Power system prediction auxiliary state estimation method and device based on deep learning

Abstract

The invention discloses an electric power system prediction auxiliary state estimation method and device based on deep learning. The method comprises the following steps: firstly, collecting power system data and inputting the power system data into a pre-training neural network integrated with a CNN, an attention mechanism and an LSTM to predict a system state; and an innovation correction amount is determined by using a difference value between a predicted value and an actually measured value, so that a measurement and process noise covariance matrix is adjusted. And finally, in combination with the adjusted covariance matrix and a Kalman filtering algorithm, real-time and accurate estimation of the state of the power system is realized, so that the problem that the state of the power system cannot be accurately estimated in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of power system prediction auxiliary state estimation based on deep learning, and in particular to a power system prediction auxiliary state estimation method and device based on deep learning. Background Art

[0002] State estimation is the core of power system monitoring and is crucial to EMS. With the widespread use of new energy and power electronic equipment, the randomness and volatility of the system have increased, and traditional state estimation faces challenges in accuracy and real-time performance. The development of PMU and WAMS provides new data sources that help capture the spatiotemporal characteristics of the system and improve state estimation.

[0003] Dynamic state estimation relies on Kalman filtering to predict system states through measurement data, but prediction accuracy and computational efficiency remain issues. The FASE method improves prediction capabilities by simplifying the model, but has limited effectiveness when dealing with complex data. The innovation process and exponential smoothing methods improve prediction performance, while deep learning-based methods show higher prediction accuracy.

[0004] The identification of bad data and network topology is a challenge, and the innovation difference performs better as a neural network input. Although the FASE algorithm with robust performance has been proposed, the estimation and filtering effects are still reduced in the face of non-normal distribution and complex working conditions, resulting in the inability of existing technologies to accurately evaluate the state of the power system. Summary of the invention

[0005] The present invention provides a method and device for power system prediction auxiliary state estimation based on deep learning to solve the problem that the state of the power system cannot be accurately evaluated in the prior art.

[0006] In a first aspect, the present application provides a power system prediction auxiliary state estimation method based on deep learning, comprising:

[0007] Acquire electrical data of power systems;

[0008] Inputting the electrical data into a trained neural network so that the neural network outputs a predicted state value of the electrical system; wherein the neural network is constructed by combining a convolutional neural network with an attention mechanism and a long short-term memory network;

[0009] Calculating the new information correction amount according to the difference between the predicted state value and the actual measured state value;

[0010] According to the new information correction amount, the measurement noise covariance matrix and the process noise covariance matrix are corrected to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix;

[0011] The state estimation value of the first power system is calculated based on the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm.

[0012] This application uses a deep learning model that integrates convolutional neural networks, attention mechanisms, and long short-term memory networks to process electrical data of the power system, thereby achieving high-precision prediction of the system state. By comparing the predicted state value and the actual measured value, calculating the new information correction amount, and adjusting the measurement and process noise covariance matrix accordingly, this method not only improves the accuracy of state prediction, but also enhances the system's adaptability to noise and mutations in actual operation. Finally, using the corrected covariance matrix and Kalman filtering algorithm, real-time and accurate estimation of the power system state is achieved to solve the problem that the prior art cannot accurately evaluate the state of the power system.

[0013] As a preferred embodiment of the first aspect, after obtaining the corrected measurement noise covariance matrix and the corrected process noise covariance matrix, the method further includes:

[0014] The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least square method.

[0015] The state estimation value of the power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least squares method, specifically:

[0016] The first calculation formula of the least squares method is:

[0017]

[0018] In the formula, z is the actual measured state value, x is the estimated state value of the power system, is the system state prediction value, H is the measurement matrix, which reflects the relationship between the system measurement value and the state value, R is the measurement noise covariance matrix, and M is the covariance matrix of the prediction value;

[0019] When the first calculation formula takes the minimum value, a second calculation formula is obtained, and the second formula is:

[0020]

[0021] According to the second calculation formula, an estimated value of the state of the power system is calculated.

[0022] In this preferred embodiment, the present application combines the predicted state value and the actual measured state value of the deep learning model, uses the modified measurement noise covariance matrix to improve the reliability of the measurement data, and the modified process noise covariance matrix to enhance the adaptability of the system dynamic characteristics, and further uses the least squares algorithm to optimize the real-time tracking capability of the power system state. This method calculates the initial state estimate value by the least squares method, ensuring the mathematical rigor of the estimation process, thereby achieving higher accuracy and robustness in the power system prediction auxiliary state estimation based on deep learning. In this way, the present invention not only improves the accuracy of state estimation, but also enhances the system's resistance to noise and system mutations, and ultimately achieves accurate grasp of the power system state, providing strong technical support for the stable operation and effective management of the power grid.

[0023] As a preferred embodiment of the first aspect, the innovation correction amount is calculated according to the difference between the predicted state value and the actual measured state value, specifically:

[0024] The innovation correction amount is calculated according to the covariance matrix of the innovation correction amount.

[0025] In this preferred embodiment, the present application calculates the difference between the predicted state value and the actual measured state value as the initial new information. This step can directly reflect the deviation between the prediction and the actual, and provides a clear starting point for subsequent corrections. Subsequently, the initial new information is standardized to obtain the new information correction amount. This process makes the correction amount more accurate and practical by considering the size of the deviation relative to the measurement uncertainty. The present application uses the new information correction amount to correct the measurement noise covariance matrix and the process noise covariance matrix, which can more accurately reflect the noise characteristics and dynamic changes of the actual system, thereby improving the accuracy and robustness of state estimation. Ultimately, this state estimation method based on the new information correction amount not only improves the accuracy of power system state monitoring, but also enhances the system's adaptability to abnormal data and environmental changes, and realizes a more reliable and accurate estimation of the power system state, providing solid technical support for the safe, stable and optimized operation of the power grid.

