Power supply station area line loss anomaly analysis method based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent power grids, and discloses a power supply station area line loss anomaly analysis method based on artificial intelligence, which comprises the following steps: step 1, acquiring power input power, output power, power grid topological information and environmental factor data of a power supply station area, and performing missing value filling, abnormal value elimination and normalization processing on the acquired data to obtain a power supply station area line loss anomaly analysis result; obtaining normalized power data and topological data; 2, a power grid graph structure is constructed based on the normalized power grid topological data, a variational adjacency matrix is established through a variational inference method, the variational adjacency matrix is used for representing the power grid connection relation between station intervals, modeling is conducted on the uncertainty of power grid topology, and power grid topological data after the variational adjacency matrix is obtained. A dynamic adjacency matrix is constructed through a variational inference method, self-adaptive modeling of dynamic changes of a power grid topological structure is achieved, and the anomaly detection capability which can still keep high accuracy in a topological change environment is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grids, and specifically to a method for analyzing abnormal line losses in power supply station areas based on artificial intelligence. Background Art

[0002] The analysis of abnormal line losses in power supply station areas is a key issue in smart grid management. Abnormal line losses are caused by factors such as power theft, equipment aging, and line faults. If not detected and effective measures are not taken in time, it will cause waste of power resources and affect the stable operation of the power grid. Currently, the analysis of abnormal line losses in station areas mainly relies on methods based on statistics and traditional machine learning, which have limitations in the face of complex power grid environments.

[0003] Traditional power grid line loss analysis methods usually rely on static power grid topology modeling and cannot handle changes in topology structure that occur during the operation of the power grid.

[0004] Existing graph neural network models mainly rely on fixed adjacency matrices for calculation and cannot adapt to the dynamically changing power grid structure, resulting in a decline in detection performance.

[0005] The load data of power supply areas has non-linear characteristics and is affected by factors such as temperature, humidity, and user behavior. Traditional dimensionality reduction methods cannot effectively retain the non-linear characteristics of the data.

[0006] Existing methods will lose the local structure information of the data during dimensionality reduction, resulting in a decline in the accuracy of anomaly detection. In the case of high complexity of load data, false alarms and missed alarms are likely to occur.

[0007] Most existing anomaly detection methods are based on fixed thresholds for classification, and the detection strategy cannot adapt to the load patterns of different station areas and different time periods, resulting in high false alarm rates and missed alarm rates.

[0008] Power theft users adopt adversarial means to avoid detection, making abnormal behaviors concealed, and the method of fixed thresholds cannot cope with such strategies.

[0009] Traditional alarm triggering mechanisms usually use fixed thresholds. When the load pattern changes, false alarms and missed alarms are likely to occur, affecting the operation and maintenance efficiency.

[0010] Existing anomaly scoring calculation methods lack an adaptive adjustment mechanism and cannot dynamically optimize the detection threshold according to historical data, resulting in the detection system being unable to maintain high-efficiency operation for a long time.

[0011] In response to the above problems, existing research has tried to introduce deep learning methods, but the methods mainly focus on time series prediction and do not fully utilize the topological structure information of the power grid, and there are still high false alarm rates and missed alarm rates. In addition, some research uses reinforcement learning to optimize the detection strategy, but there are still challenges in terms of computational complexity and convergence speed.

[0012] Therefore, those skilled in the art provide a method for analyzing abnormal line losses in power supply station areas based on artificial intelligence to solve the above-mentioned problems. Summary of the Invention

[0013] In view of the deficiencies of the prior art, the present invention provides a method for analyzing abnormal line losses in power supply station areas based on artificial intelligence to solve the problems raised in the above background technology.

