Electricity consumption anomaly detection method based on artificial intelligence

Abstract

The invention relates to the technical field of electricity utilization detection, in particular to an electricity utilization anomaly detection method based on artificial intelligence, which comprises the following steps: acquiring a power distribution network topological structure of a target area, and constructing a graph structure of a power distribution network based on the power distribution network topological structure; performing feature extraction on a node set in the graph structure of the power distribution network to obtain a node feature corresponding to each node in the node set, and weighting the graph structure of the power distribution network based on the node features to obtain a weighted graph structure of the power distribution network; and performing anomaly detection on the weighted graph structure of the power distribution network based on a pre-trained region anomaly detection model. According to the method, the graph structure of the power distribution network is constructed, the node set is subjected to feature extraction, the operation state and network topology information of equipment in the power distribution network can be effectively captured, the model can fully consider different influence factors of each piece of equipment and region through node feature extraction and weighted graph construction, and therefore the anomaly detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power consumption detection, and particularly to a method for detecting abnormal power consumption based on artificial intelligence. Background Art

[0002] Traditional methods rely on simple rules and threshold judgments. These methods usually judge abnormalities based on preset conditions or thresholds. For example, when the current and voltage values exceed the set standards, they will be judged as abnormal. This method ignores the complex relationships between various devices and regions in the distribution network, and is prone to false alarms or missed alarms. Especially when the operating state of the power grid is relatively complex, it is difficult to adapt to the dynamically changing environment; and traditional methods usually cannot be refined to specific sub-regions, so some local abnormal problems may be missed, and only global abnormalities can be identified, resulting in poor diagnostic accuracy; moreover, most traditional methods rely on simple fault speculation or experience-based judgments, often requiring a large amount of manual intervention. This method can often only locate the general area of the abnormality, lacking precise positioning of the specific fault source, which may lead to untimely handling of the problem by the operation and maintenance personnel and delay the repair; and traditional methods still need to rely on the experience and manual judgment of the operation and maintenance personnel. Especially in a complex fault environment, misjudgment or missed judgment may occur due to human errors. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a method for detecting abnormal power consumption based on artificial intelligence.

[0004] The technical solution adopted to solve the above technical problem is: A method for detecting abnormal power consumption based on artificial intelligence, including:

[0005] Obtain the topological structure of the distribution network in the target area, and construct a graph structure of the distribution network based on the topological structure of the distribution network;

[0006] Extract features from the node set in the graph structure of the distribution network to obtain the node features corresponding to each node in the node set, and weight the graph structure of the distribution network based on the node features to obtain a weighted graph structure of the distribution network;

[0007] Perform abnormal detection on the weighted graph structure of the distribution network based on a pre-trained regional abnormal detection model to obtain abnormal detection labels for each sub-region in the target area, where the abnormal detection labels include regional abnormal labels and regional normal labels;

[0008] Locate the root cause of the abnormality for the sub-regions in the target area with abnormal detection labels of regional abnormal labels to obtain the root cause of the abnormality corresponding to the sub-regions.

[0009] Preferably, the graph structure of the distribution network includes a node set and an edge set. Among them, the node set includes an equipment node set and a user node set. The equipment nodes correspond to the power equipment in the distribution network, the user nodes correspond to the electricity users in the distribution network, and the edges correspond to the power lines between the nodes.

[0010] Preferably, feature extraction is performed on the node set in the graph structure of the distribution network to obtain the node features corresponding to each node in the node set, including:

[0011] Feature extraction is performed on the user node set in the graph structure of the distribution network to obtain the user node features corresponding to each user node in the user node set;

[0012] Feature extraction is performed on the equipment node set in the graph structure of the distribution network to obtain the equipment node features corresponding to each equipment node in the equipment node set. Among them, the equipment node features include the voltage level of the power equipment, the load of the power equipment, the current state of the power equipment, and the switch state of the power equipment.

[0013] Preferably, feature extraction is performed on the user node set in the graph structure of the distribution network to obtain the user node features corresponding to each user node in the user node set, including:

[0014] The change coefficient of electricity consumption of each user node in the user node set in the graph structure of the distribution network is extracted to obtain the change coefficient of electricity consumption corresponding to each user node in the user node set. Among them, the expression of the change coefficient of electricity consumption is as follows:

[0015]

[0016] where u i represents the change coefficient of electricity consumption corresponding to the i-th user node in the user node set, Q d represents the electricity consumption data of the user on the d-th day, and x represents the number of days before and after the statistics of the electricity consumption data;

[0017] The decrease coefficient of electricity consumption of each user node in the user node set in the graph structure of the distribution network is extracted to obtain the decrease coefficient of electricity consumption corresponding to each user node in the user node set. Among them, the expression of the decrease coefficient of electricity consumption is as follows:

[0018]

[0019] where M iDenote the power consumption decline coefficient corresponding to the \(i\)-th user node in the user node set. Let \(D(d)\) represent whether there is a power consumption decline on the \(d\)-th day. When the power consumption decline coefficient on the \(d\)-th day is less than that on the \((i - 1)\)-th day, \(D(d)\) takes the value of 1; otherwise, \(D(d)\) takes the value of 0.

