A dual-attention CNN-BiLSTM method for power system state estimation

Abstract

This invention discloses a dual-attention CNN-BiLSTM method for power system state estimation. This method combines a convolutional neural network (CNN), a bidirectional long short-term memory neural network (BiLSTM), and an attention mechanism to form a dual-attention CNN-BiLSTM neural network, which improves model performance and enables rapid and accurate state estimation under simulated actual measurements. The method consists of two stages: an offline stage, where the model is built, parameters are set, and historical measurement datasets are input for training, and the resulting state estimation model is saved; and an online estimation stage, where real-time measurement data is input into the trained model for state estimation, yielding the state estimation results. Finally, the effectiveness of the proposed method is verified through simulation. A detailed and objective experimental analysis using a test simulation model demonstrates the effectiveness and feasibility of the proposed method.

Description

Technical Field

[0001] This invention relates to the fields of power technology and information technology, and in particular to a power system state estimation method based on dual attention CNN-BiLSTM. Background Technology

[0002] With the development and advancement of renewable energy sources such as wind, nuclear, and photovoltaic power, building a new power system with new energy sources as the mainstay and accelerating the green and low-carbon transformation have become hot topics. However, as the proportion of new energy sources connected to the grid increases, while alleviating the energy shortage problem, the power system is becoming increasingly complex and large-scale, leading to increasingly complex system operation and control methods, which poses a great challenge to the safe and stable operation of the power system.

[0003] To ensure the safe, economical, and reliable operation of the power system, accurate forecasting, sound decision-making, precise control, and lean management of the entire power system are essential requirements. This necessitates the development and construction of an Energy Management System (EMS) capable of unified management and scientific dispatch. However, the realization of the EMS's multi-functionality requires accurate and comprehensive grid operation data. In reality, inaccurate or insufficient measurement data inevitably occurs, preventing the collected data from fully reflecting the current operating state of the power system. For example, during the transmission of various measurement information to the EMS through transmission channels, noise and instrument errors in the measuring devices themselves can lead to the generation of bad data. Therefore, power system state estimation is needed to estimate the collected measurement information to obtain real-time and reliable power system operation data, ensuring the normal operation of the power system.

[0004] Currently, state estimation methods based on Weighted Least Squares (WLS) are widely used. Under ideal conditions where measurement noise follows a Gaussian distribution, the calculated result is the unbiased minimum variance. However, real measurement data contains noise and bad data, making WLS calculations difficult to converge and resulting in large errors. To address this problem, scholars both domestically and internationally have proposed robust state estimation methods, which improve the robustness of state estimation by changing the objective function. These include Weighted Least Absolute Value (WLAV) estimation, Quadratic Linear (QL) estimation, Quadratic Constant (QC) estimation, and Exponential Objective Function (MES) estimation. Although these methods are better at handling bad data than WLS, they suffer from complex objective functions and difficulty in solving them. To address this, some literature proposes a dynamic partitioning robust estimation method that identifies and partitions the measurement data, adaptively selecting either WLS or WLAV for estimation, improving efficiency while maintaining good accuracy and robustness. Other literature proposes transforming WLAV estimation into a bilinear problem through nonlinear transformation, thereby improving solution efficiency.

[0005] However, the traditional state estimation methods mentioned above rely on line parameters and the accuracy of measurement data, resulting in relatively long iteration times. Furthermore, they cannot fully extract valuable information from historical data for state estimation given the massive amounts of multi-source data in current power systems. With the development of various machine learning algorithms, using artificial intelligence to solve complex problems in power systems has become increasingly popular, and state estimation is no exception. How to extract valuable information from the massive amounts of data in power system databases and use data-driven methods to solve state estimation problems has gradually become a hot research topic. Summary of the Invention

[0006] In view of this, the purpose of this invention is to propose a power system state estimation method based on dual attention CNN-BiLSTM that is reliable in implementation, provides good reference results, and is highly feasible.

