A method for data recovery in power distribution networks in response to spoofing data injection attacks.
The Spatiotemporal Mamba Graph Convolutional Network (ST-MambaGCN) addresses the issues of real-time performance, integrity, and physical consistency in data recovery from Data Fictitious Injection Attacks (FDIA) in power distribution networks. It enables collaborative recovery of measurement data, improving the reliability and engineering applicability of the recovery results.
Patent Information
- Application Number
- CN202611361252.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
AI Technical Summary
Existing data recovery methods for distribution network spurious data injection attacks (FDIA) have poor adaptability to topology changes, parameter uncertainties, or complex measurement noise, high computational complexity, and are difficult to meet real-time recovery requirements. Furthermore, the recovery results fail to fully integrate the long-term evolution patterns of measurement data, the topological coupling relationships between nodes, and the correlation characteristics between multiple physical quantities, which affects the reliability of subsequent services.
The Spatiotemporal Mamba Graph Convolutional Network (ST-MambaGCN) is adopted. By constructing an input mapping layer, a temporal feature extraction module, a spatial feature extraction module, a temporal attention aggregation layer, and a multi-path recovery output head, and combining the physical constraint term in the loss function, the temporal dynamic features and spatial topological correlations of distribution network measurement data are recovered collaboratively.
It achieves accurate recovery of measurement data of attacked nodes and their related branches. The recovery results are consistent at both the data and physical levels, reducing computational complexity and improving recovery accuracy and generalization ability under different attack scenarios, thus ensuring the safe and stable operation of the distribution network.
Smart Images

Figure CN123122941A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power distribution network security protection technology, and in particular to a method for recovering power distribution network data in response to false data injection attacks. Background Technology
[0002] With the continuous advancement of new power system construction, distribution networks are rapidly evolving towards deep cyber-physical integration. The accuracy of measurement data such as node voltage and power directly determines the effectiveness of dispatching decisions, including power flow calculation, economic dispatch, and security analysis. False data injection attacks (FDIA) can effectively circumvent traditional malicious data detection mechanisms by tampering with power system measurement data, thereby misleading state estimation results and posing a serious threat to the safe and stable operation of the distribution network.
[0003] Currently, existing methods for FDIA data recovery in distribution networks are mainly divided into optimization-based recovery methods and data-driven recovery methods. Optimization-based methods typically model data recovery as an optimization problem with power flow constraints or state variable boundary constraints; data-driven methods, such as GAN, Transformer, CNN, and LSTM, can utilize historical measurement data to learn the nonlinear distribution characteristics of the system's operating state and predict and reconstruct attacked measurements. However, existing methods still have the following shortcomings:
[0004] (1) Optimization-based recovery methods usually rely on accurate distribution network topology, line parameters and power flow models. When the system topology changes, parameters are uncertain or measurement noise is complex, the recovery results are easily affected by model errors and have poor adaptability. (2) Some optimization-based recovery methods need to solve non-convex optimization problems with power flow constraints, which have high computational complexity and are difficult to meet the real-time data recovery requirements of large-scale distribution networks after being subjected to FDIA. (3) Most existing data-driven methods independently model time dynamic features or node spatial correlation features, and fail to fully integrate the long-term evolution law of measurement data, the topological coupling relationship between nodes and the correlation characteristics between multiple physical quantities, which limits the data recovery capability under complex attack scenarios. (4) Most methods lack explicit constraints on the physical consistency of the recovery results, which makes it difficult to ensure that the recovered measurement data meets the AC power flow equation and power balance relationship, affecting the reliability of subsequent online power flow calculation, economic dispatch and security analysis.
[0005] Therefore, a data recovery method for distribution network spoofed data injection (FDIA) attacks based on spatiotemporal Mamba graph convolutional networks (ST-MambaGCN) is needed to address the above-mentioned shortcomings, while taking into account the requirements of real-time performance, integrity, spatiotemporal coordination, and physical consistency in distribution network FDIA data recovery. Summary of the Invention
[0006] The purpose of this invention is to propose a method for data recovery in power distribution networks against spoofed data injection attacks, comprising:
[0007] Collect FDIA detection datasets from the power distribution network and preprocess the datasets;
[0008] A spatiotemporal Mamba graph convolutional network is constructed, consisting of an input mapping layer, a temporal feature extraction module, a spatial feature extraction module, a temporal attention aggregation layer, and a multi-path recovery output head connected in series.