[0020] Extract the power consumption difference coefficients between each user node and other user nodes in the user node set of the graph structure of the distribution network to obtain the power consumption difference coefficients between each user node and other user nodes in the user node set. Among them, the expression of the power consumption difference coefficient is as follows:

[0021]

[0022] Among them, \(\alpha\) i \((j, d)\) represents the power consumption difference coefficient between the \(i\)-th user node and the \(j\)-th user node in the user node set. represents the power consumption at the \(h\)-th hour on the \(d\)-th day, and \(r\) represents the number of user nodes with similar power consumption patterns.

[0023] Preferably, the regional anomaly detection model includes an embedding layer, a graph attention layer, a propagation aggregation layer, and a prediction layer. Among them, the embedding layer is used to embed the node features in the weighted graph structure of the distribution network to obtain a node feature embedding matrix. The graph attention layer is used to calculate the attention weights between each node in the weighted graph structure of the distribution network through an attention mechanism, capture the first-order information of the weighted graph structure of the distribution network by weighted representation based on the attention weights between each node in the weighted graph structure of the distribution network, and fuse the first-order information of the weighted graph structure of the distribution network through a fusion gate to obtain a fusion embedding vector corresponding to the weighted graph structure of the distribution network. The propagation aggregation layer is used to perform message passing on the fusion embedding vector corresponding to the weighted graph structure of the distribution network by using a graph neural network to obtain the representations of each layer corresponding to the weighted graph structure of the distribution network. The prediction layer is used to perform an inner product after connecting the representations of each layer with a single vector to obtain a prediction result.

[0024] Preferably, the expression of the attention weight is as follows:

[0025] \(a = \text{softmax}(\hat{a})\);

[0026] Among them, \(a\) represents the attention weight, \(\text{softmax}\) represents the softmax activation function, \(\hat{a}=(\mathbf{W}\) r \(\mathbf{e}\) u )\cdot\tanh(\mathbf{W}\) r \(\mathbf{e}\) i +\mathbf{e}\) r ), \(\mathbf{W}\) r represents the projection of the node from a \(d\) e -dimensional node space to a \(d\)r Dimensional relationship space, W r e i +e r Indicates the feature embedding matrix e of the i-th node i Through the transformation matrix W of edge r r Perform a linear transformation, and then add the relationship vector e of edge r r Offset, tanh represents the hyperbolic tangent activation function;

[0027] The expression of the first-order information of the weighted graph structure of the distribution network is as follows:

[0028]

[0029] Among them, Represents the first-order information of node i of category c.

[0030] Preferably, the expression of the fused embedding vector is as follows:

[0031]

[0032] Among them, g i Represents the fused embedding vector, σ represents the activation function, W c And W k Represents learnable transformation parameters, Represents the first-order information of node i of category k;

[0033] The expressions of the representations of each layer corresponding to the weighted graph structure of the distribution network are as follows:

[0034]

[0035] Among them, Represents the representation of the (l + 1)-th layer in the weighted graph structure of the distribution network, LeakyReLU represents the LeakyReLU activation function, Represents the fused embedding vector of the l-th layer in the weighted graph structure of the distribution network, Represents the fused embedding vector after message passing of the l-th layer in the weighted graph structure of the distribution network.

[0036] Preferably, for the sub-regions with abnormal detection labels as regional abnormal labels in the target region, perform abnormal root cause localization to obtain the corresponding abnormal root causes of the sub-regions, including:

[0037] Based on the graph structure of the distribution network corresponding to the sub-regions with abnormal detection labels as regional abnormal labels in the target region and the preset fault record nodes, construct a fault knowledge graph;

[0038] Performing information dissemination and node representation on the fault knowledge graph based on a pre-trained graph neural network to update the entity node representation of the fault knowledge graph;

[0039] Calculating the anomaly score of each entity node in the fault knowledge graph based on the entity node representation of the updated fault knowledge graph;

[0040] Locating the fault source based on the anomaly scores of each entity node in the fault knowledge graph to obtain the anomaly root cause corresponding to the sub-region.