[0007] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:

[0008] A power system state estimation method based on dual-attention CNN-BiLSTM, comprising:

[0009] S01. Perform power flow calculations from the historical operating load of the power system to obtain multi-section historical power flow data, and then use it as the true value;

[0010] S02. Based on the historical tidal data of multiple sections, Gaussian white noise is added to generate multi-section measurement information;

[0011] S03. Perform data normalization, dimension transformation and data tiling preprocessing on the multi-section measurement information. Then, according to the 7:3 division ratio, extract 70% of the data as the training set and 30% of the data as the test set.

[0012] S04. Construct a CNN-BiLSTM model. Before training the model, set its network structure and model parameters.

[0013] S05. Train the constructed CNN-BiLSTM model using the training set, calculate the estimation result and actual value calculation error of a single iteration, then update the model weights, and then perform the next iteration by shuffling the training set data until the preset number of iterations is reached, at which point the training ends. Then, import the test set into the trained CNN-BiLSTM model for testing. When the test results meet the requirements, save the CNN-BiLSTM model.

[0014] S06. Input the real-time cross-sectional measurement information of the power system into the trained CNN-BiLSTM model to obtain the current system state of the power system.

[0015] As one possible implementation, solution S06 further includes:

[0016] The system state variables output by the CNN-BiLSTM model are compared with the true values, performance metrics are calculated, and the results are output.

[0017] As a possible implementation, further, in this scheme S04, the constructed CNN-BiLSTM model includes the SE-CNN model and the FA-BiLSTM model;

[0018] Among them, the SE-CNN model is formed by adding a channel attention mechanism to the SENet module in the CNN network, which is used to extract the spatial features of the data;

[0019] The FA-BiLSTM model is formed by adding an FA model to a BiLSTM network, which is used to bidirectionally mine the temporal features of data.

[0020] As a preferred implementation option, the SE-CNN model described in this scheme performs classification convolution operations and compression excitation processing on the input data;

[0021] In the classification convolution operation, the input data X is divided into g groups, which are operated on with the same number of parameters to generate output feature maps.

[0022] During the excitation squeeze process, the output feature maps are stacked to obtain an unweighted feature map U. Then, through the data squeezing operation, the information contained in each channel is aggregated in spatial dimensions to form a multidimensional statistic containing information on the importance of each channel. After excitation and scaling, channel weights are assigned and the result is output.

[0023] As a preferred implementation option, this scheme preferably employs global average pooling and global max pooling during the squeezing excitation processing. The data obtained from each pooling are then summed. The pooling process involves compressing the unweighted feature map U along its spatial dimension H×W to obtain the multidimensional statistic z∈R. C The c-th element is represented as:

[0024]

[0025] In equation (1), u c (i,j) represents the element in the i-th row and j-th column of the input data;

[0026] During stimulus processing, a parameterized gating mechanism is constructed using two fully connected layers. The first fully connected layer reduces the input C-dimensional statistic z to C / r dimension, and the second fully connected layer increases it to C dimension, i.e.:

[0027]

[0028] In equation (2), W1 and W2 are the parameters of the two fully connected layers, r is the scaling parameter, and s is the weight representing the importance of each channel in the feature map U.

[0029] During scaling, the obtained weight s is multiplied by the unweighted output U along each channel to form the final output.

[0030]

[0031]

[0032] As a preferred implementation option, the FA-BiLSTM model in this scheme performs scaling and transformation processing on the data output by the SE-CNN model;

[0033] The scaling process involves aggregating the information contained in the feature map output by the SE-CNN model through two-dimensional convolution to form a one-dimensional statistic containing the importance information of each feature on the feature map. During this process, the input C1'-dimensional features... Figure X 1. The dimensionality is reduced to a one-dimensional attention weight map e containing feature information using two-dimensional convolution operations, which is represented as:

[0034] e = sigmoid(W e X1+b e (5)

[0035] In equation (5), e is the combination of attention weight coefficients corresponding to a single feature on the input feature map; W e b are the parameters of the convolutional layer. e The bias vector to be set;

[0036] Then input features Figure X Multiplying the weighted feature map e by the obtained weighted feature map yields the weighted feature map U1:

[0037] U1 = eX (6)

[0038] The transformation process involves normalizing the features of the obtained weighted feature map using the Softmax function to reduce the influence of data dimensions on the output and prevent overfitting. The normalization formula for a single feature x is:

[0039]

[0040] The normalized result x' is passed through the activation function swi sh to add nonlinear factors in order to solve the problem of insufficient expressive power of the linear model and obtain a new weighted feature map U2.

