoid misjudgment caused by small calculation errors in deep neural networks. Therefore, only when E(e 2 ) is relatively small, the formula derived by the present invention has practical application value.

[0087] Example 2

[0088] In this embodiment, the IEEE RTS-79 system, the IEEE RTS-96 system and their modified examples are used to verify the effectiveness of the method of the present invention. The example information is as follows:

[0089] Example 1: IEEE RTS-79 system. Assume that the load curve follows a normal distribution with a standard deviation of 10% of the expected value and the historical annual peak load of the IEEE RTS-79 system is the expected value.

[0090] Example 2: IEEE RTS-96 system In this system, three IEEE RTS 79 systems are connected by six transmission lines.

[0091] Example 3: Modified IEEE RTS-79 system with a 20% penetration rate of renewable energy.

[0092] Example 4: Modified IEEE RTS-96 system with a 20% penetration rate of renewable energy.

[0093] Example 5: Modified IEEE RTS-79 system with a renewable energy penetration of 25%.

[0094] Example 6: Modified IEEE RTS-79 system. In Example 1, a new branch is added from bus node 1 to bus node 5.

[0095] Example 7: Modified IEEE RTS-79 system. In Example 1, a bus and two new lines are added.

[0096] The convergence condition of the Monte Carlo method is when the variance coefficient α is less than 5% or the number of samples reaches 100,000. The reliability result of the Monte Carlo method combined with the Newton-Raphson method and the interior point method is an exact solution. The hyperparameters of the deep neural network are detailed in 0. The following indicators are used to quantify the accuracy of the power flow calculation. P vm It refers to the proportion of voltage amplitude absolute error exceeding 0.001pu. va It refers to the proportion of voltage phase angle absolute error exceeding 0.01rad. pf / P qt Refers to the proportion of the absolute error of the active power flowing into / out of the branch exceeding 5MW. M_VM, M_VA, M_PF and M_PT correspond to the above P vm , P va , P pf and P qt The root mean square error of these variables.

[0097] The accuracy of a classifier is usually measured by the following metrics:

[0098]

[0099]

[0100]

[0101]

[0102] Among them, TP = True Pos, that is, true positive class; FN = False Negative, that is, false negative class; TN = True Negative, that is, true negative class; FP = False Positive, that is, false positive class; sensitivity is recall; specificity is specificity.

[0103] The training samples can be collected through the actual operation data of the project or obtained through simulation. The present invention obtains the training samples by solving the power flow equation and the minimum load shedding model based on the Newton method and the interior point method respectively. The operating conditions are used as the input feature vectors of the deep neural network power flow module and the deep neural network optimal power flow classification module. The output feature vector of the deep neural network power flow regression module is the power system power flow distribution, including branch power and voltage amplitude, etc. The output feature of the deep neural network optimal power flow classification module is a flag vector 0 or 1, representing whether the optimal power flow will shear the load. For Examples 1 to 4, the hyperparameters of the deep neural network intelligent module are shown in Table 1.

[0104] Table 1 Hyperparameters of different deep neural network intelligent modules under different examples

[0105]

[0106] Figure 3 is the maximum value of the Jensen-Shannon divergence between the input variables of different examples. Table 2 lists the performance analysis of directly using the deep neural network of the original scene to calculate the new scene. Figure 3 As shown in the figure, the maximum Jensen-Shannon divergence values ​​between Example 1 and Example 3 and between Example 1 and Example 5 are 0.4213 and 0.4438 respectively. These values ​​are much greater than 0.1, indicating that the data distribution differences between the two examples are large. If the deep neural network of Example 1 is used directly, the accuracy of Examples 3 and 5 will be greatly reduced by 3. In contrast, the maximum Jensen-Shannon divergence value between Example 3 and Example 5 is 0.0321, which is less than 0.1, indicating that the data distribution difference between the two is small. Therefore, the reliability index directly calculated by the deep neural network of Example 3 for Example 5 can still meet the accuracy requirements, with g-mean greater than 97%, and the relative errors of LOLP and EENS no more than 2.62% and 1.95% respectively.