[0014] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for analyzing abnormal line losses in power supply station areas based on artificial intelligence includes: Step 1: Obtain the power input power, output power, grid topology information, and environmental factor data of the power supply station area, and perform missing value filling, outlier removal, and normalization processing on the collected data to obtain normalized power data and topology data; Step 2: Based on the normalized grid topology data, construct a grid graph structure, and use the variational inference method to establish a variational adjacency matrix. The above variational adjacency matrix is used to represent the grid connection relationship between areas, and model the uncertainty of the grid topology to obtain the grid topology data after the variational adjacency matrix; Step 3: Input the normalized power data and the grid topology data after the variational adjacency matrix into a graph neural network for feature calculation, use the graph neural network propagation mechanism based on the variational adjacency matrix to calculate the hidden features of each area, and use an activation function to perform a non-linear transformation on the hidden features to obtain the feature representation of each area; Step 4: Input the feature representation of each area into a manifold learning model, calculate the normalized Laplacian matrix by constructing an adjacency weight matrix, and perform eigenvalue decomposition on the normalized Laplacian matrix to extract the first several eigenvectors as the reduced-dimensional feature representation of the area; Step 5: Based on the reduced-dimensional feature representation of the area and the feature representation of the area calculated by the graph neural network, calculate the degree of difference, and classify the degree of difference according to a set threshold to obtain the abnormal score of each area; Step 6: Based on the abnormal scores of each area, establish a game model between the power supply side and the user, define the detection strategy of the power supply side and the power consumption strategy of the user, construct a loss function, and use the gradient descent method to solve the optimal detection strategy and update the classification threshold of the abnormal score; Step 7: Based on the updated classification threshold of the abnormal score, determine whether the area is abnormal, send an abnormal alarm message to the power grid dispatching system, and at the same time dynamically adjust the detection strategy based on historical alarm data to optimize the balance between the false alarm rate and the missed alarm rate.

[0015] Preferably, the power input power and output power obtained in Step 1 are denoted as and , where the line loss ratio is calculated by the formula: , where represents the input power of the kth transformer substation at time t; represents the output power of the kth transformer substation at time t; represents the line loss ratio of the kth transformer substation at time t.

[0016] Preferably, the variational adjacency matrix established in step 2 is calculated by the following variational inference method: , where represents the variational adjacency mean between transformer substations i and j, represents the variational adjacency variance between transformer substations i and j, represents that the probability distribution of the adjacency relationship follows a normal distribution with a mean of and a variance of .

[0017] Preferably, the graph neural network propagation mechanism in step 3 adopts the following update rule: , where represents the hidden feature of transformer substation p at the th layer, is the adjacency relationship weight in the variational adjacency matrix, is the weight matrix of the tth layer, is the bias term of the tth layer, represents the hidden feature of transformer substation q at the tth layer, is the non-linear activation function, is the set of adjacent transformer substations connected to transformer substation p.

[0018] Preferably, the manifold learning dimensionality reduction method in step 4 adopts Laplacian eigenmaps, and the objective function is as follows: , where and are the dimensionality-reduced feature representations of the transformer substations, is the adjacent weight: , where is the parameter for controlling the local neighborhood range, It is the high-dimensional feature representation obtained by the graph neural network of the transformer substation area r. It is the high-dimensional feature representation obtained by the graph neural network of the transformer substation area. It represents the Euclidean distance of features between the transformer substation area r and the transformer substation area s.

[0019] Preferably, the abnormal score calculation formula in step 5 is: , where, represents the abnormal score of the u-th transformer substation area, is the reduced-dimensional feature representation of the transformer substation area, is the feature representation of the transformer substation area calculated by the graph neural network; If exceeds the set threshold θ, it is determined that the transformer substation area u is abnormal, and the abnormal score is used for the optimization of the detection strategy in step 6.

[0020] Preferably, the optimization of the detection strategy in step 6 is carried out by using the differential game method, and the loss function of the power supply side is defined as: , The loss function of the user is defined as: , where, is the detection strategy of the power supply side, is the power consumption strategy of the user, is the state variable of the transformer substation area, T is the optimization time interval, is the loss function of the power supply side, is the loss function of the user, and are the instantaneous loss functions of the power supply side and the user.

[0021] Preferably, the abnormal alarm information in step 7 is triggered according to the abnormal score , and the alarm is triggered when the abnormal score meets the following conditions: , where, is the abnormal score of the transformer substation area u, is the classification threshold, is the alarm trigger adjustment coefficient; If the alarm is triggered, an alarm message is sent to the power grid dispatching system, and the alarm data is stored for historical data analysis and the optimization of the detection strategy in step 6.

[0022] Preferably, the dynamic adjustment of the detection strategy in step 7 adopts the following optimization rules: , where, is the classification threshold optimized in the t-th round, the classification threshold optimized in the round of is the learning rate, is the gradient of the power supply party's loss function with respect to the classification threshold.