[0026] As a preferred embodiment of the first aspect, the measurement noise covariance matrix and the process noise covariance matrix are corrected according to the innovation correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix, specifically:

[0027] Correcting the measurement noise covariance matrix according to the new information correction amount to obtain a corrected measurement noise covariance matrix;

[0028] Calculating the statistical distribution of a vector of new information between the predicted state value and the actual measured state value;

[0029] If the statistical distribution is greater than a preset threshold, the process noise covariance matrix is ​​corrected to obtain a corrected process noise covariance matrix.

[0030] The measurement noise covariance matrix is ​​corrected according to the new information correction amount to obtain the corrected measurement noise covariance matrix, which is specifically:

[0031] According to the robustness factor and the innovation correction amount, the measurement noise covariance matrix is ​​corrected to obtain a corrected measurement noise covariance matrix; wherein the robustness factor is:

[0032]

[0033] In the formula, Robustness factor The i-th element of is the i-th element of the standard innovation correction v, v N is the standardization of the new information correction value v, k0 is the upper threshold, k1 is the lower threshold;

[0034] The corrected measurement noise covariance matrix is:

[0035]

[0036] In the formula, is the corrected measurement noise covariance matrix, R k is the measurement noise covariance matrix, is the robustness factor.

[0037] If the statistical distribution is greater than a preset threshold, the process noise covariance matrix is ​​corrected to obtain a corrected process noise covariance matrix, which is specifically:

[0038] According to the weight factor and the new information correction amount, the process noise covariance matrix is ​​corrected to obtain the corrected process noise covariance matrix;

[0039] Wherein, the weight factor is:

[0040] d k =(1-b) / (1-b k ),

[0041] In the formula, b is the forgetting factor, d is k is the weight factor;

[0042] The modified process noise covariance matrix is:

[0043]

[0044] In the formula, the subscript k represents the time, Q kis the modified process noise covariance matrix, K k is the Kalman filter gain, v k is the new information correction amount, P k is the estimated error covariance matrix, F k is the state transfer matrix, d k is the weight factor.

[0045] In this preferred embodiment, the present application adjusts the measurement noise covariance matrix by using the new information correction amount, thereby achieving more accurate modeling of the measurement error of the power system, thereby improving the accuracy of state estimation. The new information correction amount reflects the deviation between the predicted state value and the actual measured value. By correcting the measurement noise covariance matrix, the impact of abnormal data on state estimation can be reduced and the anti-interference ability of the system can be enhanced. Furthermore, the present application can detect whether a sudden change occurs in the power system by analyzing the statistical distribution of the new information vector. This adaptive ability is crucial to dealing with the uncertainty in the operation of the power grid. When a sudden change in the system is detected, the present application further adjusts the process noise covariance matrix to reflect the changes in the dynamic characteristics of the system, so that the state estimation can adapt to the real-time changes of the power grid. Through this method, the present application not only improves the accuracy of state estimation, but also enhances the adaptability and robustness of the system to the dynamic changes of the power grid, providing a strong guarantee for the safe, reliable and economical operation of the power system.

[0046] In a second aspect, the present application provides a power system prediction auxiliary state estimation device based on deep learning. The power system prediction auxiliary state estimation device based on deep learning includes an acquisition module, an input and output module, a first calculation module, a correction module, and a second calculation module;

[0047] The acquisition module is used to acquire electrical data of the power system;

[0048] The input-output module is used to input the electrical data into a trained neural network so that the neural network outputs a predicted state value of the electrical system; wherein the neural network is constructed by combining a convolutional neural network with an attention mechanism and a long short-term memory network;

[0049] The first calculation module is used to calculate the innovation correction amount according to the innovation between the predicted state value and the actual measured state value;

[0050] The correction module is used to correct the measurement noise covariance matrix and the process noise covariance matrix according to the new information correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix;

[0051] The second calculation module is used to calculate the state estimation value of the first power system according to the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm.

[0052] This device uses five modules to divide the work and coordinate work to better estimate the state of the power system. This application uses a deep learning model that integrates convolutional neural networks, attention mechanisms, and long short-term memory networks to process the electrical data of the power system, thereby achieving high-precision prediction of the system state. By comparing the predicted state value and the actual measured value, calculating the new information correction amount, and adjusting the measurement and process noise covariance matrix accordingly, this method not only improves the accuracy of state prediction, but also enhances the system's adaptability to noise and mutations in actual operation. Finally, using the corrected covariance matrix and Kalman filtering algorithm, real-time and accurate estimation of the power system state is achieved to solve the problem that the prior art cannot accurately evaluate the state of the power system.

[0053] As a preferred embodiment of the second aspect, after obtaining the corrected measurement noise covariance matrix and the corrected process noise covariance matrix, the method further includes:

[0054] The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least square method.

[0055] In this preferred embodiment, the present application combines the predicted state value and the actual measured state value of the deep learning model, uses the modified measurement noise covariance matrix to improve the reliability of the measurement data, and the modified process noise covariance matrix to enhance the adaptability of the system dynamic characteristics, and further uses the least squares algorithm to optimize the real-time tracking capability of the power system state. This method calculates the initial state estimate value by the least squares method, ensuring the mathematical rigor of the estimation process, thereby achieving higher accuracy and robustness in the power system prediction auxiliary state estimation based on deep learning. In this way, the present invention not only improves the accuracy of state estimation, but also enhances the system's resistance to noise and system mutations, and ultimately achieves accurate grasp of the power system state, providing strong technical support for the stable operation and effective management of the power grid.

[0056] As a preferred embodiment of the second aspect, the measurement noise covariance matrix and the process noise covariance matrix are corrected according to the innovation correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix, specifically:

[0057] Correcting the measurement noise covariance matrix according to the new information correction amount to obtain a corrected measurement noise covariance matrix;

[0058] Calculating the statistical distribution of a vector of new information between the predicted state value and the actual measured state value;

[0059] If the statistical distribution is greater than a preset threshold, the process noise covariance matrix is ​​corrected to obtain a corrected process noise covariance matrix.