[0009] The spatiotemporal Mamba graph convolutional network is trained in the following ways: During training, the input samples first pass through an input mapping layer, which maps the original multidimensional measurement features to a latent space suitable for deep feature extraction. Then, the temporal feature extraction module learns the dynamic evolution of distribution network measurement data within historical time windows, capturing the temporal deviation features between attacked and normal measurements. Next, the spatial feature extraction module models the spatial coupling relationship between nodes based on the distribution network topology, enabling the association information between attacked nodes and their adjacent nodes and related branches to participate in the recovery decision. Finally, the key features at different time steps are weighted and fused by a temporal attention aggregation layer, and the data recovery results are output by multiple recovery output heads.
[0010] The data recovery performance of spatiotemporal Mamba graph convolutional networks is evaluated.
[0011] Furthermore, the FDIA inspection dataset of the distribution network is collected, and the dataset is preprocessed, specifically including:
[0012] Based on the distribution network topology and power operator load data, time-series load scenarios are generated, and historical measurement data of distribution network operation are collected.
[0013] Construct FDIA attack vectors and superimpose random offsets that follow a uniform distribution onto the measurement data of the corresponding nodes. Design FDIA attack scenarios with different attack intensities, attack ratios, and number of attack nodes.
[0014] Measurement data including normal and attack scenarios are labeled and divided into training, validation and test sets according to proportions;
[0015] Historical measurement data are standardized to eliminate dimensional differences, and the original time-series measurement data are transformed into fixed-length window samples using a sliding time window method.
[0016] Furthermore, historical measurement data includes three types of node measurements: voltage amplitude at each node, active power injected into the node, and reactive power; and four types of branch measurements: active and reactive power flow at the beginning of the line, and active and reactive power flow at the end of the line.
[0017] Furthermore, the input mapping layer is composed of linear transformation, layer normalization, and ReLU activation function concatenated. For each scenario to be recovered, based on the current time t, the historical node measurement sequence of the previous W time steps and the measurement at the current time are extracted and concatenated in the time dimension to form a complete time series input of length W+1. The historical node measurement sequence provides the model with a time series prior reference, and the measurement at the current time step contains the attack deviation information to be recovered. The node feature dimension of each time step in the concatenated time series input is D0, corresponding to three types of measurements: node voltage amplitude, node injected active power, and reactive power.
[0018] Furthermore, the temporal feature extraction module is composed of several Mamba blocks connected in series. In each Mamba block, the input features are normalized in the first layer and then divided into two paths, a signal branch and a gated branch, through gated projection. The signal branch is processed by SSM and then multiplied element-wise with the soft gated vector obtained by sigmoid activation of the gated branch. After output projection and second-layer normalization, the output is connected through residuals.
[0019] Furthermore, the feature matrix of the (l+1)th layer node in the spatial feature extraction module is represented as follows:
[0020] ;
[0021] in, Let l be the feature matrix of the nodes in the l-th layer. For learnable weights, It is the ReLU activation function. This is the symmetric normalized adjacency matrix of node relationships.
[0022] Furthermore, the attention weights of the temporal attention aggregation layer are generated by the linear layer through Softmax normalization. Features at different times are assigned different weights based on their relevance to the current recovery task, and then weighted and summed to form a shape... Feature tensor compression to Node-level feature representation.
[0023] Furthermore, each output head of the multi-path recovery output head is composed of a linear layer, a layer normalization layer, a ReLU activation layer, and an output linear layer connected in series.
[0024] Furthermore, training the spatiotemporal Mamba graph convolutional network also includes:
[0025] The Adam optimizer combined with the backpropagation algorithm is used to iteratively update the parameters of the spatiotemporal Mamba graph convolutional network. The model recovery performance in each round is evaluated using the validation set. The training stability is improved by combining adaptive learning rate adjustment and early stopping mechanism. The model with the smallest recovery error in the validation set and the best performance in satisfying physical constraints is saved as the inference model for online recovery of distribution network measurement data after FDIA attack.