[0041] Then, U2 is grouped and convolved, and pointwise convolution is performed using the VAIL calculation method to further extract relevant features. These features are then normalized (followed by a Swish activation function) to obtain the final output feature map.

[0042] Finally, feature maps The data is flattened using the flinten function to obtain a feature sequence with dynamically assigned weights, which is then input into the Bi LSTM model layer.

[0043] This proposed method combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) neural networks, and an attention mechanism to form a dual-attention CNN-BiLSTM neural network. This improves model performance and enables rapid and accurate state estimation under simulated real-world measurements. The method consists of two stages: an offline stage, where the model is built, parameters are set, and historical measurement datasets are input for training, resulting in a saved state estimation model; and an online estimation stage, where real-time measurement data is input into the trained model for state estimation, yielding the estimated state results. Finally, simulations validate the effectiveness of the proposed method. Detailed and objective experimental analysis using a test simulation model demonstrates the effectiveness and feasibility of the proposed method.

[0044] Because the safe and stable operation of power systems requires accurate and reliable data for EMS dispatch and management, power system state estimation is essential. However, traditional state estimation methods rely on line parameters and measurement data, resulting in relatively long iteration times. Furthermore, given the massive amounts of multi-source data in current power systems, they cannot fully extract valuable information from historical data for state estimation. Therefore, this technical solution trains a state estimation model by fitting the mapping relationship between input measurements and output state variables, extracting and utilizing valuable information from historical data. It omits the traditional method's analysis of the internal mechanisms of the power system and eliminates the need for line parameters in the solution process. Performing state estimation under simulated actual measurements yields accurate results quickly and demonstrates better estimation accuracy, robustness, and computational efficiency compared to traditional methods.

[0045] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0046] (1) This technical solution does not rely on line parameters. By mining the mapping pattern between measurement and state variables in historical data to train the model, it can effectively utilize the effective information in historical data to perform real-time state estimation and has good estimation accuracy and robustness, providing a reference for future power system state estimation.

[0047] (2) The present solution is tested under simulated actual measurement and bad data conditions to fully verify that the model does not require any noise preprocessing measures and is superior to the traditional WLS algorithm and other data-driven algorithms in terms of estimation accuracy and robustness.

[0048] (3) The calculation time of the online estimation of this technical solution is less affected by the expansion of the system scale, while the calculation time of the traditional method varies greatly. This indicates that the calculation efficiency of this technical solution is more suitable for the needs of large-scale power grid state estimation and has good development prospects. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a simplified implementation flowchart of the proposed method.

[0051] Figure 2 This is a structural diagram of the CNN-BiLSTM network model of the proposed method;

[0052] Figure 3 This is a schematic diagram of the SE-CNN network module structure in the CNN-BiLSTM network model of this scheme;

[0053] Figure 4 This is a schematic diagram of the FA-BiLSTM network module structure in the CNN-BiLSTM network model of this scheme;

[0054] Figure 5 This is a comparison of the state estimation results of 33 nodes at 50 sections in the comparative test of this scheme;

[0055] Figure 6 This is a comparison of the absolute errors of the state estimation results for section 50 in the comparative test of this scheme;

[0056] Figure 7 This is a comparison of the absolute error of the state estimation results for the 150-200 section in the comparative test of this scheme;

[0057] Figure 8 This is a comparison of the state estimation results of 50 sections and 33 nodes containing bad data in the comparative test of this scheme;

[0058] Figure 9This is a comparison of the absolute errors of the state estimation results for 50 sections containing bad data in the comparative test of this scheme;