[0060] In this preferred embodiment, the present application adjusts the measurement noise covariance matrix by using the new information correction amount, thereby achieving more accurate modeling of the measurement error of the power system, thereby improving the accuracy of state estimation. The new information correction amount reflects the deviation between the predicted state value and the actual measured value. By correcting the measurement noise covariance matrix, the impact of abnormal data on state estimation can be reduced and the anti-interference ability of the system can be enhanced. Furthermore, the present application can detect whether a sudden change occurs in the power system by analyzing the statistical distribution of the new information vector. This adaptive ability is crucial to dealing with the uncertainty in the operation of the power grid. When a sudden change in the system is detected, the present application further adjusts the process noise covariance matrix to reflect the changes in the dynamic characteristics of the system, so that the state estimation can adapt to the real-time changes of the power grid. Through this method, the present application not only improves the accuracy of state estimation, but also enhances the adaptability and robustness of the system to the dynamic changes of the power grid, providing a strong guarantee for the safe, reliable and economical operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 : A flowchart of an embodiment of a power system prediction auxiliary state estimation method based on deep learning provided in this application;

[0062] Figure 2 : A schematic diagram of the structure of an embodiment of a convolutional long short-term memory neural network model with an attention mechanism added provided in the present application;

[0063] Figure 3 : A schematic diagram of the structure of an embodiment of the long short-term memory neural network provided by the present application;

[0064] Figure 4 : A structural diagram of an embodiment of the deep learning-based power system prediction auxiliary state estimation device provided in this application. DETAILED DESCRIPTION

[0065] 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.

[0066] Embodiment 1

[0067] Please refer to Figure 1 , which is a deep learning-based power system prediction auxiliary state estimation method provided in an embodiment of the present invention.

[0068] In this embodiment, the process of the power system prediction auxiliary state estimation method based on deep learning in this application is described in detail through steps S01-S04.

[0069] S01: Obtain electrical data of the power system.

[0070] As a preferred embodiment of implementation one, the acquiring of electrical data of the power system is specifically:

[0071] A massive amount of electrical data is obtained from the PMU. The electrical data contains the dynamic time correlation and spatial distribution characteristics of the power system based on physical constraints. By analyzing the electrical data, the present application can extract key spatiotemporal features and predict the operating status of the power system accordingly.

[0072] S02: Inputting the electrical data into a trained neural network so that the neural network outputs a predicted state value of the electrical system; wherein the neural network is constructed by combining a convolutional neural network with an attention mechanism and a long short-term memory network.

[0073] As a preferred embodiment of the first embodiment, the electrical data is input into a trained neural network so that the neural network outputs a predicted state value of the electrical system, specifically:

[0074] This application uses the electrical data obtained by the PMU as input and performs state prediction based on the convolutional long short-term memory neural network (CNN-LSTM-ATT) model with an attention mechanism. Figure 2 As shown in the figure, the model first captures the spatial features of the input sequence through a convolutional neural network, then captures the temporal correlation through a long short-term memory neural network, and finally uses the attention layer to better understand the internal associations of the sequence. This comprehensive structure can more comprehensively process sequence data with spatiotemporal correlation and obtain more accurate state prediction values.

[0075] This application uses a one-dimensional convolutional neural network to process time series data to effectively extract the spatial connection between different feature values ​​of the data. Through sparse connections and shared weights, convolutional neural networks can effectively extract significant feature information from input data and filter out other insignificant information, thereby reducing the number of parameters and computational complexity and improving the generalization ability of the model.

[0076] Convolutional neural networks mainly include convolutional layers and pooling layers. The function of the convolutional layer is to extract features from the input data. It contains multiple convolution kernels, which perform convolution operations on the input data of each given time step by sliding convolution to filter local information. The convolution operation is expressed as:

[0077]

[0078] In the formula, represents the convolution operation; σ represents the activation function in the convolution operation, and the ReLU function is used here; x c is the result of the convolution operation; x t-L+1,t is the input time series of length L; ω c and b c are the weight vector and bias vector respectively.

[0079] The feature information is then reduced in dimension through the pooling layer, which has the ability to filter the strong and weak features of the convolutional layer output results, thereby reducing the size of the parameter matrix and preventing overfitting. Finally, the padding layer can be used to adjust the size of the output information to ensure that the input and output remain the same size.

[0080] Based on the convolutional neural network, the input data passes through the convolution layer to amplify the features, and then passes through the maximum pooling layer to reduce the data information and filter out the information with less effect, so as to extract the spatial information from different features.

[0081] Recurrent neural networks are neural networks that are specifically designed to process sequence information. By adding a self-feedback connection, the current input and the previous state jointly determine the current output. However, when there are too many sequences to remember, gradient vanishing or gradient explosion will occur, resulting in a limited length of sequence to be processed. Long short-term memory neural networks use a gating mechanism to control the flow and memory of time series information, thereby extracting feature information of long time series. The structure of the long short-term memory neural network is shown in the figure below. Figure 3 shown.

[0082] The long short-term memory neural network uses hidden state h to store short-term memory and cell state c to store long-term memory. The current input state g t From the current input data x t and the hidden state h at the previous moment t-1 The calculation results are:

[0083] g t =tanh(ω m ·[h t-1 ,x t ]+b m ) (2)

[0084] In the formula, the tanh function is used to extract the candidate state at the current moment; ω f and b f are weight and bias respectively;

[0085] The gating mechanism mainly includes the forget gate, input gate and output gate. The forget gate controls how much information of the cell state at the previous moment needs to be retained; the input gate controls how much information of the input state at the current moment needs to be retained; the output gate controls how much information of the cell state at the current moment needs to be output to the hidden state at the current moment. The proportional coefficients of the forget gate, input gate and output gate are calculated from the current input and the hidden state at the previous moment:

[0086] f t =σ sigmoid (ω f ·[h t-1 ,x t ]+b f ) (3)

[0087] i t =σ sigmoid (ω i ·[h t-1 ,x t ]+b i ) (4)

[0088] o t =σ sigmoid (ω o ·[h t-1 ,x t ]+b o ) (5)

[0089] In the formula, σ represents the activation function. Here, the sigmoid function is used to ensure that the output is between [0, 1]; ω f ,ω i ,ω o and b f 、b i 、b o They are the weights and offsets of the forget gate, input gate, and output gate respectively;