[0026] Furthermore, the loss function for training the spatiotemporal Mamba graph convolutional network is:
[0027] ;
[0028] ;
[0029] in, , , , , , , The data fitting losses for four types of power flow measurements—node voltage, node injected active power, node injected reactive power, and branch head-to-end flow—are all implemented using the Smooth L1 loss function; the threshold parameter β... huber =1, the absolute value of the error is less than β. huber When calculated in L2 form, it is greater than β. huber Calculated in L1 form. The auxiliary supervision loss for voltage residual prediction is α, with a weighting coefficient of α. The supervision loss for the confidence output head is β, with a weighting coefficient of β. The physical constraint loss is represented by the weighting coefficient λ. , Inject the recovery value directly predicted by the power output head into the node; , The value recovered from the node voltage The calculated theoretical injection estimate, , , , The predicted recovery value for the branch power flow output head. , , , The value recovered from the node voltage , The calculated theoretical power flow estimates for the first and last branches; N is the number of nodes, and L is the number of branches; , As weight.
[0030] The beneficial effects of this invention are as follows:
[0031] 1. The method of this invention designs a spatiotemporal Mamba graph convolutional network (ST-MambaGCN) and constructs a data recovery framework for distribution network FDIA. Based on the obtained attack detection and location results, it can accurately recover and reconstruct the measurement data of the attacked node and its related branches. By comprehensively utilizing the Mamba module's ability to model the dynamic evolution of long-term time-series measurement data and the graph convolutional network's ability to extract the topological spatial coupling relationship between distribution network nodes, it can achieve collaborative recovery of multiple types of measurement data, overcoming the shortcomings of existing methods that only focus on node measurements or model a single spatiotemporal feature.
[0032] 2. The method of this invention introduces physical constraint terms of AC power flow equations into the loss function, so that the recovery results not only closely approximate the actual measured values at the data level, but also satisfy the power balance relationship of the distribution network at the physical level, thereby improving the physical consistency and engineering usability of the recovery results.
[0033] 3. Compared with existing optimization-based recovery methods, the method of this invention avoids the high computational overhead caused by complex non-convex optimization problems. Compared with conventional data-driven recovery methods, this method can simultaneously take into account temporal dynamic characteristics, spatial topological correlation, and physical constraint consistency. It has good recovery accuracy and generalization ability under different attack intensities, number of attacking nodes, attack ratios, and distribution network scale scenarios. It can effectively reduce the continuous impact of FDIA on subsequent online power flow calculation, economic dispatch, and security analysis, and provide reliable support for the safe recovery and stable operation of distribution network cyber-physical systems after being subjected to FDIA. Attached Figure Description
[0034] Figure 1 A flowchart illustrating a method for data recovery in a power distribution network against a fake data injection attack.
[0035] Figure 2 This is the structure of the ST-MambaGCN model.
[0036] Figure 3 (a), (b), and (c) represent the node measurement recovery results for voltage amplitude, active power, and reactive power in the IEEE 33-node system, respectively.
[0037] Figure 4 Verification of KCL power balance residuals in the IEEE 33-bus system.
[0038] Figure 5 To improve the model recovery performance under different numbers of attacking nodes in the IEEE 33-node system.
[0039] Figure 6Performance metrics analysis of ST-MambaGCN and benchmark models. Detailed Implementation
[0040] This invention proposes a method for data recovery in power distribution networks in response to spoofed data injection attacks. The invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0041] This invention provides a method for FDIA data recovery of power distribution networks based on a spatiotemporal Mamba graph convolutional network (ST-MambaGCN), taking into account the following key factors:
[0042] Temporal characteristics of measurement data: Measurement data such as voltage and power at distribution network nodes change dynamically over time, and anomalies caused by FDIA exhibit temporal evolution patterns;
[0043] Spatial topology characteristics of distribution networks: There is spatial coupling relationship between the same type of measurement data in different nodes, and multi-node coordinated attacks will show specific abnormal distributions at the topology level;
[0044] Physical consistency of the restored results: The restored measurement data should not only be close to the actual operating state in terms of numerical value, but also satisfy the power balance relationship described by the AC power flow equations to ensure that the restored results have engineering usability;
[0045] Diversity of attack scenarios: FDIA attacks vary in terms of attack intensity, number of attacking nodes, attack ratio, and target distribution network scale. Data recovery methods need to have good generalization ability and scenario adaptability to meet the reliable recovery requirements of attacked measurement data under different FDIA attack conditions.