[0107] Table 2 Performance of directly using the deep neural network of the original scene to calculate the new scene

[0108]

[0109] For the two new scenarios of Example 6 and Example 7, since they cause the dimension of the deep neural network feature vector to change, they must be updated. Table 3 is the performance analysis of Example 6 and Example 7 after adopting the deep neural network update strategy. According to the deep neural network fast update strategy based on transfer learning proposed by the present invention, only 1000 training samples are used, and the total update time of Example 6 and Example 7 does not exceed 10 seconds respectively. The relative error of the reliability index calculated after the update is less than 5%, which has a high accuracy. Therefore, for deep neural networks that are no longer applicable, the deep neural network can be quickly updated with the help of the reliability assessment "knowledge" that the deep neural network has learned, without learning from scratch, which can effectively save the training sample volume and training time of the deep neural network.

[0110] Table 3 Performance analysis of Example 6 and Example 7 after using the deep neural network update strategy

[0111]

[0112] Table 4 shows the effectiveness verification results of the deep neural network adaptability criterion for a single sample in Examples 1-4. The two training samples closest to the test sample are measured by Euclidean distance. If the error level of the training sample is high, the working sample is considered to be an unreliable sample, and the deep network is not used for analysis to improve the calculation accuracy of the working sample. It can be seen from Table 4 that the total number of unreliable samples in all examples is very small compared to the total number of tens of thousands of working samples, and does not exceed 243. In addition, among all unreliable samples, about 30.7%-78.8% of the samples will get wrong classification conclusions when analyzed by deep neural networks. It can be seen that the proposed adaptability criterion for a single sample can effectively detect samples with large calculation errors of deep neural networks at the cost of only a small computational burden, significantly improving the calculation accuracy.

[0113] Table 4. Validation of the adaptability criteria of deep neural networks for a single sample under Examples 1 to 4.

[0114]

[0115] Table 5 shows the computational performance of the deep neural network using the relaxation factor in Examples 1-4. The present invention designs λ to be equal to 10%, and the relaxation factor ξ value is calculated according to the formula. The results of the deep neural network regression model classification problem samples calculated are specifically shown in Table 5. The classification results of the deep neural network without the relaxation factor are compared, that is, when λ is equal to 0. It can be found from Table 5 that in Examples 1, 2, and 4, the deep neural network regression results judged that the problem samples were not misjudged due to small calculation errors, so the accuracy of the two indicators ACC and SEN did not change before and after the relaxation factor was used. In Example 2, the relaxation factor effectively avoided the misclassification caused by the small calculation error of the deep neural network, and increased the ACC from 99.13% to 99.45%, and the SEN from 99.82% to 100%.

[0116] Table 5 Computational performance of deep neural networks using relaxation factors in Examples 1-4

[0117]

[0118] Table 6 shows the computational performance of the proposed safety classification method using the deep neural network in Examples 1-4. The reliability of the deep neural network classification results is quantified by the deviation between the deep neural network prediction value and the predefined classification label. Unreliable samples are called unreliable samples. As can be seen from Table 6, the unreliable samples identified by the proposed method in Examples 1-4 range from dozens to hundreds or even thousands. In the power flow regression module, the use of deep neural networks for analysis will result in a proportion of erroneous results of about 40%, while in the optimal power flow module, the proportion reaches more than 80% or even 97.5%. It can be seen that the proposed method can effectively predict samples with inaccurate calculations of the deep neural network at a relatively low computational cost, thereby improving the computational accuracy of subsequent analysis. In particular, compared with the power flow classification module, the proposed method has more prominent advantages in the optimal power flow intelligent classification module.

[0119] Table 6 Computational performance of deep neural network using the proposed security classification method in Examples 1-4

[0120]

[0121] In summary, the present invention proposes a deep neural network security strategy for large power grid reliability assessment. Compared with the existing data-driven reliability assessment method, the present invention considers the weak interpretability of deep neural networks for the first time, and proposes a response strategy to effectively ensure the credibility of data-driven reliability assessment results. First, the adaptability criterion of the deep neural network for reliability assessment proposed in the present invention can determine whether the deep neural network can be used for reliability assessment for new scenarios and single samples respectively. For new scenarios that do not meet the adaptability criterion, the present invention proposes a deep neural network reliability model based on transfer learning to achieve rapid updating of deep neural networks, saving training samples and training time. In addition, the present invention also proposes a deep neural network output correction strategy for large power grid reliability assessment from the two aspects of regression model and classification model. Based on theoretical derivation and rule design, samples with unreasonable calculation results are prejudged and corrected. The proposed method can effectively judge the credibility of single sample classification and further increase the credibility of the calculation results.