[0090] Current cell state c t The update of is mainly achieved through the forget gate and the input gate, where the forget gate determines how much information to retain, and the input gate is responsible for adding new information to the cell state to form long-term memory. It is expressed as:

[0091] c t =f t ·c t-1 +i t ·g t (6)

[0092] The current hidden state h t The update is mainly realized through the output gate, which controls the cell state c at the current moment. t How much information needs to be output to the hidden state h at the current moment? t , and finally forms short-term memory, that is, the output of the long short-term memory neural network at the current moment:

[0093] h t =o t ·tanh(c t ) (7)

[0094] The attention mechanism can capture the correlation between elements within a sequence. When processing sequence data, each element can be associated with other elements in the sequence, rather than just relying on elements in adjacent positions. The model is optimized through the attention mechanism, and the relative importance between elements, that is, the probability distribution of attention, is calculated to select key features from a large number of features, thereby improving the computational efficiency of the model.

[0095] The output of the LSTM neural network is used as the input of the attention layer. For each element in the sequence, the self-attention mechanism calculates its similarity with other elements:

[0096] s t =σ tanh (h t ·ω h +b h ) (8)

[0097] In the formula, ω h and b h are the weight and bias of the attention mechanism respectively;

[0098] And normalize these similarities to attention weights:

[0099] α t =σ softmax (exp(s t ) / ∑exp(s t )) (9)

[0100] In the formula, σ represents the activation function, and the softmax function is used here for normalization.

[0101] Then, the output of the attention mechanism can be obtained by weighted summing each element with the corresponding attention weight:

[0102]

[0103] The fully connected layer takes the output of the attention layer Map it to the prediction result y and output the final state prediction value:

[0104]

[0105] In the formula, is the final output state prediction value; ω c and b c are weights and biases respectively.

[0106] S03: Calculating the new information correction amount according to the difference between the predicted state value and the actual measured state value;

[0107] According to the innovation correction amount, the measurement noise covariance matrix and the process noise covariance matrix are corrected to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix.

[0108] As a preferred embodiment of the first embodiment, the innovation correction amount is calculated according to the difference between the predicted state value and the actual measured state value, specifically:

[0109] When there is bad data in the measurement data, the deviation between the predicted value and the actual value will increase significantly. This phenomenon can be used to identify and deal with bad measurement data, because the predicted state value can essentially be regarded as an additional measurement reference. In addition, measurement value anomalies in power systems do not always mean data problems, they sometimes also indicate actual sudden changes in the power grid state, such as system failures or sharp changes in load.

[0110] The difference between the measured value and the state prediction is the new information, denoted as v k+1 , defined as follows:

[0111]

[0112] In the formula, is the measured predicted value calculated from the state predicted value:

[0113]

[0114] Similarly, it is reasonable to assume that the new information conforms to a zero-mean Gaussian distribution, and the covariance matrix of the new information can be calculated as follows:

[0115]

[0116] The standardized new information can be obtained as follows:

[0117]

[0118] Therefore, the standard innovation conforms to the normal distribution, and the standard innovation can be judged by statistical hypothesis testing. When the standard innovation of a certain measurement point exceeds the threshold, the measurement point can be judged as an outlier. If the measurement value is a single measurement point of a single plant or an independent topological node, these outliers can be considered as bad data; if most or all of the standard innovations of a plant or directly connected measurement value exceed the threshold, it may be a sudden change in the system state.

[0119] As a preferred embodiment of the first embodiment, the measurement noise covariance matrix and the process noise covariance matrix are corrected according to the innovation correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix, specifically:

[0120] First, the measurement noise co-defense error matrix is ​​corrected: In order to avoid the adverse effects of bad data on the state estimation results, the robustness factor is introduced to correct the measurement noise co-defense error matrix R and reduce the weight of bad data in the state estimation. The robustness factor is generated using the IGGⅢ weight function:

[0121]

[0122] In the formula, for The i-th element in ; k0 and k1 are control factors used to measure the quality of the measurement value. The degree of process noise correction can be appropriately adjusted according to the measurement redundancy.

[0123] Compared with the standard residual, the standard innovation takes into account the predicted state and avoids residual pollution. From the elements of the weight function, it can be seen that the weight of the measurement value in the state estimation depends on the size of the standard innovation. The larger the standard innovation, the smaller the weight of the corresponding measurement value. When the standard innovation is too large, the corresponding measurement value will be eliminated. The final corrected measurement noise covariance matrix is:

[0124]

[0125] When the system state changes suddenly, the accuracy and reliability of the predicted value will be greatly reduced, and the process noise covariance needs to be corrected to adjust the weight of the predicted value in the state estimation. To this end, the process noise covariance sequence is adjusted using the latest value of the new information. By introducing the fading memory exponential weighting method, the new data can be weighted according to the degree of change of the system, which can better reflect the real-time changes of the system process noise.

[0126] Process noise is the difference between the true state value and the predicted state value. Since the true state of the system is unknown, the estimated value of the state can be used to approximate the process noise:

[0127]

[0128] Where ε is the estimation error;

[0129] When the number of samples in the sample space is N, the unbiased estimate of the process noise covariance is:

[0130]

[0131] Among these N samplings, the latest sampling can better reflect the time-varying characteristics of the current process noise, so the weight factor d is introduced to give weight to the latest sampling. The weight factor d is:

[0132] d k =(1-b) / (1-b k ) (twenty one)

[0133] In the formula, b is the forgetting factor. The greater the change of system process noise, the smaller the value of b.

[0134] Therefore, the process noise covariance can be modified as:

[0135]

[0136] S04: Calculate and obtain a state estimation value of the first power system according to the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm.

[0137] As a preferred embodiment of the first embodiment, the state estimation value of the first power system is calculated based on the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm, specifically:

[0138] When the state prediction at time k+1 is obtained and the measured value z k+1 Afterwards, according to the Kalman filter algorithm, the calculation method of the system state value is as follows:

[0139]

[0140] In the formula, K k+1 That is the filter gain matrix, P k+1 To estimate the covariance matrix, is the corrected measurement noise covariance matrix. Using the measured values ​​and predicted values, a new state estimate is obtained, and the filter gain matrix is ​​also updated at the same time.