[0046] Figure 1 This is a flowchart illustrating a method for data recovery in distribution networks against spoofed data injection attacks. It achieves reliable data recovery from FDIA (Digital Data Interpretation) attacks in distribution networks. This method fully leverages the temporal characteristics, spatial topology features, and physical consistency of distribution network measurement data, addressing the technical limitations of existing methods. The steps are as follows:
[0047] Step 1: Construct and preprocess the FDIA detection dataset for the power distribution network.
[0048] 1. Measurement Data Acquisition: Based on the distribution network topology and PJM power operator load data, time-series load scenarios are generated. Simultaneously, historical measurement data of the distribution network operation are collected, including three types of node measurements: voltage amplitude at each node, active power injected into the node, and reactive power; and four types of branch measurements: active and reactive power flow at the beginning and end of the line.
[0049] 2. Attack Scenario Construction: Construct FDIA attack vectors, and superimpose uniformly distributed random offsets onto the measurement data of the corresponding nodes. Design FDIA attack scenarios with different attack intensities (c=0.01 / 0.015 / 0.02 / 0.025 / 0.03 / 0.05), attack ratios (15% / 30%), and the number of attack nodes (1-5).
[0050] 3. Dataset labeling and partitioning: The measurement data containing normal and attack scenarios are labeled, with 0 representing normal samples and 1 representing attacked samples. The dataset is divided into training set, validation set, and test set in a ratio of 70%:15%:15%.
[0051] 4. Data Standardization and Windowing: Seven types of features, including node voltage amplitude, node injected active power, and reactive power, are standardized to eliminate dimensional differences. The sliding time window method is used to transform the original time series measurement data into fixed-length window samples, enabling the model to perceive the time evolution of the measurement data.
[0052] Step 2: Construct the Spatiotemporal Mamba Graph Convolutional Network (ST-MambaGCN)
[0053] The ST-MambaGCN model combines state-space sequence modeling with graph convolution spatial feature extraction, enabling it to simultaneously capture the dynamic evolution of distribution network measurement data over time and the spatial topological coupling relationships between nodes, automatically learning the differences between tampered measurements and normal operating states. The model takes labeled measurement sequences containing historical time-series windows as input, recovering measurements of nodes marked as attacked and their adjacent branches, while keeping the original values of measurements of non-attacked nodes unchanged. The model consists of five sequentially connected parts: an input mapping layer, a temporal feature extraction module, a spatial feature extraction module, a temporal attention aggregation layer, and a multi-path recovery output header.
[0054] 1. Input Mapping Layer: The input mapping layer is responsible for mapping the time-series measurement inputs to a high-dimensional latent feature space. For each scenario to be recovered, the model uses the current time t as a reference, extracts the historical node measurement sequences of the previous W time steps and the current time measurement, and concatenates them in the time dimension to form a complete time-series input of length W+1. The historical sequences provide the model with a time-series prior reference, while the current time measurement contains the attack bias information to be recovered. The node feature dimension of each time step in the concatenated time-series input is D0, corresponding to three types of measurements: node voltage amplitude, node injected active power, and reactive power. The mapping layer consists of a linear transformation, layer normalization, and a ReLU activation function concatenated to map the original feature dimension D0 to the hidden dimension D:
[0055] ;
[0056] Where B is the batch size, W is the historical window length, and M is the number of nodes. The concatenated temporal input tensor The latent feature tensor after mapping;
[0057] 2. Temporal Feature Extraction Module: The temporal feature extraction module consists of multiple Mamba blocks strung together to model the long-range dependencies of measurement data over time. In each Mamba block, input features... After the first layer of normalization, the signal branch and the gated branch are split into two paths through gated projection. The signal branch is processed by SSM and then multiplied element-wise with the soft-gated vector obtained by the gated branch through Sigmoid activation, achieving adaptive filtering of historical state information. Finally, after output projection and the second layer of normalization, the output is connected through residual connections to ensure the effective propagation of gradients in the deep network.