[0059] Figure 10 This is a comparison of the absolute error of the state estimation results for the 150-200 section containing bad data in the comparative test of this scheme. Detailed Implementation

[0060] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] This embodiment presents a power system state estimation method based on dual-attention CNN-BiLSTM, which includes:

[0062] S01. Perform power flow calculations from the historical operating load of the power system to obtain multi-section historical power flow data, and then use it as the true value;

[0063] S02. Based on the historical tidal data from multiple sections, Gaussian white noise with a mean of 0 and a standard deviation of 2% is added to generate multi-section measurement information.

[0064] S03. Perform data normalization (to unify data units and prevent overfitting), dimension transformation, and data tiling on the multi-section measurement information (to generate two-dimensional data that conforms to convolution, since a two-dimensional grouped convolutional layer is followed). Then, according to the 7:3 division ratio, extract 70% of the data as the training set and 30% of the data as the test set.

[0065] S04. Construct a CNN-BiLSTM model. Before training the model, set its network structure and model parameters.

[0066] S05. Train the constructed CNN-BiLSTM model using the training set, calculate the estimation result and actual value calculation error of a single iteration, then update the model weights, and then perform the next iteration by shuffling the training set data until the preset number of iterations is reached, at which point the training ends. Then, import the test set into the trained CNN-BiLSTM model for testing. When the test results meet the requirements, save the CNN-BiLSTM model.

[0067] S06. Input the real-time cross-sectional measurement information of the power system into the trained CNN-BiLSTM model to obtain the current system state of the power system.

[0068] To facilitate model evaluation, S06 of this scheme also includes:

[0069] The system state variables output by the CNN-BiLSTM model are compared with the true values, and the performance metrics mean absolute error (MAE), maximum absolute error (MaxAE), and root mean square error (RMSE) are calculated to evaluate the model's state estimation performance. The formulas for calculating the three metrics are as follows:

[0070]

[0071]

[0072]

[0073] In the formula, Let x be the estimated value of the i-th state vector at section t. i (t) represents the true value of the i-th state vector at section t, and T represents the total number of time sections in the test set.

[0074] This approach can also compare and verify the performance metrics of the above three metrics with those calculated from the output results of SE-CNN, FA-BiLSTM, and WLS.

[0075] Based on the above technologies, the CNN-BiLSTM model constructed in this solution S04 includes the SE-CNN model and the FA-BiLSTM model;

[0076] Among them, the SE-CNN model is formed by adding a channel attention mechanism to the SENet module in the CNN network, which is used to extract the spatial features of the data;

[0077] The FA-BiLSTM model is formed by adding an FA model to a BiLSTM network, which is used to bidirectionally mine the temporal features of data.

[0078] This approach enhances the correlation between input and target features in the CNN-BiLSTM network through a dual attention mechanism, thereby improving model performance.

[0079] As a preferred implementation option, the SE-CNN model described in this scheme performs classification convolution operations and compression excitation processing on the input data;

[0080] In the classification convolution operation, the input data X is divided into g groups, which are operated on with the same number of parameters to generate output feature maps (which can generate output feature maps that are g times larger than those of regular convolution);

[0081] During the squeeze excitation process, the output feature maps are stacked to obtain an unweighted feature map U. Then, through the data squeeze operation, the information contained in each channel is aggregated in spatial dimensions to form a multidimensional statistic containing the importance information of each channel. After excitation and scaling processing, channel weights are assigned and the result is output.

[0082] As a preferred implementation option, this scheme preferably employs global average pooling and global max pooling for the squeeze stimulus processing. The data obtained from each pooling method are then summed. The pooling process involves compressing the unweighted feature map U along its spatial dimension H×W to obtain the multidimensional statistics z∈R. C The c-th element is represented as:

[0083]

[0084] In equation (1), u c (i,j) represents the element in the i-th row and j-th column of the input data;

[0085] During excitation processing, a parameterized gating mechanism is constructed using two fully connected layers. The first fully connected layer reduces the input C-dimensional statistic z to C / r dimension (followed by a ReLU activation function), and the second fully connected layer increases it to C dimension (followed by a si gmoid activation function).