[0023] Preferably, the historical alarm data analysis in step 7 adopts a time window method to calculate the average anomaly score within the historical window : , where is the average anomaly score within window W, is the window length, is the anomaly score at the w-th time point within the window.

[0024] The present invention provides an artificial intelligence-based method for analyzing abnormal line losses in power supply substations. It has the following beneficial effects: 1. The present invention constructs a dynamic adjacency matrix through variational inference method to achieve adaptive modeling of the dynamic changes in the power grid topology structure, and obtains an anomaly detection ability that can still maintain high accuracy in a topological change environment.

[0025] 2. The present invention performs manifold learning dimensionality reduction through Laplacian eigenmaps to achieve feature extraction of non-linear load data in the power supply substation area, obtains a low-dimensional feature representation that retains local geometric information, and improves the accuracy of anomaly detection.

[0026] 3. The present invention constructs a differential game model between the power supply party and users to achieve dynamic optimization and adjustment of the detection strategy, obtains an intelligent anomaly detection mechanism that takes into account the balance between false alarm rate and missed alarm rate, and improves the reliability and practicality of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is the flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0028] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0029] The present invention will be described in detail below with reference to the drawings: Embodiment

[0030] Please refer to the attachedFigure 1 , an embodiment of the present invention provides an abnormal analysis method for the line loss of a power supply station area based on artificial intelligence, including: Step 1: Obtain the power input power, output power, grid topology information, and environmental factor data of the power supply station area, and perform missing value filling, outlier removal, and normalization processing on the collected data to obtain normalized power data and topology data; Step 2: Based on the normalized grid topology data, construct a grid graph structure, and use the variational inference method to establish a variational adjacency matrix. The above variational adjacency matrix is used to represent the grid connection relationship between regions, and model the uncertainty of the grid topology to obtain the grid topology data after the variational adjacency matrix; Step 3: Input the normalized power data and the grid topology data after the variational adjacency matrix into a graph neural network for feature calculation, use the graph neural network propagation mechanism based on the variational adjacency matrix to calculate the hidden features of each region, and perform a non-linear transformation on the hidden features using an activation function to obtain the feature representation of each region; Step 4: Input the feature representation of each region into a manifold learning model, calculate the normalized Laplacian matrix by constructing an adjacency weight matrix, and perform eigenvalue decomposition on the normalized Laplacian matrix to extract the first several eigenvectors as the feature representation of the region after dimensionality reduction; Step 5: Calculate the difference degree based on the feature representation of the region after dimensionality reduction and the feature representation of the region calculated by the graph neural network, and classify the difference degree according to a set threshold to obtain the abnormal score of each region; Step 6: Based on the abnormal scores of each region, establish a game model between the power supply side and the user, define the detection strategy of the power supply side and the power consumption strategy of the user, construct a loss function, and use the gradient descent method to solve the optimal detection strategy and update the classification threshold of the abnormal score; Step 7: Based on the updated classification threshold of the abnormal score, determine whether the region is abnormal, and send an abnormal alarm message to the power grid dispatching system. At the same time, dynamically adjust the detection strategy based on historical alarm data to optimize the balance between the false alarm rate and the missed alarm rate.

[0031] By collecting power input power, output power, grid topology information, and environmental factor data, comprehensively obtain the key factors affecting the line loss of the region, and ensure the integrity of the data.

[0032] Performing missing value filling, outlier removal, and normalization processing can improve the data quality, reduce noise interference, and provide high-quality input data for subsequent analysis.

[0033] Using the variational inference method to establish a variational adjacency matrix can dynamically represent the grid connection relationship between regions, enable the model to adapt to changes in the grid topology, and improve the robustness of detection.

[0034] By modeling the uncertainty of the power grid topology, the adaptability to complex power grid structures is enhanced, and the accuracy of anomaly detection is improved.

[0035] Combined with the variational adjacency matrix for graph neural network calculation, it can utilize the power grid topology information and improve the representation ability of substation area features.

[0036] Using an activation function for non-linear transformation enhances the model's ability to express complex non-linear relationships and improves the detection accuracy.

[0037] Using Laplacian eigenmaps for dimensionality reduction can retain the local structure information of substation area data and improve the interpretability of features.