[0058] ;
[0059] in, and These are the input and output feature tensors of the l-th Mamba block, respectively. and These are signal branches and gated branches, This indicates element-wise multiplication. This represents the Sigmoid activation function;
[0060] 3. Spatial Feature Extraction Module: The spatial feature extraction module consists of a multi-layer graph convolutional network that explicitly models the spatial coupling relationships between distribution network nodes. In FDIA data recovery scenarios, it enables the features of each node to fuse with the state information of its topological neighbors, thereby perceiving the indirect impact of attacking nodes on the measurements of neighboring nodes through the distribution network topology. Batch normalization and Dropout regularization are applied after each graph convolution layer to ensure training stability. Stacking multiple layers of graph convolutions expands the model's receptive field to higher-order domains, providing sufficient spatial context support for measurement data recovery.
[0061] ;
[0062] in, Let l be the feature matrix of the nodes in the l-th layer. For learnable weights, It is the ReLU activation function. This is the symmetric normalized adjacency matrix of node relationships;
[0063] 4. Temporal Attention Aggregation Layer: After the temporal and spatial features are extracted by the above two-level modules, the temporal dimension is weighted and compressed through the temporal attention aggregation layer. Attention weights are generated by the linear layer through Softmax normalization. Features at different times are assigned different weights based on their relevance to the current recovery task, and then weighted and summed to form a shape... Feature tensor compression to Node-level feature representation;
[0064] 5. Multi-channel recovery output head: The multi-channel recovery output head completes the recovery prediction of seven types of measurements based on node-level features. Each output head consists of a linear layer, layer normalization, ReLU activation, and an output linear layer. For node voltage amplitude, the model adopts a residual prediction strategy, outputting the corrected deviation between the current attacked measurement and the true value, and superimposing it on the original attacked measurement to obtain the recovered value, thereby improving the recovery accuracy of small voltage deviations.
[0065] Step 3: Training the ST-MambaGCN model
[0066] Based on the constructed and preprocessed distribution network FDIA data recovery dataset, the ST-MambaGCN model was trained under supervision. Standardized historical measurement samples organized using a sliding time window were input into the ST-MambaGCN model. The input samples contained seven types of measurement data, including node voltage amplitudes at each time step, along with the distribution network topology and attack node labels. The sample labels were the uncontaminated real measurement data corresponding to the attack scenario, used to supervise the model's learning of the recovery mapping relationship between attacked and normal measurement data.
[0067] During training, the input samples first pass through an input mapping layer, mapping the original multidimensional measurement features to a latent space suitable for deep feature extraction. Then, the Mamba temporal feature extraction module learns the dynamic evolution of distribution network measurement data within historical time windows, capturing the temporal deviation features between attacked and normal measurements. Next, a graph convolutional network is used to model the spatial coupling relationship between nodes based on the distribution network topology, enabling the association information between attacked nodes and their adjacent nodes and related branches to participate in the recovery decision. Afterward, the key features at different time steps are weighted and fused by a temporal attention aggregation layer, and finally, the recovery results of measurement data such as node voltage, node injected power, and related branch power flow are output by a multi-path recovery output head.
[0068] Regarding the loss function, the total loss during the training phase consists of two parts: a data fitting term and a physical constraint term.
[0069] ;
[0070] in, , , , , , , The data fitting losses for four types of power flow measurements—node voltage, node injected active power, node injected reactive power, and branch head-to-end flow—are all performed using the Smooth L1 loss function. The threshold parameter β... huber =1, the absolute value of the error is less than β. huber When calculated in L2 form, it is greater than β. huber The calculation is performed in L1 form, which maintains high sensitivity to small errors while suppressing the gradient dominance effect of large error samples, thereby improving training stability. The auxiliary supervision loss for voltage residual prediction is weighted by α, and is applied to the residual prediction output and the true residual in the normalized space. Applying Smooth L1 loss to the deviation between them enhances the model's accuracy in estimating the direction and magnitude of the attack deviation; The supervision loss of the confidence output head is β, and the mean squared error is used to measure the difference between the confidence output and the supervision target mentioned above. The physical constraint loss is represented by the weighting coefficient λ.