[0141] The covariance matrix of the predicted value can be given by the following formula according to the Kalman filter method:

[0142]

[0143] Where F is the state transfer matrix, which can be obtained by online estimation using the parameter estimation method to complete the update of the covariance matrix of the predicted value, and Q is the corrected process noise covariance matrix.

[0144] In this preferred embodiment, the present application adjusts the measurement noise covariance matrix by using the new information correction amount, thereby achieving more accurate modeling of the measurement error of the power system, thereby improving the accuracy of state estimation. The new information correction amount reflects the deviation between the predicted state value and the actual measured value. By correcting the measurement noise covariance matrix, the impact of abnormal data on state estimation can be reduced and the anti-interference ability of the system can be enhanced. Furthermore, the present application can detect whether a sudden change occurs in the power system by analyzing the statistical distribution of the new information vector. This adaptive ability is crucial to dealing with the uncertainty in the operation of the power grid. When a sudden change in the system is detected, the present application further adjusts the process noise covariance matrix to reflect the changes in the dynamic characteristics of the system, so that the state estimation can adapt to the real-time changes of the power grid. Through this method, the present application not only improves the accuracy of state estimation, but also enhances the adaptability and robustness of the system to the dynamic changes of the power grid, providing a strong guarantee for the safe, reliable and economical operation of the power system.

[0145] As a preferred embodiment of the first embodiment, after obtaining the corrected measurement noise covariance matrix and the corrected process noise covariance matrix, the method further includes:

[0146] The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least square method.

[0147] The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least squares method. Specifically, after the state prediction value is obtained, the state prediction value and the measured value at the latest moment can achieve the best state estimation. Based on the basic idea of ​​least squares, the following objective function is proposed on the basis of the static state estimation method:

[0148]

[0149] In the formula, z is the system measurement value, x is the system state value to be obtained, is the predicted value of the system state, H is the measurement matrix, reflecting the relationship between the measured value and the state value of the system, R is the measurement noise covariance matrix, M is the covariance matrix of the predicted value, when the objective function is minimized:

[0150]

[0151] In the formula, is the estimated value of the system state. Substituting the above formula into the solution, we can get the state estimate at the latest moment, as shown in the following formula:

[0152]

[0153] The above is a direct solution method for the state estimation value of the second power system, which does not require iterative calculation and has high calculation stability and calculation performance.

[0154] In this preferred embodiment, the present application combines the predicted state value and the actual measured state value of the deep learning model, uses the modified measurement noise covariance matrix to improve the reliability of the measurement data, and the modified process noise covariance matrix to enhance the adaptability of the system dynamic characteristics, and further uses the least squares algorithm to optimize the real-time tracking capability of the power system state. This method calculates the initial state estimate value by the least squares method, ensuring the mathematical rigor of the estimation process, thereby achieving higher accuracy and robustness in the power system prediction auxiliary state estimation based on deep learning. In this way, the present invention not only improves the accuracy of state estimation, but also enhances the system's resistance to noise and system mutations, and ultimately achieves accurate grasp of the power system state, providing strong technical support for the stable operation and effective management of the power grid.

[0155] This application uses a deep learning model that integrates convolutional neural networks, attention mechanisms, and long short-term memory networks to process electrical data of the power system, thereby achieving high-precision prediction of the system state. By comparing the predicted state value and the actual measured value, calculating the new information correction amount, and adjusting the measurement and process noise covariance matrix accordingly, this method not only improves the accuracy of state prediction, but also enhances the system's adaptability to noise and mutations in actual operation. Finally, using the corrected covariance matrix and Kalman filtering algorithm, real-time and accurate estimation of the power system state is achieved to solve the problem that the prior art cannot accurately evaluate the state of the power system.

[0156] Embodiment 2

[0157] Please refer to Figure 4 , which is a deep learning-based power system prediction auxiliary state estimation device provided in an embodiment of the present application.

[0158] In this embodiment, the power system prediction auxiliary state estimation device based on deep learning includes an acquisition module 10, an input and output module 20, a first calculation module 30, a correction module 40 and a second calculation module 50.

[0159] The acquisition module 10 is used to acquire electrical data of the power system.

[0160] As a preferred embodiment of implementation 2, the obtaining of electrical data of the power system is specifically:

[0161] A massive amount of electrical data is obtained from the PMU. The electrical data contains the dynamic time correlation and spatial distribution characteristics of the power system based on physical constraints. By analyzing the electrical data, the present application can extract key spatiotemporal features and predict the operating status of the power system accordingly.

[0162] The input-output module 20 is used to input the electrical data into a trained neural network so that the neural network outputs a predicted state value of the electrical system; wherein the neural network is constructed by combining a convolutional neural network with an attention mechanism and a long short-term memory network.

[0163] As a preferred embodiment of the second embodiment, the electrical data is input into a trained neural network so that the neural network outputs a predicted state value of the electrical system, specifically:

[0164] This application uses the electrical data obtained by the PMU as input and performs state prediction based on the convolutional long short-term memory neural network (CNN-LSTM-ATT) model with an attention mechanism. Figure 2 As shown in the figure, the model first captures the spatial features of the input sequence through a convolutional neural network, then captures the temporal correlation through a long short-term memory neural network, and finally uses the attention layer to better understand the internal associations of the sequence. This comprehensive structure can more comprehensively process sequence data with spatiotemporal correlation and obtain more accurate state prediction values.

[0165] This application uses a one-dimensional convolutional neural network to process time series data to effectively extract the spatial connection between different feature values ​​of the data. Through sparse connections and shared weights, convolutional neural networks can effectively extract significant feature information from input data and filter out other insignificant information, thereby reducing the number of parameters and computational complexity and improving the generalization ability of the model.