[0038] By constructing an adjacency weight matrix, the topological relationship between data is maintained, ensuring that the data after dimensionality reduction still has the ability to distinguish anomalies.

[0039] By comparing the substation area feature representation after dimensionality reduction and the substation area feature representation calculated by the graph neural network, the anomaly degree of the substation area is effectively measured, and the accuracy of anomaly detection is improved.

[0040] Using the calculation of the degree of difference combined with the set threshold classification can ensure the reliability of the anomaly score and provide a basis for optimizing subsequent detection strategies.

[0041] By establishing a game model between the power supply side and users, dynamic optimization can be carried out according to the strategies of both parties, improving the intelligence of the detection strategy.

[0042] Using the gradient descent method to solve the optimal detection strategy achieves a good balance between the false alarm rate and the miss rate, improving the detection effect.

[0043] By updating the anomaly score classification threshold, the anomaly detection standard can be dynamically adjusted according to real-time data, improving the adaptability of the detection.

[0044] Combined with the analysis of historical alarm data, the detection strategy is optimized to reduce false alarms and missed alarms, enhancing the long-term stability of the detection system.

[0045] The power input and output powers obtained in Step 1 are denoted as and , where the line loss ratio The calculation formula is: , where represents the input power of the kth substation area at time t; represents the output power of the kth substation area at time t; represents the line loss ratio of the kth substation area at time t.

[0046] By calculating the line loss ratio , this method can accurately measure the power loss of the substation area, providing reliable data support for anomaly detection. This calculation method has normalization processing to improve the calculation stability of the model, is applicable to the dynamic power grid environment, lays a foundation for the subsequent graph neural network feature calculation and anomaly score calculation, and further improves the intelligent level and adaptability of the overall detection system.

[0047] The variational adjacency matrix established in step 2 is calculated through the following variational inference method: , where represents the variational adjacency mean between substation area i and substation area j, represents the variational adjacency variance between substation area i and substation area j, represents that the probability distribution of the adjacency relationship follows a normal distribution with a mean of and a variance of .

[0048] The traditional adjacency matrix adopts a fixed structure and cannot adapt to the changes in the power grid topology. The present invention uses the variational inference method to calculate the variational adjacency matrix , and models the connection relationship between substation areas through probability distribution, making the power grid topology modeling have dynamic adaptability.

[0049] The variational adjacency mean represents the average connection strength between substation area i and substation area j, which can reflect the long-term power transmission mode; the variational adjacency variance characterizes the uncertainty of the topological structure, enabling the model to tolerate short-term fluctuations in the connection relationship between substation areas and improving the stability of the detection system.

[0050] The traditional adjacency matrix is usually constructed based on the physical topology of the power grid, while this method learns the probability distribution of the adjacency relationship through variational inference, enabling the adjacency matrix to be adaptively adjusted, accurately representing the actual power transmission mode between substation areas, and reducing the modeling error.

[0051] Modeling the adjacency relationship through the normal distribution effectively avoids the influence of abnormal connections on the topological structure, makes the substation area feature calculation more robust, and improves the accuracy of calculating the substation area features by the graph neural network.

[0052] In the actual operation of the power grid, due to load fluctuations, topological changes, and data noise, the adjacency relationship will change in the short term. Using a fixed adjacency matrix will lead to misjudgment of the detection system. This method models the uncertainty of the adjacency relationship through the variational distribution , enabling the system to maintain stable detection ability in a complex environment.

[0053] The graph neural network propagation mechanism in step 3 adopts the following update rules: , where, represents the hidden feature of substation area p at the -th layer, is the adjacency relation weight in the variational adjacency matrix, is the weight matrix of the t-th layer, is the bias term of the t-th layer, represents the hidden feature of substation area q at the t-th layer, is the non-linear activation function, is the set of adjacent substation areas connected to substation area p.

[0054] The present invention adopts a graph neural network propagation mechanism based on a variational adjacency matrix, improves the utilization rate of topological information, and enhances the expression ability of substation area features. By introducing a non-linear activation function and a trainable weight matrix, the adaptive ability of the model is improved, and abnormal substation areas can be accurately identified. At the same time, multi-layer feature propagation is adopted to reduce the influence of single substation area noise on the detection result, improve the detection stability and robustness of the system, and provide efficient and reliable technical support for the abnormal detection of smart grids.