[0071] Physical constraints Based on the system node admittance matrix, linearized AC power flow equations are constructed. The AC power flow equations are:
[0072] ;
[0073] The power flow equations at the beginning and end of branch l are:
[0074] ;
[0075] Since the phase angle difference between nodes is generally small during normal operation of the power grid, a zero-phase-angle approximation is used to constrain the injected power at the nodes, i.e. The branch power flow constraint adopts a small angle approximation, and the specific constraint formula is shown below:
[0076] ;
[0077] in, , These are the conductance and susceptance components in the nodal admittance matrix. , Here are the branch admittance parameters. The total physical constraint loss is a weighted sum of the node injection constraint loss and the branch power flow constraint loss:
[0078] ;
[0079] in, , Inject the recovery value directly predicted by the power output head into the node; , The value recovered from the node voltage Substituting the theoretical injection estimate into the constraint formula, , , , The predicted recovery value for the branch power flow output head. , , , The value recovered from the node voltage , The theoretical head-to-tail branch power flow estimates are obtained by substituting the values into the constraint formula. N represents the number of nodes, and L represents the number of branches.
[0080] Subsequently, the Adam optimizer combined with the backpropagation algorithm was used to iteratively update the parameters of the ST-MambaGCN model, jointly optimizing the weight parameters and bias terms in the input mapping layer, Mamba temporal module, graph convolutional layer, temporal attention aggregation layer, and multi-path recovery output header. During training, the model recovery performance in each round was evaluated using the validation set, and adaptive learning rate adjustment and early stopping mechanisms were combined to improve training stability and prevent the model from overfitting to the training set. Finally, the model with the smallest recovery error on the validation set and the best performance in satisfying physical constraints was saved as the inference model for online recovery of distribution network measurement data after FDIA attacks.
[0081] Step 4: Evaluation and Verification of Detection and Location Results
[0082] After the model was trained, the performance of the ST-MambaGCN data recovery model was evaluated on the test set. To effectively measure the model's data recovery capability, the following four evaluation metrics were used to assess the recovery effect.
[0083] Mean Absolute Error (MAE) measures the average absolute magnitude of the deviation between the recovered value and the true value, and intuitively reflects the overall recovery accuracy of the model.
[0084] ;
[0085] in, The set of measurements being attacked. Its sample size, The recovery value of the i-th measurement. This corresponds to the actual measured value.
[0086] The root mean square error (RMSE) imposes a higher penalty on larger errors, effectively characterizing the worst-case performance of the model on samples with large recovery errors.
[0087] ;
[0088] Coefficient of Determination (R²) 2 This measures the degree of fit between the recovered value and the true value's variation pattern, with a value range of [value range missing]. The closer the value is to 1, the higher the linear agreement between the recovery result and the actual measurement.
[0089] ;
[0090] in, This is the mean of the true values measured under attack.
[0091] KCL power residuals are used to assess the physical justification of the restored measurement. The formula for calculating KCL power balance residuals is the same as the physical constraint term in step three. Consistent. In the formula, the weights are... , The smaller the KCL power balance residual, the more the recovered measurement conforms to the physical operation of the distribution network. The difference between the two is that during training, the KCL power balance residual participates in gradient backpropagation as part of the loss, while during evaluation, it is used as an independent indicator to quantify the physical consistency of the recovered measurement.
[0092] Of the four evaluation metrics above, MAE and RMSE assess recovery accuracy from the perspective of error magnitude, while R... 2 The KCL power balance residual reflects the recovery quality from the perspective of fit consistency, and verifies the engineering applicability of the recovery results from the perspective of physical feasibility. The four types of indicators complement each other and together constitute the evaluation system for the recovery performance of the ST-MambaGCN model.