[0086]

[0087] In equation (2), W1 and W2 are the parameters of the two fully connected layers, r is the scaling parameter, and s is the weight representing the importance of each channel in the feature map U.

[0088] During scaling, the obtained weight s is multiplied by the unweighted output U channel by channel to form the final output.

[0089]

[0090]

[0091] As a preferred implementation option, the FA-BiLSTM model in this scheme performs scaling and transformation processing on the data output by the SE-CNN model;

[0092] The scaling process involves aggregating the information contained in the feature map output by the SE-CNN model through two-dimensional convolution to form a one-dimensional statistic containing the importance information of each feature on the feature map. During this process, the input C1'-dimensional features... Figure X1. A two-dimensional convolution operation (followed by a si gmoid activation function) is used to reduce the dimensionality to a one-dimensional attention weight map e containing feature information, which is represented as:

[0093] e = sigmoid(W e X1+b e (5)

[0094] In equation (5), e is the combination of attention weight coefficients corresponding to a single feature on the input feature map; W e b are the parameters of the convolutional layer. e The bias vector to be set;

[0095] Then input features Figure X Multiplying the weighted feature map e by the obtained weighted feature map yields the weighted feature map U1:

[0096] U1 = eX (6)

[0097] The transformation process involves normalizing the features of the obtained weighted feature map using the Softmax function to reduce the influence of data dimensions on the output and prevent overfitting. The normalization formula for a single feature x is:

[0098]

[0099] The normalized result x' is passed through the activation function swi sh to add nonlinear factors in order to solve the problem of insufficient expressive power of the linear model and obtain a new weighted feature map U2.

[0100] Then, U2 is grouped and convolved, and pointwise convolution is performed using the VAIL calculation method to further extract relevant features. These features are then normalized (followed by a Swish activation function) to obtain the final output feature map.

[0101] Finally, feature maps The data is flattened using the flinten function to obtain a feature sequence with dynamically assigned weights, which is then input into the Bi LSTM model layer.

[0102] The data processing logic principle of the model in this paper is as follows: The input node active power and branch active power data are n-dimensional one-dimensional data.

[0103] 1) Multi-channel two-dimensional data is generated through data preprocessing and input into a two-dimensional grouped convolutional layer. Using the same number of parameters as regular convolution, two-dimensional grouped convolution can generate g times the two-dimensional feature maps, which is helpful for subsequent feature extraction;

[0104] 2) The input pooling layer is used to extract features to obtain a one-dimensional feature vector. Then, through the self-selection gate mechanism composed of two fully connected layers, a one-dimensional weight vector with the same dimension as the input feature channel is generated. This vector is used to represent the importance of the feature channel.

[0105] 3) Multiply the obtained one-dimensional weight vector with the corresponding feature channels of the input data to obtain the weighted feature map. This enhances the feature channels that are useful for the current task and suppresses the feature channels that are not very useful for the current task, thus solving the problem that it is difficult to characterize the correlation between channels in conventional CNNs.

[0106] 4) The weighted feature map is then processed using a two-dimensional conventional convolution to aggregate the information contained in the feature map, forming a one-dimensional statistic that includes information on the importance of each feature on the feature map. This statistic is then multiplied with the input feature map to obtain the weighted feature map. This achieves weighting of individual features on a single channel, effectively solving the problem of lost feature association information that occurs during the selection of input features using traditional feature correlation analysis methods.

[0107] 5) Normalize the feature map that performs single feature weighting to prevent overfitting, and input it into a two-dimensional grouped convolutional layer (followed by normalization, to unify the data units of the data that generate multiple feature maps again to prevent overfitting) for point-by-point convolution to further extract features. Finally, input the obtained multi-channel feature map into a flattening layer to obtain a one-dimensional feature vector that has undergone channel weighting and feature weighting through a dual attention mechanism.