[0055] The manifold learning dimensionality reduction method in step 4 adopts Laplacian eigenmaps, and the objective function is as follows: , where, and are the dimensionality-reduced substation area feature representations, is the adjacency weight: , where, is the parameter for controlling the local neighborhood range, is the high-dimensional feature representation calculated by the graph neural network of substation area r, is the high-dimensional feature representation calculated by the graph neural network of substation area s, represents the Euclidean distance of features between substation area r and substation area s.

[0056] Traditional dimensionality reduction methods are mainly based on global linear transformation and easily ignore the local geometric information of data. This method adopts Laplacian eigenmaps and maintains the local similarity of substation area features through the adjacency weight matrix so that the dimensionality-reduced features can still reflect the power relationship between substation areas.

[0057] This method can maintain the coherence of the topological structure, make the features of abnormal substation areas prominent, and improve the accuracy of abnormal detection.

[0058] Use a graph neural network to calculate the high-dimensional substation area features, and construct the adjacency weights based on their Euclidean distances , which can ensure that the graph structure information can still be retained after dimensionality reduction, and improve the feature representation ability of the substation area after dimensionality reduction.

[0059] The abnormal score calculation formula in step 5 is as follows: , where represents the abnormal score of the u-th substation area, is the feature representation of the substation area after dimensionality reduction, is the feature representation of the substation area calculated by the graph neural network; If exceeds the set threshold θ, it is determined that the substation area u is abnormal, and the abnormal score is used to optimize the detection strategy in step 6.

[0060] Traditional adjacency matrices usually rely on static structures and cannot cope with changes in the power grid topology. The variational adjacency matrix constructed by the variational inference method can reflect the dynamic changes in the power grid connection relationship through probability distributions, enabling the power grid topology to be adaptively adjusted over time. The dynamic adaptability ensures that the detection system can maintain high accuracy and robustness when the power grid topology changes.

[0061] With the variational adjacency matrix, instead of simply using fixed adjacency relationships, the adjacency mean and variance calculated through variational inference can describe the changes in the power grid topology structure. This enables the model to perform effective anomaly detection under different topological states, and through uncertainty modeling, improves the adaptability to the complexity of the power grid environment.

[0062] The variational adjacency mean and variational adjacency variance of the variational adjacency matrix can respectively characterize the stable connection strength between substations and the uncertainty of the connection relationship. The introduction of uncertainty modeling helps to accurately judge abnormal states, especially when facing complex power grid structure changes and changes in the relationship between substations, and can flexibly process data. As a result, the accuracy and stability of the detection model are improved.

[0063] The detection strategy optimization in step 6 is optimized using differential game methods. The loss function of the power supply side is defined as: , The loss function of the user is defined as: , where is the detection strategy of the power supply side, is the electricity consumption strategy of the user, is the substation area state variable, T is the optimization time interval, is the loss function of the power supply side, is the loss function of the user, and are the instantaneous loss functions of the power supply side and the user, respectively.

[0064] Through the propagation mechanism of the graph neural network, this method can make full use of the dynamic topological information provided by the variational adjacency matrix to improve the anomaly detection ability under the change of power grid topology. The non-linear activation function enhances the feature expression ability of the model, while the propagation mechanism of the adjacent substation area set optimizes the information flow to improve the accuracy and sensitivity of the overall anomaly detection. The design of multi-level feature extraction enhances the robustness and adaptability of the model, providing strong support for the efficient monitoring of smart grids.

[0065] The anomaly warning information in step 7 is triggered according to the anomaly score and triggers an alarm when the anomaly score meets the following conditions: , where is the anomaly score of substation area u, is the classification threshold, is the alarm trigger adjustment coefficient; If an alarm is triggered, an alarm message is sent to the power grid dispatching system, and the alarm data is stored for historical data analysis and the detection strategy optimization in step 6.

[0066] The anomaly warning information trigger mechanism of the present invention dynamically adjusts the threshold based on the anomaly score, which can improve the accuracy of the warning, reduce false alarms and missed alarms. At the same time, a real-time warning mechanism is adopted to ensure a quick response when an anomaly occurs and reduce the power grid risk. In addition, the storage and historical analysis of warning data can optimize the detection strategy, making the power grid anomaly detection system have self-adaptability, stability and intelligence level, providing technical guarantee for the efficient management of smart grids.