[0166] Convolutional neural networks mainly include convolutional layers and pooling layers. The function of the convolutional layer is to extract features from the input data. It contains multiple convolution kernels, which perform convolution operations on the input data of each given time step by sliding convolution to filter local information. The convolution operation is expressed as:

[0167]

[0168] In the formula, represents the convolution operation; σ represents the activation function in the convolution operation, and the ReLU function is used here; x c is the result of the convolution operation; x t-L+1,t is the input time series of length L; ω c and bc are the weight vector and bias vector respectively.

[0169] The feature information is then reduced in dimension through the pooling layer, which has the ability to filter the strong and weak features of the convolutional layer output results, thereby reducing the size of the parameter matrix and preventing overfitting. Finally, the padding layer can be used to adjust the size of the output information to ensure that the input and output remain the same size.

[0170] Based on the convolutional neural network, the input data passes through the convolution layer to amplify the features, and then passes through the maximum pooling layer to reduce the data information and filter out the information with less effect, so as to extract the spatial information from different features.

[0171] Recurrent neural networks are neural networks that are specifically designed to process sequence information. By adding a self-feedback connection, the current input and the previous state jointly determine the current output. However, when there are too many sequences to remember, gradient vanishing or gradient explosion will occur, resulting in a limited length of sequence to be processed. Long short-term memory neural networks use a gating mechanism to control the flow and memory of time series information, thereby extracting feature information of long time series. The structure of the long short-term memory neural network is shown in the figure below. Figure 3 shown.

[0172] The long short-term memory neural network uses hidden state h to store short-term memory and cell state c to store long-term memory. The current input state g t From the current input data x t and the hidden state h at the previous moment t-1 The calculation results are:

[0173] g t =tanh(ω m ·[h t-1 ,x t ]+b m ) (13)

[0174] In the formula, the tanh function is used to extract the candidate state at the current moment; ω f and b f are weight and bias respectively;

[0175] The gating mechanism mainly includes the forget gate, input gate and output gate. The forget gate controls how much information of the cell state at the previous moment needs to be retained; the input gate controls how much information of the input state at the current moment needs to be retained; the output gate controls how much information of the cell state at the current moment needs to be output to the hidden state at the current moment. The proportional coefficients of the forget gate, input gate and output gate are calculated from the current input and the hidden state at the previous moment:

[0176] f t =σ sigmoid (ω f·[h t-1 ,x t ]+b f ) (14)

[0177] i t =σ sigmoid (ω i ·[h t-1 ,x t ]+b i ) (15)

[0178] o t =σ sigmoid (ω o ·[h t-1 ,x t ]+b o ) (16)

[0179] In the formula, σ represents the activation function. Here, the sigmoid function is used to ensure that the output is between [0, 1]; ω f ,ω i ,ω o and b f 、b i 、b o are the weights and offsets of the forget gate, input gate, and output gate respectively;

[0180] Current cell state c t The update of is mainly achieved through the forget gate and the input gate, where the forget gate determines how much information to retain, and the input gate is responsible for adding new information to the cell state to form long-term memory. It is expressed as:

[0181] c t =f t ·c t-1 +i t ·g t (17)

[0182] The current hidden state h t The update is mainly realized through the output gate, which controls the cell state c at the current moment. t How much information needs to be output to the hidden state h at the current moment? t , and finally forms short-term memory, that is, the output of the long short-term memory neural network at the current moment:

[0183] h t =o t ·tanh(c t ) (18)

[0184] The attention mechanism can capture the correlation between elements within a sequence. When processing sequence data, each element can be associated with other elements in the sequence, rather than just relying on elements in adjacent positions. The model is optimized through the attention mechanism, and the relative importance between elements, that is, the probability distribution of attention, is calculated to select key features from a large number of features, thereby improving the computational efficiency of the model.

[0185] The output of the LSTM neural network is used as the input of the attention layer. For each element in the sequence, the self-attention mechanism calculates its similarity with other elements:

[0186] s t =σ tanh (h t ·ω h +b h ) (19)

[0187] In the formula, ω h and b h are the weight and bias of the attention mechanism respectively;

[0188] And normalize these similarities to attention weights:

[0189] α t =σ softmax (exp(s t ) / ∑exp(s t )) (20)

[0190] In the formula, σ represents the activation function, and the softmax function is used here for normalization.

[0191] Then, the output of the attention mechanism can be obtained by weighted summing each element with the corresponding attention weight:

[0192]

[0193] The fully connected layer takes the output of the attention layer Map it to the prediction result y and output the final state prediction value:

[0194]

[0195] In the formula, is the final output state prediction value; ω c and b c are weights and biases respectively.

[0196] The first calculation module 30 is used to calculate the new information correction value according to the difference between the predicted state value and the actual measured state value;

[0197] The correction module 40 is used to correct the measurement noise covariance matrix and the process noise covariance matrix according to the innovation correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix.

[0198] As a preferred embodiment of the second embodiment, the innovation correction amount is calculated according to the difference between the predicted state value and the actual measured state value, specifically:

[0199] When there is bad data in the measurement data, the deviation between the predicted value and the actual value will increase significantly. This phenomenon can be used to identify and deal with bad measurement data, because the predicted state value can essentially be regarded as an additional measurement reference. In addition, measurement value anomalies in power systems do not always mean data problems, they sometimes also indicate actual sudden changes in the power grid state, such as system failures or sharp changes in load.

[0200] The difference between the measured value and the state prediction is the new information, denoted as v k+1 , defined as follows:

[0201]

[0202] In the formula, is the measured predicted value calculated from the state predicted value:

[0203]

[0204] Similarly, it is reasonable to assume that the new information conforms to a zero-mean Gaussian distribution, and the covariance matrix of the new information can be calculated as follows:

[0205]

[0206] The standardized new information can be obtained as follows:

[0207]

[0208] Therefore, the standard innovation conforms to the normal distribution, and the standard innovation can be judged by statistical hypothesis testing. When the standard innovation of a certain measurement point exceeds the threshold, the measurement point can be judged as an outlier. If the measurement value is a single measurement point of a single plant or an independent topological node, these outliers can be considered as bad data; if most or all of the standard innovations of a plant or directly connected measurement value exceed the threshold, it may be a sudden change in the system state.