[0041] Preferably, the fault knowledge graph includes an entity node set and a relationship edge set. Among them, the entity nodes include device nodes, user nodes, and fault record nodes, the relationship edges include connection relationships, fault propagation relationships, and status relationships. The device nodes correspond to the power equipment in the distribution network, the user nodes correspond to the electricity users in the distribution network, the fault record nodes correspond to the records of power equipment fault information and electricity user anomaly occurrence information, the connection relationships correspond to the power lines between nodes, the fault propagation relationships correspond to the influence relationships of device node failures on other entity nodes, and the status relationships correspond to the status connection relationships between device nodes.

[0042] Preferably, the expression for updating the entity node representation of the fault knowledge graph is as follows:

[0043]

[0044] Among them, represents the feature representation of entity node v i at the k+1 layer, N(i) represents the set of neighbor entity nodes of entity node v i , c ij represents a normalization factor used to adjust the weight of the relationship edge, W (k) and b (k) represent the learned weight matrix and bias term;

[0045] The expression for the anomaly score of each entity node in the fault knowledge graph is as follows:

[0046]

[0047] Among them, score(v i ) represents the anomaly score of entity node v i , represents the feature representation of entity node v i at the last layer, represents entity node v iIn the initial feature representation, |||| represents the Euclidean distance calculation function.

[0048] The beneficial effects of the present invention are as follows: (1) By constructing the graph structure of the distribution network and extracting features from the node set, the present invention can effectively capture the operating status of devices and network topology information in the distribution network. The extraction of node features and the construction of the weighted graph enable the model to fully consider different influencing factors of each device and area, thereby improving the accuracy of anomaly detection. By weighting the graph structure, the importance of different devices or areas can be reflected, and the setting of weights can strengthen the anomaly detection of key devices or areas in the network, thereby reducing false alarms and improving the accuracy rate; (2) Based on the pre-trained regional anomaly detection model, the present invention can intelligently perform anomaly detection on the distribution network and judge the anomaly and normal states of each sub-region. Therefore, not only can the overall anomaly situation be identified, but also it can be refined to specific sub-regions, facilitating the accurate identification of the location where the problem occurs. And by introducing regional anomaly labels, the system can classify according to the operating conditions of specific regions, reducing the risk that global detection cannot identify local anomalies; (3) For the detected abnormal regions, the method further locates the root cause of the anomaly. By analyzing the topological structure and node features of the region, the root cause of the fault or anomaly can be quickly located. This process combines artificial intelligence and graph neural networks, making anomaly detection and root cause location more accurate and efficient. The root cause location can quickly identify the specific reasons leading to regional anomalies, helping maintenance personnel respond faster and reducing the risk of power outages and equipment damage. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic flow chart of the overall method in an embodiment proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Embodiment 1, as Figure 1 shown, an abnormal power consumption detection method based on artificial intelligence proposed by the present invention includes:

[0051] S1. Obtain the topological structure of the distribution network in the target area, and construct the graph structure of the distribution network based on the topological structure of the distribution network;

[0052] S2. Extract features from the node set in the graph structure of the distribution network to obtain the node features corresponding to each node in the node set, and weight the graph structure of the distribution network based on the node features to obtain the weighted graph structure of the distribution network;

[0053] S3. Perform anomaly detection on the weighted graph structure of the distribution network based on the pre-trained regional anomaly detection model to obtain the anomaly detection labels of each sub-region in the target area, where the anomaly detection labels include regional anomaly labels and regional normal labels;

[0054] S4. Locate the root cause of the anomaly for the sub-regions with abnormal detection labels as regional anomaly labels in the target region to obtain the root cause of the anomaly corresponding to the sub-regions.

[0055] In the present invention, the distribution network topology refers to the connection relationship and layout structure of various devices (such as transformers, switches, wires, loads, etc.) in the distribution network. It is a structure diagram that describes the nodes (such as power supply devices) and the edges (such as power lines) connecting them in the distribution network. The distribution network topology reflects the physical or electrical connections between the various components within the power system; the regional anomaly label indicates that there are abnormal conditions in the power equipment or network within a specific region, which may be caused by equipment failures, load overloading, etc.; the regional normal label indicates that the power equipment or network within a specific region is operating normally without any abnormal conditions.

[0056] Embodiment 2. An abnormal power consumption detection method based on artificial intelligence proposed by the present invention. Compared with Embodiment 1, this embodiment further includes: The graph structure of the distribution network includes a node set and an edge set. Among them, the node set includes an equipment node set and a user node set. The equipment nodes correspond to the power equipment in the distribution network, and the user nodes correspond to the power consumption users in the distribution network. The edges correspond to the power lines between the nodes.