[0067] The dynamic adjustment of the detection strategy in step 7 adopts the following optimization rules: , where is the classification threshold optimized in the t-th round, is the classification threshold optimized in the -th round, is the learning rate, is the gradient of the power supply side loss function with respect to the classification threshold.

[0068] This method uses gradient descent to optimize the classification threshold, enabling the detection strategy to dynamically adapt to the changes in the power grid load, improving the sensitivity and accuracy of anomaly detection. By optimizing the loss function of the power supply side, false alarms and missed alarms are reduced, enhancing the stability of the detection strategy. In addition, by combining historical alarm data to optimize the classification threshold, the detection strategy is made scientific and reasonable, improving the detection effect in the long term. At the same time, this method reduces manual intervention, improves the intelligent level of detection, and enhances the safety and reliability of the power grid.

[0069] The historical alarm data analysis in step 7 uses the time window method to calculate the average anomaly score within the historical window : , where is the average anomaly score within window W, is the window length, and is the anomaly score at the w-th time point within the window.

[0070] By calculating the difference between the reduced-dimensional substation area feature representation and the substation area feature representation calculated by the graph neural network, the anomaly degree of the substation area is effectively quantified. The difference calculation provides a quantification method that can accurately reflect whether there is an anomaly in the substation area, providing a basis for subsequent anomaly scoring.

[0071] This calculation method ensures that the anomaly degree of abnormal substations can be accurately captured, improving the accuracy of the detection results.

[0072] Using the set threshold to classify the difference degree, determine the anomaly score of each substation area, and classify the substation area as normal and abnormal. This process is flexible and can dynamically adjust the threshold according to the real-time data of the power grid, ensuring the adaptability of the model to different power grid environments.

[0073] This method ensures that substations with different anomaly degrees can be accurately distinguished, avoiding false alarms and missed alarms, and improving the effectiveness of the anomaly detection system.

[0074] Through the method based on difference degree calculation and threshold classification, the system can automatically adapt to different operating states of the power grid. The difference degree combines the substation area features after dimensionality reduction and the features calculated by the graph neural network, making the anomaly score reflect local anomalies and being able to identify complex anomaly patterns in the global power grid.

[0075] The self-adaptability of this method provides real-time detection capabilities for the dynamic changes in the power grid operation, enhancing the reliability of the system in complex environments.

[0076] By inputting the anomaly scores of each substation area into the subsequent detection strategy optimization step, the dispatching strategy and detection threshold of the power grid are optimized according to the score values. It provides strong decision-making support between the power supply side and users, ensuring the timely identification and handling of abnormal power transmission.

[0077] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing abnormal line loss in power supply substations based on artificial intelligence, characterized in that: include: Step 1: Obtain the power input power, output power, grid topology information and environmental factor data of the power supply substation, and fill in missing values, remove outliers and normalize the collected data to obtain normalized power data and topology data; Step 2: construct a power grid graph structure based on the normalized power grid topology data, and use a variational inference method to establish a variational adjacency matrix. The variational adjacency matrix is ​​used to represent the power grid connection relationship between substations, and the uncertainty of the power grid topology is modeled to obtain the power grid topology data after the variational adjacency matrix. Step 3: Input the normalized power data and the grid topology data after the variational adjacency matrix into the graph neural network for feature calculation. The graph neural network propagation mechanism based on the variational adjacency matrix is ​​used to calculate the hidden features of each substation, and the activation function is used to perform nonlinear transformation on the hidden features to obtain the feature representation of each substation. Step 4: Input the feature representation of each area into the manifold learning model, calculate the normalized Laplace matrix by constructing an adjacency weight matrix, perform eigenvalue decomposition on the normalized Laplace matrix, and extract the first several eigenvectors as the feature representation of the area after dimensionality reduction; Step 5: Calculate the difference based on the feature representation of the area after dimensionality reduction and the feature representation of the area calculated by the graph neural network, and classify the difference according to the set threshold to obtain the abnormality score of each area; Step 6: Based on the anomaly score of each substation, a game model between the power supplier and the user is established, the detection strategy of the power supplier and the power consumption strategy of the user are defined, a loss function is constructed, the gradient descent method is used to solve the optimal detection strategy, and the classification threshold of the anomaly score is updated; Step 7: Based on the updated abnormal score classification threshold, determine whether the substation is abnormal, and send abnormal alarm information to the power grid dispatching system. At the same time, dynamically adjust the detection strategy based on historical alarm data to optimize the balance between false alarm rate and missed alarm rate.