[0209] As a preferred embodiment of the second embodiment, the measurement noise covariance matrix and the process noise covariance matrix are corrected according to the innovation correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix, specifically:

[0210] First, the measurement noise co-defense error matrix is ​​corrected: In order to avoid the adverse effects of bad data on the state estimation results, the robustness factor is introduced to correct the measurement noise co-defense error matrix R and reduce the weight of bad data in the state estimation. The robustness factor is generated using the IGGⅢ weight function:

[0211]

[0212] In the formula, for The i-th element in ; k0 and k1 are control factors used to measure the quality of the measurement value. The degree of process noise correction can be appropriately adjusted according to the measurement redundancy.

[0213] Compared with the standard residual, the standard innovation takes into account the predicted state and avoids residual pollution. From the elements of the weight function, it can be seen that the weight of the measurement value in the state estimation depends on the size of the standard innovation. The larger the standard innovation, the smaller the weight of the corresponding measurement value. When the standard innovation is too large, the corresponding measurement value will be eliminated. The final corrected measurement noise covariance matrix is:

[0214]

[0215] When the system state changes suddenly, the accuracy and reliability of the predicted value will be greatly reduced, and the process noise covariance needs to be corrected to adjust the weight of the predicted value in the state estimation. To this end, the process noise covariance sequence is adjusted using the latest value of the new information. By introducing the fading memory exponential weighting method, the new data can be weighted according to the degree of change of the system, which can better reflect the real-time changes of the system process noise.

[0216] Process noise is the difference between the true state value and the predicted state value. Since the true state of the system is unknown, the estimated value of the state can be used to approximate the process noise:

[0217]

[0218] Where ε is the estimation error;

[0219] When the number of samples in the sample space is N, the unbiased estimate of the process noise covariance is:

[0220]

[0221] Among these N samplings, the latest sampling can better reflect the time-varying characteristics of the current process noise, so the weight factor d is introduced to give weight to the latest sampling. The weight factor d is:

[0222] d k =(1-b) / (1-b k ) (twenty one)

[0223] In the formula, b is the forgetting factor. The greater the change of system process noise, the smaller the value of b.

[0224] Therefore, the process noise covariance can be modified as:

[0225]

[0226] The second calculation module 50 is used to calculate the state estimation value of the first power system according to the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm.

[0227] As a preferred embodiment of the second embodiment, the state estimation value of the first power system is calculated based on the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm, specifically:

[0228] When the state prediction at time k+1 is obtained and the measured value z k+1 Afterwards, according to the Kalman filter algorithm, the calculation method of the system state value is as follows:

[0229]

[0230]

[0231] In the formula, K k+1 That is the filter gain matrix, P k+1 To estimate the covariance matrix, is the corrected measurement noise covariance matrix. Using the measured values ​​and predicted values, a new state estimate is obtained, and the filter gain matrix is ​​also updated at the same time.

[0232] The covariance matrix of the predicted value can be given by the following formula according to the Kalman filter method:

[0233]

[0234] Where F is the state transfer matrix, which can be obtained by online estimation using the parameter estimation method to complete the update of the covariance matrix of the predicted value, and Q is the corrected process noise covariance matrix.

[0235] In this preferred embodiment, the present application adjusts the measurement noise covariance matrix by using the new information correction amount, thereby achieving more accurate modeling of the measurement error of the power system, thereby improving the accuracy of state estimation. The new information correction amount reflects the deviation between the predicted state value and the actual measured value. By correcting the measurement noise covariance matrix, the impact of abnormal data on state estimation can be reduced and the anti-interference ability of the system can be enhanced. Furthermore, the present application can detect whether a sudden change occurs in the power system by analyzing the statistical distribution of the new information vector. This adaptive ability is crucial to dealing with the uncertainty in the operation of the power grid. When a sudden change in the system is detected, the present application further adjusts the process noise covariance matrix to reflect the changes in the dynamic characteristics of the system, so that the state estimation can adapt to the real-time changes of the power grid. Through this method, the present application not only improves the accuracy of state estimation, but also enhances the adaptability and robustness of the system to the dynamic changes of the power grid, providing a strong guarantee for the safe, reliable and economical operation of the power system.

[0236] As a preferred embodiment of the second embodiment, after obtaining the corrected measurement noise covariance matrix and the corrected process noise covariance matrix, the method further includes:

[0237] The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least square method.

[0238] The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least squares method. Specifically, after the state prediction value is obtained, the state prediction value and the measured value at the latest moment can achieve the best state estimation. Based on the basic idea of ​​least squares, the following objective function is proposed on the basis of the static state estimation method:

[0239]

[0240] In the formula, z is the system measurement value, x is the system state value to be obtained, is the system state prediction value, H is the measurement matrix, which reflects the relationship between the system measurement value and the state value, R is the measurement noise covariance matrix, and M is the covariance matrix of the prediction value. When the objective function is minimized, we have:

[0241]

[0242] In the formula, is the estimated value of the system state. Substituting the above formula into the solution, we can get the state estimate at the latest moment, as shown in the following formula:

[0243]

[0244] The above is a direct solution method for the state estimation value of the second power system, which does not require iterative calculation and has high calculation stability and calculation performance.

[0245] In this preferred embodiment, the present application combines the predicted state value and the actual measured state value of the deep learning model, uses the modified measurement noise covariance matrix to improve the reliability of the measurement data, and the modified process noise covariance matrix to enhance the adaptability of the system dynamic characteristics, and further uses the least squares algorithm to optimize the real-time tracking capability of the power system state. This method calculates the initial state estimate value by the least squares method, ensuring the mathematical rigor of the estimation process, thereby achieving higher accuracy and robustness in the power system prediction auxiliary state estimation based on deep learning. In this way, the present invention not only improves the accuracy of state estimation, but also enhances the system's resistance to noise and system mutations, and ultimately achieves accurate grasp of the power system state, providing strong technical support for the stable operation and effective management of the power grid.