[0057] In this embodiment, the nodes in the equipment node set are mainly related to the following equipment: Transformers: used to change the voltage level to ensure that electricity is transmitted at an appropriate voltage; switches: used to control the switching and distribution of electricity to avoid overload or failure; circuit breakers: used to disconnect the circuit in case of a failure to protect other equipment in the distribution network.

[0058] In an optional embodiment, feature extraction is performed on the node set in the graph structure of the distribution network to obtain the node features corresponding to each node in the node set, including:

[0059] A1. Perform feature extraction on the user node set in the graph structure of the distribution network to obtain the user node features corresponding to each user node in the user node set;

[0060] A2. Perform feature extraction on the equipment node set in the graph structure of the distribution network to obtain the equipment node features corresponding to each equipment node in the equipment node set. Among them, the equipment node features include the voltage level of the power equipment, the load of the power equipment, the current state of the power equipment, and the switch state of the power equipment.

[0061] In an optional embodiment, feature extraction is performed on the user node set in the graph structure of the distribution network to obtain the user node features corresponding to each user node in the user node set, including:

[0062] B1. Extract the electricity consumption change coefficient for each user node in the user node set in the graph structure of the distribution network to obtain the electricity consumption change coefficient corresponding to each user node in the user node set. The expression of the electricity consumption change coefficient is as follows:

[0063]

[0064] where, u i represents the electricity consumption change coefficient corresponding to the i-th user node in the user node set, Q d represents the user electricity consumption data on the d-th day, and x represents the number of days before and after the electricity consumption data statistics;

[0065] B2. Extract the electricity consumption decline coefficient for each user node in the user node set in the graph structure of the distribution network to obtain the electricity consumption decline coefficient corresponding to each user node in the user node set. The expression of the electricity consumption decline coefficient is as follows:

[0066]

[0067] where, M i represents the electricity consumption decline coefficient corresponding to the i-th user node in the user node set, D(d) represents whether there is a decline in electricity consumption on the d-th day. When the electricity consumption decline coefficient on the d-th day is less than that on the (i - 1)-th day, D(d) takes the value of 1; otherwise, D(d) takes the value of 0;

[0068] B3. Extract the electricity consumption difference coefficient between each user node and other user nodes in the user node set in the graph structure of the distribution network to obtain the electricity consumption difference coefficient between each user node and other user nodes in the user node set. The expression of the electricity consumption difference coefficient is as follows:

[0069]

[0070] where, α i (j, d) represents the electricity consumption difference coefficient between the i-th user node and the j-th user node in the user node set, represents the electricity consumption at the h-th hour on the d-th day, and r represents the number of user nodes with similar electricity consumption patterns.

[0071] In an optional embodiment, the regional anomaly detection model includes an embedding layer, a graph attention layer, a propagation aggregation layer, and a prediction layer. Among them, the embedding layer is used to embed the node features in the weighted graph structure of the distribution network to obtain a node feature embedding matrix. The graph attention layer is used to calculate the attention weights between each node in the weighted graph structure of the distribution network through an attention mechanism, capture the first-order information of the weighted graph structure of the distribution network through weighted representation based on the attention weights between each node in the weighted graph structure of the distribution network, and fuse the first-order information of the weighted graph structure of the distribution network through a fusion gate to obtain a fusion embedding vector corresponding to the weighted graph structure of the distribution network. The propagation aggregation layer is used to perform message passing on the fusion embedding vector corresponding to the weighted graph structure of the distribution network by using a graph neural network to obtain the representations of each layer corresponding to the weighted graph structure of the distribution network. The prediction layer is used to perform an inner product after connecting the representations of each layer with a single vector to obtain a prediction result.

[0072] It should be noted that the embedding layer is a process of converting the original node features into a low-dimensional dense vector space. These node features can be the attribute information of each node in the distribution network. Through embedding, each node in the distribution network can be expressed in a more compact representation (vector) without using the original high-dimensional sparse data. The propagation aggregation layer is usually a part of the graph neural network, used to propagate and aggregate information between nodes in the graph. In this layer, information is "transmitted" through the edges between nodes, and their feature information is aggregated according to the adjacency relationship of each node. The fusion gate usually refers to a mechanism for fusing multiple types of information in the model. In this model, the fusion gate is used to fuse the first-order information (features between nodes and their direct neighbors) and the graph structure information (global structure information of nodes). It can determine the fusion ratio of different information through learning to generate the final fusion embedding vector. The graph neural network is a type of neural network used to process graph-structured data. GNNs transmit information between nodes in the graph through a message passing mechanism and update the feature representation of each node layer by layer. GNNs are particularly effective in processing data with complex structures and can capture the complex dependencies between nodes.