[0093] The following specific examples illustrate the effectiveness of this invention. Using hourly load data from the US power operator PJM over five years (2020-2024), 100,000 time-series load scenarios were generated through interpolation. FDIA attack scenario quantity measurement datasets with different attack intensities (c=0.01 / 0.015 / 0.02 / 0.025 / 0.03 / 0.05), attack ratios (15% / 30%), and the number of attacking nodes (1-5) were generated on IEEE 33-node and IEEE 69-node systems. An ST-MambaGCN model was constructed, the structure of which is as follows... Figure 2 As shown.
[0094] Figure 3 (a), (b), and (c) represent the node measurement recovery results for voltage amplitude, active power, and reactive power in the IEEE 33-node system, respectively. Figure 4 Verification of KCL power balance residuals in the IEEE 33-bus system. Figure 5 To assess the model recovery performance under different numbers of attacking nodes in the IEEE 33-node system, the ST-MambaGCN model was used to recover data from the FDIA-affected dataset. The data recovery results for the IEEE 33-node system are shown in Tables 1, 2, and 4. Figures 3-5 As shown in Table 3, the data recovery results under the IEEE 69-node system are presented. Analysis reveals that the ST-MambaGCN recovery framework exhibits stable recovery capabilities and physical reliability, verifying the broad applicability of ST-MambaGCN in various FDIA attack scenarios.
[0095] Table 1 Overall Evaluation Indicators of Model Recovery Performance in the IEEE 33-Node System
[0096]
[0097] Table 2. Model recovery performance under different attack intensities in the IEEE 33-node system.
[0098]
[0099] Table 3. Model recovery performance under different attack intensities in the IEEE 69-node system
[0100]
[0101] Table 4. Model recovery performance under different attack ratios in the IEEE 33-node system
[0102]
[0103] at last, Figure 6 Performance analysis of ST-MambaGCN and benchmark models. Performance metrics of ST-MambaGCN are compared with GAN and VAE on an IEEE 69-node system. Experimental results show that ST-MambaGCN outperforms GAN and VAE in MAE, RMSE, and R... 2 The model achieved optimal results across all evaluation metrics, and its overall recovery performance was superior to the comparative model.
[0104] In summary, the method of this invention can accurately recover and reconstruct the measurement data of attacked nodes and their related branches, and can achieve collaborative recovery of multiple types of measurement data. It overcomes the shortcomings of existing methods that only focus on node measurements or model single spatiotemporal features, improving the physical consistency and engineering usability of the recovery results. Furthermore, it can simultaneously consider temporal dynamic characteristics, spatial topological correlation, and physical constraint consistency, effectively reducing the continued impact of FDIA on subsequent online power flow calculations, economic dispatching, and security analysis, providing reliable support for the safe recovery and stable operation of the distribution network cyber-physical system after being subjected to FDIA.
Claims
1. A power distribution network data recovery method against false data injection attack, characterized in that, include: Collect FDIA detection datasets from the power distribution network and preprocess the datasets; A spatiotemporal Mamba graph convolutional network is constructed, consisting of an input mapping layer, a temporal feature extraction module, a spatial feature extraction module, a temporal attention aggregation layer, and a multi-path recovery output head connected in series. The spatiotemporal Mamba graph convolutional network is trained by first passing the input sample through an input mapping layer to map the original multidimensional measurement features to a latent space suitable for deep feature extraction. Then, the temporal feature extraction module learns the dynamic evolution of distribution network measurement data within historical time windows, capturing the temporal deviation features between attacked and normal measurements. Finally, the spatial feature extraction module models the spatial coupling relationship between nodes based on the distribution network topology, enabling the association information between attacked nodes and their adjacent nodes and related branches to participate in recovery decisions. Then, the key features at different time steps are weighted and fused by the temporal attention aggregation layer, and finally the data recovery results are output by the multi-path recovery output head. The data recovery performance of spatiotemporal Mamba graph convolutional networks is evaluated.
2. The power distribution network data recovery method against false data injection attack according to claim 1, characterized in that, Collecting FDIA inspection datasets from the power distribution network and preprocessing the datasets specifically includes: Based on the distribution network topology and power operator load data, time-series load scenarios are generated, and historical measurement data of distribution network operation are collected. Construct FDIA attack vectors and superimpose random offsets that follow a uniform distribution onto the measurement data of the corresponding nodes. Design FDIA attack scenarios with different attack intensities, attack ratios, and number of attack nodes. Measurement data including normal and attack scenarios are labeled and divided into training, validation and test sets according to proportions; Historical measurement data are standardized to eliminate dimensional differences, and the original time-series measurement data are transformed into fixed-length window samples using a sliding time window method.