[0246] This device uses five modules to divide the work and coordinate work to better estimate the state of the power system. This application uses a deep learning model that integrates convolutional neural networks, attention mechanisms, and long short-term memory networks to process the electrical data of the power system, thereby achieving high-precision prediction of the system state. By comparing the predicted state value and the actual measured value, calculating the new information correction amount, and adjusting the measurement and process noise covariance matrix accordingly, this method not only improves the accuracy of state prediction, but also enhances the system's adaptability to noise and mutations in actual operation. Finally, using the corrected covariance matrix and Kalman filtering algorithm, real-time and accurate estimation of the power system state is achieved to solve the problem that the prior art cannot accurately evaluate the state of the power system.

[0247] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A power system prediction auxiliary state estimation method based on deep learning, characterized in that: include: Acquire electrical data of power systems; Inputting the electrical data into a trained neural network so that the neural network outputs a predicted state value of the electrical system; wherein the neural network is constructed by combining a convolutional neural network with an attention mechanism and a long short-term memory network; Calculating the new information correction amount according to the difference between the predicted state value and the actual measured state value; According to the new information correction amount, the measurement noise covariance matrix and the process noise covariance matrix are corrected to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix; The state estimation value of the first power system is calculated based on the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm.

2. The power system prediction auxiliary state estimation method based on deep learning according to claim 1 is characterized in that: After obtaining the modified measurement noise covariance matrix and the modified process noise covariance matrix, the method further includes: The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least square method.

3. The power system prediction auxiliary state estimation method based on deep learning according to claim 2 is characterized in that: The state estimation value of the power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least squares method, specifically: The first calculation formula of the least squares method is: In the formula, z is the actual measured state value, x is the estimated state value of the power system, is the system state prediction value, H is the measurement matrix, which reflects the relationship between the system measurement value and the state value, R is the measurement noise covariance matrix, and M is the covariance matrix of the prediction value; When the first calculation formula takes the minimum value, a second calculation formula is obtained, and the second formula is: According to the second calculation formula, an estimated value of the state of the power system is calculated.

4. The power system prediction auxiliary state estimation method based on deep learning according to claim 1, characterized in that: The new information correction amount is calculated based on the difference between the predicted state value and the actual measured state value, specifically: The innovation correction amount is calculated according to the covariance matrix of the innovation correction amount.

5. The power system prediction auxiliary state estimation method based on deep learning according to claim 1 is characterized in that: According to the innovation correction amount, the measurement noise covariance matrix and the process noise covariance matrix are corrected to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix, which are specifically: According to the new information correction amount, the measurement noise covariance matrix is ​​corrected to obtain a corrected measurement noise covariance matrix; Calculating the statistical distribution of a vector of new information between the predicted state value and the actual measured state value; If the statistical distribution is greater than a preset threshold, the process noise covariance matrix is ​​corrected to obtain a corrected process noise covariance matrix.

6. The power system prediction auxiliary state estimation method based on deep learning according to claim 5 is characterized in that: The measurement noise covariance matrix is ​​corrected according to the new information correction amount to obtain the corrected measurement noise covariance matrix, which is specifically: According to the robustness factor and the innovation correction amount, the measurement noise covariance matrix is ​​corrected to obtain a corrected measurement noise covariance matrix; wherein the robustness factor is: In the formula, Robustness factor The i-th element of is the i-th element of the standard innovation correction v, v N is the standardization of the new information correction value v, k0 is the upper threshold, k1 is the lower threshold; The corrected measurement noise covariance matrix is: In the formula, is the corrected measurement noise covariance matrix, R k is the measurement noise covariance matrix, is the robustness factor.

7. The power system prediction auxiliary state estimation method based on deep learning according to claim 5 is characterized in that: If the statistical distribution is greater than a preset threshold, the process noise covariance matrix is ​​corrected to obtain a corrected process noise covariance matrix, which is specifically: According to the weight factor and the new information correction amount, the process noise covariance matrix is ​​corrected to obtain the corrected process noise covariance matrix; Wherein, the weight factor is: d k =(1-b) / (1-b k ), In the formula, b is the forgetting factor, d is k is the weight factor; The modified process noise covariance matrix is: In the formula, the subscript k represents the time, Q k is the corrected process noise covariance matrix, K k is the Kalman filter gain, v k is the new information correction amount, P k is the estimated error covariance matrix, F k is the state transfer matrix, d k is the weight factor.

8. A power system prediction auxiliary state estimation device based on deep learning, characterized in that: It includes an acquisition module, an input and output module, a first calculation module, a correction module and a second calculation module; The acquisition module is used to acquire electrical data of the power system; The input-output module is used to input the electrical data into a trained neural network so that the neural network outputs a predicted state value of the electrical system; wherein the neural network is constructed by combining a convolutional neural network with an attention mechanism and a long short-term memory network; The first calculation module is used to calculate the innovation correction amount according to the innovation between the predicted state value and the actual measured state value; The correction module is used to correct the measurement noise covariance matrix and the process noise covariance matrix according to the new information correction amount to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix; The second calculation module is used to calculate the state estimation value of the first power system according to the corrected measurement noise covariance matrix, the corrected process noise covariance matrix and the Kalman filter algorithm.

9. The power system prediction auxiliary state estimation device based on deep learning according to claim 8, characterized in that: After obtaining the modified measurement noise covariance matrix and the modified process noise covariance matrix, the method further includes: The state estimation value of the second power system is calculated based on the predicted state value, the actual measured state value, the corrected measurement noise covariance matrix and the least square method.

10. The power system prediction auxiliary state estimation device based on deep learning according to claim 8, characterized in that: According to the innovation correction amount, the measurement noise covariance matrix and the process noise covariance matrix are corrected to obtain a corrected measurement noise covariance matrix and a corrected process noise covariance matrix, which are specifically: According to the new information correction amount, the measurement noise covariance matrix is ​​corrected to obtain a corrected measurement noise covariance matrix; Calculating the statistical distribution of a vector of new information between the predicted state value and the actual measured state value; If the statistical distribution is greater than a preset threshold, the process noise covariance matrix is ​​corrected to obtain a corrected process noise covariance matrix.

Images