[0108] 6) Input the one-dimensional weighted feature vector into the Bi LSTM layer to bidirectionally mine the temporal features of the data and output a one-dimensional vector.

[0109] 7) Finally, the output one-dimensional vector is input into two fully connected layers for classification to obtain the output results of the corresponding dimension.

[0110] Comparative Test

[0111] This scheme uses the IEEE 33-node system as the simulation system and London, UK household load data as the historical database. A sample section is used every 5 minutes, generating 3000 sets of actual loads. The power flow results obtained by reducing the actual loads across multiple sections to nodal loads are used as the true values. Verification is performed by adding different levels of Gaussian noise to the true power flow values ​​to simulate actual measurements and bad data. The training and test sets are divided in a 7:3 ratio. The specific measurement configuration involves injecting active and reactive power into all nodes and the active and reactive power at the head of branches 1-2, 2-3, 4-5, 7-8, 9-10, 12-13, 15-16, 20-21, 24-25, 26-27, and 29-30. The models are trained to output nodal voltage amplitude and nodal voltage phase angle. For ease of comparison, the test set measurement data is used as the actual measurement data for testing. Existing WLS, SE-CNN, and FA-BiLSTM models are used as comparison objects, and the following comparisons are performed.

[0112] (1) Estimation accuracy analysis

[0113] Gaussian noise with a mean of 0 and a standard deviation of 2% is added to multi-section power flow data to simulate actual measurements, and the proposed method is compared with WLS, SE-CNN, FA-BiLSTM and other methods.

[0114] State estimation was performed using data from the IEEE 33-node system at the 50th randomly selected cross-section. The results are as follows: Figure 5 As shown.

[0115] from Figure 5 As can be seen, the state estimation results of the method of the present invention have high accuracy. To more intuitively compare the differences between different methods, the state estimation performance is represented by mean absolute error, maximum absolute error, and root mean square error.

[0116] Figure 6 The absolute error diagram is obtained by calculating the state estimation results of each method and the true value of the power flow under the 50th section.

[0117] Depend on Figure 6 The data shows that the state estimation method proposed in this invention has a relatively small overall error in a single data cross-section under Gaussian white noise. Table 1 shows the state estimation results of different algorithms for IEEE 33 nodes in the 50th data cross-section:

[0118] Table 1. State estimation results and indices for section 50

[0119]

[0120] Comparing the mean absolute errors of the proposed method with WLS in Table 1, the voltage amplitude and voltage phase angle errors of the overall state at the 50th data segment are reduced by 56.07% and 40.52%, respectively, compared to WLS. The other two error metrics of the proposed method are also lower than those of WLS. While the mean absolute errors of SE-CNN and FA-BiLSTM, which extract single features, are lower than those of WLS, their maximum absolute errors can exceed those of WLS, potentially leading to significant deviations in the state estimation results. The proposed method addresses this drawback compared to methods that extract single features, and exhibits a smaller overall estimation error. Therefore, the proposed method significantly improves the state estimation accuracy at a single data segment compared to the traditional WLS method and other data-driven algorithms.

[0121] To verify the accuracy of the method of the present invention under different data sections, its accuracy is determined by taking partial cross-sectional data of a single node. Figure 7 The absolute error diagram is obtained by calculating the state estimation results of various methods and the true power flow values ​​for randomly selected node 30 in data sections 150-200.

[0122] Depend on Figure 7 The data shows that the state estimation method proposed in this invention has a relatively small overall error across multiple data sections under Gaussian white noise. Table 2 shows the state estimation results of node 30 using different algorithms in the 150-200 data section:

[0123] Table 2. State estimation results and indices for section 150-200

[0124]

[0125]

[0126] By comparing the mean absolute errors of the method of this invention with those of WLS in Table 2, the voltage amplitude and voltage phase angle errors of node 30 in the 150-200 data cross-section are reduced by 25.47% and 18.66% respectively compared to WLS, and the other two error indicators are also lower than those of WLS. Furthermore, all error indicators are lower than those of the two methods for extracting single features. It can be concluded that the state estimation accuracy of the method of this invention in multiple data cross-sections is significantly improved compared to the traditional WLS method and other data-driven algorithms.