2. According to the artificial intelligence-based power supply substation line loss anomaly analysis method of claim 1, it is characterized in that: The electric input power and output power obtained in step 1 are recorded as and , where line loss ratio The calculation formula is: ， in, represents the input power of the kth station at time t; represents the output power of the kth station at time t; Represents the line loss ratio of the kth station area at time t.

3. According to the artificial intelligence-based power supply substation line loss anomaly analysis method of claim 1, it is characterized in that: The variational adjacency matrix established in step 2 It is calculated by the following variational inference method: ， in, represents the variational neighboring mean between area i and area j, represents the variational neighbor variance between area i and area j, The probability distribution of the adjacency relationship follows the mean , the variance is The normal distribution of .

4. According to the artificial intelligence-based power supply substation line loss anomaly analysis method of claim 1, it is characterized in that: The graph neural network propagation mechanism in step 3 adopts the following update rules: ， in, Indicates that the station p is Hidden features of the layer, is the adjacency weight in the variational adjacency matrix, is the weight matrix of the tth layer, is the bias term of the tth layer, represents the hidden features of area q at the tth layer, is a nonlinear activation function, is the set of adjacent areas connected to area p.

5. According to the artificial intelligence-based power supply substation line loss anomaly analysis method of claim 1, it is characterized in that: The manifold learning dimensionality reduction method in step 4 adopts Laplace eigenmap, and the objective function is as follows: ， in, and is the feature representation of the area after dimensionality reduction, is the adjacency weight: ， in, To control the parameters of the local neighborhood range, is the high-dimensional feature representation calculated by the graph neural network of area r, It is a high-dimensional feature representation calculated by the graph neural network. Represents the characteristic Euclidean distance between area r and area s.

6. The method for analyzing abnormal line loss in power supply substations based on artificial intelligence according to claim 1 is characterized in that: The calculation formula of the abnormality score in step 5 is: ， in, represents the abnormal score of the u-th station, is the feature representation of the area after dimensionality reduction, The station feature representation calculated for graph neural network; like If it exceeds the set threshold θ, the station u is judged to be abnormal, and the abnormal score is used to optimize the detection strategy in step 6.

7. The method for analyzing abnormal line loss in power supply substations based on artificial intelligence according to claim 6 is characterized in that: The detection strategy optimization in step 6 is optimized using the differential game method, and the loss function of the power supplier is defined as: ， The user's loss function is defined as: ， in, The detection strategy for the power supplier is: For users' electricity usage strategies, is the state variable of the station area, T is the optimization time interval, is the loss function of the power supplier, is the user’s loss function, and is the instantaneous loss function of the power supplier and the user.

8. The method for analyzing abnormal line loss in power supply substation area based on artificial intelligence according to claim 7 is characterized in that: The abnormal alarm information in step 7 is based on the abnormal score If the abnormal score meets the following conditions, an alarm is triggered: ， in, is the abnormal score of area u, is the classification threshold, Adjustment factor for alarm triggering; If the alarm is triggered, the alarm information is sent to the power grid dispatching system, and the alarm data is stored for historical data analysis and detection strategy optimization in step 6.

9. The method for analyzing abnormal line loss in power supply substations based on artificial intelligence according to claim 8 is characterized in that: The dynamic adjustment of the detection strategy in step 7 adopts the following optimization rules: ， in, is the classification threshold after the tth round of optimization, No. The classification threshold after round optimization, is the learning rate, is the gradient of the power supplier loss function with respect to the classification threshold.

10. The method for analyzing abnormal line loss in power supply substation area based on artificial intelligence according to claim 9 is characterized in that: The historical alarm data analysis in step 7 adopts the time window method to calculate the historical window Mean anomaly score within: ， in, is the mean of the anomaly scores within the window W, is the window length, is the abnormality score at the wth time point in the window.

Images