[0122] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for ensuring the credibility of deep neural network results for large power grid reliability assessment, characterized in that: The following steps are involved: According to the data distribution, it is determined whether the neural network model needs to update parameters in the original scene and the new scene. That is, it is determined whether the deep neural network needs to be updated according to the difference in data distribution between the original scene and the new scene. If the value of Jensen-Shannon of the two scenes is 0, no update is required, otherwise it needs to be updated. The value of Jensen-Shannon of the two scenes is expressed as: Among them, D JS (D o ||D e ) represents the original scene data distribution D o With the new scene data distribution D e Jensen-Shannon value between; D KL (D o ||D e ) is the original scene data distribution D o With the new scene data distribution D e The divergence between Ω is the variable set of all input and output of the reliability assessment model; If the parameters need to be updated, historical data of the new scene is obtained to quickly update the deep neural network based on transfer learning technology; For the updated neural network model, when the sample to be analyzed is input, the sample to be tested is compared with the historical data. If there is similarity between the sample and the historical data, the prediction result of the neural network is used as the power flow analysis result of the sample. Otherwise, the power flow analysis of the sample is performed using a purely model-driven numerical method.

2. According to claim 1, a method for ensuring the credibility of deep neural network results for large power grid reliability assessment is characterized in that: Performing fast updates on a deep neural network involves the following steps: Preprocess the raw data; Use the trained deep neural network parameters to initialize the new deep neural network in the evolution scenario; Based on the input raw data, all parameters of the new deep neural network are fine-tuned using the Rmsprop algorithm; Whether to end fine-tuning training is determined when the number of epoch iterations reaches the set value or the early stopping condition is met during training. If the training is ended, the update of the neural network is completed.

3. According to claim 2, a method for ensuring the credibility of deep neural network results for large power grid reliability assessment is characterized in that: The process of preprocessing raw data includes: Among them, v mean and v std are the mean and standard deviation of the original data vector V respectively.

4. According to claim 1, a method for ensuring the credibility of deep neural network results for large power grid reliability assessment is characterized in that: The process of determining whether a sample uses the prediction result of the neural network as its power flow analysis includes: The historical data is divided into k groups based on the K-means clustering method, and the Euclidean distance between the sample to be analyzed and the center of each cluster is calculated; Get the class with the smallest Euclidean distance to the sample to be analyzed, and find two samples with the smallest Euclidean distance to the sample to be analyzed as similar samples; If the training error of the two most similar training samples exceeds a set threshold, the deep neural network is not used to analyze the sample to be analyzed, but a pure model-driven method is used to analyze the sample to be analyzed; otherwise, a deep neural network is used to analyze the sample.

5. The method for ensuring the credibility of deep neural network results for large power grid reliability assessment according to any one of claims 1 to 4, characterized in that: In the process of deep neural network processing optimal power flow calculation, the classification prediction result of the sample output by the neural network model is a probability. The value of the probability is used to judge whether the result output by the neural network model is credible, and the power flow analysis is performed on unreliable samples through a purely model-driven numerical method. The process of judging whether the sample is credible is as follows: Among them, y i is the output of the deep neural network optimal power flow module.

6. The method for ensuring the credibility of deep neural network results for large power grid reliability assessment according to any one of claims 1 to 4, characterized in that: In the process of processing power flow calculation by deep neural network, the relaxation factor ξ is introduced into the neural network to ensure that the samples with a proportion of 1-λ will not be misjudged as problematic samples due to a small error e. The relationship between the relaxation factor ξ and the errors e and λ is expressed as: Among them, M is the system power flow parameter; M and They correspond to the upper and lower limits of the system power flow parameters respectively.

Images