[0127] (2) Robustness analysis

[0128] Since bad data inevitably exists in power system measurement data, this invention simulates bad data in real-world measurements by adding measurement noise with a standard deviation of 10% to some branches. Branches 29-30 (including both active and reactive power) are added, and the noise added to the remaining branches is the same as in the previous section.

[0129] Among them, the data state estimation results of the IEEE 33-node system under the 50th section containing bad data are as follows: Figure 8 As shown.

[0130] from Figure 8 As can be seen, even with bad data, the state estimation results of the method of the present invention still have high accuracy.

[0131] Figure 9 The absolute error diagram is obtained by calculating the state estimation results of each method and the true power flow value under the 50th section containing bad data.

[0132] Depend on Figure 9 The data shows that the state estimation method proposed in this invention has a smaller overall error in a single data section when bad data is present. Table 3 shows the state estimation results of different algorithms for IEEE 33 nodes under the 50th data section containing bad data.

[0133] Table 3. State estimation results for 50 sections containing bad data.

[0134]

[0135] As shown in Table 3, when bad data is present, the voltage amplitude and voltage phase angle errors of the method of this invention are smaller than those of the other three methods. Furthermore, comparing the mean absolute errors of each method with those in the previous section, the errors of WLS with bad data are increased by 118.55% and 61.79% respectively compared to those with 2% Gaussian noise, while the errors of the method of this invention are increased by 31.77% and 9.15%. It can be concluded that the method proposed in this invention has higher accuracy and robustness in estimating the state of a single data cross-section in system measurements with bad data than traditional WLS and other data-driven algorithms.

[0136] Figure 10 The absolute error map is obtained by calculating the state estimation results of node 30 containing bad data using various methods and the true value of the power flow in data section 150-200.

[0137] Depend on Figure 10 The data shows that the state estimation method proposed in this invention has a smaller overall error across multiple data sections when bad data is present. Table 4 shows the state estimation results of node 30 with bad data under different algorithms in the 150-200 data section:

[0138] Table 4. State estimation results for the 150-200 section containing bad data.

[0139]

[0140] As shown in Table 4, under different data cross-sections containing bad data, the voltage amplitude and voltage phase angle errors of the method proposed in this invention are smaller than those of the other three methods. Furthermore, comparing the mean absolute error of each method with that in the previous section, the error of WLS with bad data increases by 115.12% and 85.30% respectively compared to the error with 2% Gaussian noise, while the error of the method proposed in this invention increases by 14.39% and 19.41%. It can be concluded that the method proposed in this invention has higher accuracy and robustness in state estimation under multiple data cross-sections containing bad data than traditional WLS and other data-driven algorithms.

[0141] (3) Computational efficiency analysis

[0142] The estimation time was tested using two methods on different systems. When the test system was scaled up from a 33-node system to a 118-node system, the estimation time of the method in this invention increased from 0.0469s to 0.0566s, a 20.68% increase, while the estimation time of the traditional WLS algorithm increased from 0.0321s to 0.0941s, a 193.14% increase. This shows that the estimation time of the physical model-based WLS changes significantly with the increase in system size; while the method in this invention is more affected by the scale change in offline training time, with little change in online estimation time. The excessively long offline training time can be addressed using multi-threading. This indicates that the method is more efficient than traditional methods in large-scale power system state estimation.