3. The power distribution network data recovery method against false data injection attack according to claim 2, characterized in that, The historical measurement data includes three types of node measurements: voltage amplitude at each node, active power injected into the node, and reactive power; and four types of branch measurements: active and reactive power flow at the beginning of the line, and active and reactive power flow at the end of the line.
4. The power distribution network data recovery method against false data injection attack according to claim 3, characterized in that, The input mapping layer consists of a linear transformation, layer normalization, and a ReLU activation function concatenated. For each scenario to be recovered, based on the current time t, the historical node measurement sequences of the previous W time steps and the current time measurement are extracted and concatenated in the time dimension to form a complete time series input of length W+1. The historical node measurement sequences provide the model with a time series prior reference, and the current time measurement contains the attack deviation information to be recovered. The node feature dimension of each time step in the concatenated time series input is D0, corresponding to three types of measurements: node voltage amplitude, node injected active power, and reactive power.
5. The power distribution network data recovery method against false data injection attack according to claim 3 or 4, characterized in that, The temporal feature extraction module consists of several Mamba blocks connected in series. In each Mamba block, the input features are normalized at the first layer and then divided into a signal branch and a gated branch by gated projection. The signal branch is processed by SSM and then multiplied element-wise with the soft gated vector obtained by sigmoid activation of the gated branch. After output projection and second-layer normalization, the output is connected through residuals.
6. The power distribution network data recovery method against false data injection attack according to claim 5, characterized in that, The feature matrix of the (l+1)th layer node in the spatial feature extraction module is represented as follows: ; wherein, is the l-th layer node feature matrix, is the learnable weight, is the ReLU activation function, is the symmetric normalized node relation adjacency matrix.
7. The distribution network data recovery method for false data injection attacks according to claim 6, characterized in that, The attention weights of the temporal attention aggregation layer are generated by a linear layer with Softmax normalization, and the features at different time instants are weighted and summed after being given different weights according to their relevance to the current recovery task, compressing the feature tensor with a shape of into a node-level feature representation with a shape of .
8. The distribution network data recovery method for false data injection attacks according to claim 7, characterized in that, Each output head of the multi-path recovery output head is composed of a linear layer, a layer normalization layer, a ReLU activation layer, and an output linear layer connected in series.
9. The distribution network data recovery method for false data injection attacks according to claim 2, characterized in that, Training a spatiotemporal Mamba graph convolutional network also includes: The Adam optimizer combined with the backpropagation algorithm is used to iteratively update the parameters of the spatiotemporal Mamba graph convolutional network. The model recovery performance in each round is evaluated using the validation set. The training stability is improved by combining adaptive learning rate adjustment and early stopping mechanism. The model with the smallest recovery error in the validation set and the best performance in satisfying physical constraints is saved as the inference model for online recovery of distribution network measurement data after FDIA attack.
10. The distribution network data recovery method for false data injection attacks according to claim 9, characterized in that, The loss function for training a spatiotemporal Mamba graph convolutional network is: ; ; in, , , , , , , The data fitting losses for four types of power flow measurements—node voltage, node injected active power, node injected reactive power, and branch head-to-end flow—are all implemented using the Smooth L1 loss function; the threshold parameter β... huber =1, the absolute value of the error is less than β. huber When calculated in L2 form, it is greater than β. huber Calculated in L1 form. The auxiliary supervision loss for voltage residual prediction is α, with a weighting coefficient of α. The supervision loss for the confidence output head is β, with a weighting coefficient of β. The physical constraint loss is represented by the weighting coefficient λ. , Inject the recovery value directly predicted by the power output head into the node; , The value recovered from the node voltage The calculated theoretical injection estimate, , , , The predicted recovery value for the branch power flow output head. , , , The value recovered from the node voltage , The calculated theoretical power flow estimates for the first and last branches; N is the number of nodes, and L is the number of branches; , For weights.