[0143] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A power system state estimation method using dual-attention CNN-BiLSTM, characterized in that, It includes: S01. Perform power flow calculations from the historical operating load of the power system to obtain multi-section historical power flow data, and then use it as the true value; S02. Based on the historical tidal data of multiple sections, Gaussian white noise is added to generate multi-section measurement information; S03. Perform data normalization, dimension transformation and data tiling preprocessing on the multi-section measurement information. Then, according to the 7:3 division ratio, extract 70% of the data as the training set and 30% of the data as the test set. S04. Construct a CNN-BiLSTM model. Before training the model, set its network structure and model parameters. S05. Train the constructed CNN-BiLSTM model using the training set, calculate the estimation result and actual value calculation error of a single iteration, then update the model weights, and then perform the next iteration by shuffling the training set data until the preset number of iterations is reached, at which point the training ends. Then, import the test set into the trained CNN-BiLSTM model for testing. When the test results meet the requirements, save the CNN-BiLSTM model. S06. Input the real-time cross-sectional measurement information of the power system into the trained CNN-BiLSTM model to obtain the current system state of the power system. In S04, the constructed CNN-BiLSTM model includes the SE-CNN model and the FA-BiLSTM model; Among them, the SE-CNN model is formed by adding a channel attention mechanism to the SENet module in the CNN network, which is used to extract the spatial features of the data; The FA-BiLSTM model is formed by adding an FA model to a BiLSTM network, and it is used to bidirectionally mine the temporal features of data. The SE-CNN model performs classification convolution operations and compression excitation processing on the input data; In the classification convolution operation, the input data X is divided into g groups, which are operated on with the same number of parameters to generate output feature maps. During the excitation squeezing process, the output feature maps are stacked to obtain an unweighted feature map U. Then, through the data squeezing operation, the information contained in each channel is aggregated in spatial dimension to form a multidimensional statistic containing the importance information of each channel. After excitation and scaling processing, channel weights are assigned and the output is then processed. In the squeeze excitation processing, global average pooling and global max pooling are used. The data after pooling by these two methods are added together. The pooling process involves compressing the unweighted feature map U along its spatial dimension H×W to obtain the multidimensional statistic z∈R. C The c-th element is represented as: In equation (1), u c (i,j) represents the element in the i-th row and j-th column of the input data; During stimulus processing, a parameterized gating mechanism is constructed using two fully connected layers. The first fully connected layer reduces the input C-dimensional statistic z to C / r dimension, and the second fully connected layer increases it to C dimension, i.e.: In equation (2), W1 and W2 are the parameters of the two fully connected layers, r is the scaling parameter, and s is the weight representing the importance of each channel in the feature map U. During scaling, the obtained weight s is multiplied by the unweighted output U along each channel to form the final output.

2. The power system state estimation method using dual-attention CNN-BiLSTM as described in claim 1, characterized in that, S06 also includes: The system state variables output by the CNN-BiLSTM model are compared with the true values, performance metrics are calculated, and the results are output.

3. The power system state estimation method using dual-attention CNN-BiLSTM as described in claim 1, characterized in that, The FA-BiLSTM model performs scaling and transformation processing on the data output by the SE-CNN model. The scaling process involves aggregating the information contained in the feature map output by the SE-CNN model through two-dimensional convolution to form a one-dimensional statistic containing the importance information of each feature on the feature map. During this process, the input C1'-dimensional feature map X1 is reduced to a one-dimensional attention weight map e containing feature information using two-dimensional convolution operations, which is represented as: e=sigmoid(W e X1+b e ) (5) In equation (5), e is the combination of attention weight coefficients corresponding to a single feature on the input feature map; W e b are the parameters of the convolutional layer. e The bias vector to be set; Then, the input feature map X is multiplied by the obtained weight map e to obtain the weighted feature map U1: U1 = eX (6) The transformation process involves normalizing the features of the obtained weighted feature map using the Softmax function to reduce the influence of data dimensions on the output and prevent overfitting. The normalization formula for a single feature x is: The normalized result x' is passed through the activation function swish to add nonlinear factors in order to solve the problem of insufficient expressive power of the linear model and obtain a new weighted feature map U2. Then, U2 is grouped and convolved, and pointwise convolution is performed using the Vaild calculation method to further extract relevant features. After normalization, the final output feature map is obtained. Finally, feature maps The data is flattened using the flatten function to obtain a feature sequence with dynamically assigned weights, which is then input into the BiLSTM model layer.

Images