[0073] In an optional embodiment, the expression of the attention weight is as follows:

[0074] a = softmax(á);

[0075] where a represents the attention weight, softmmax represents the softmmax activation function, and á = (W r e u )·tanh(W r e i +e r ), W r represents that the node is composed of d eProject the d-dimensional node space onto r the d-dimensional relationship space, W r e i + e r denotes the feature embedding matrix e of the i-th node i through the transformation matrix W of edge r r for linear transformation, and then add the relationship vector e of edge r r offset, where tanh represents the hyperbolic tangent activation function;

[0076] The expression for the first-order information of the weighted graph structure of the distribution network is as follows:

[0077]

[0078] where represents the first-order information of node i of class c.

[0079] In an alternative embodiment, the expression for the fused embedding vector is as follows:

[0080]

[0081] where gi represents the fused embedding vector, σ represents the activation function, W c and W k represent learnable transformation parameters, represents the first-order information of node i of class k;

[0082] The expressions for the representations of each layer corresponding to the weighted graph structure of the distribution network are as follows:

[0083]

[0084] where represents the representation of the (l + 1)-th layer in the weighted graph structure of the distribution network, LeakyReLU represents the LeakyReLU activation function, represents the fused embedding vector of the l-th layer in the weighted graph structure of the distribution network, represents the fused embedding vector after message passing of the l-th layer in the weighted graph structure of the distribution network.

[0085] In an alternative embodiment, for the sub-regions with the anomaly detection label as the regional anomaly label in the target region, perform anomaly root cause localization to obtain the anomaly root cause corresponding to the sub-regions, including:

[0086] C1. Based on the graph structure of the distribution network corresponding to the sub-regions with the anomaly detection label as the regional anomaly label in the target region and the preset fault record nodes, construct a fault knowledge graph;

[0087] C2. Propagate information and represent nodes in the fault knowledge graph based on a pre-trained graph neural network to update the entity node representations in the fault knowledge graph;

[0088] C3. Calculate the anomaly scores for each entity node in the fault knowledge graph based on the updated entity node representations in the fault knowledge graph;

[0089] C4. Locate the fault source based on the anomaly scores of each entity node in the fault knowledge graph to obtain the anomaly root cause corresponding to the sub-region.

[0090] It should be noted that a knowledge graph is a data structure that represents entities and their relationships through a graph structure. The fault knowledge graph is constructed based on the power distribution network graph structure in the target area and preset fault record nodes. These entities may include equipment nodes (such as transformers, switches, lines, etc.) and fault record nodes (such as historical fault occurrence locations, types, etc.) in the power distribution network. A graph is formed through the connection relationships between these nodes. Fault source location refers to determining the root cause of the area anomaly in the power distribution network based on the anomaly scores. Based on the anomaly scores of each entity node in the fault knowledge graph, nodes with higher scores are found to determine the fault source. Usually, these nodes are related to key equipment or fault history records in the power distribution network and may be the root cause of the anomaly.

[0091] In an optional embodiment, the fault knowledge graph includes an entity node set and a relationship edge set. Among them, the entity nodes include equipment nodes, user nodes, and fault record nodes, and the relationship edges include connection relationships, fault propagation relationships, and status relationships. The equipment nodes correspond to the power equipment in the power distribution network, the user nodes correspond to the electricity users in the power distribution network, the fault record nodes correspond to the nodes that record the fault information of power equipment and the abnormal occurrence information of electricity users, the connection relationships correspond to the power lines between nodes, the fault propagation relationships correspond to the influence relationships of equipment node failures on the remaining entity nodes, and the status relationships correspond to the status connection relationships between equipment nodes.

[0092] In an optional embodiment, the expression for updating the entity node representation of the fault knowledge graph is as follows:

[0093]

[0094] Where, represents the feature representation of entity node v i at the k + 1 layer, N(i) represents the set of neighbor entity nodes of entity node v i , c ij represents the normalization factor used to adjust the weight of the relationship edge, W (k) and b (k)Denote the learned weight matrix and bias term;

[0095] The expression for the anomaly score of each entity node in the fault knowledge graph is as follows:

[0096]

[0097] where score(v i ) represents the anomaly score of entity node v i , represents the feature representation of entity node v i at the last layer, represents the initial feature representation of entity node v i , and |||| represents the Euclidean distance calculation function.

[0098] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art to which the present invention pertains.

Claims

1. A method for detecting abnormal power consumption based on artificial intelligence, characterized in that: include: Acquire a distribution network topology structure of a target area, and construct a distribution network graph structure based on the distribution network topology structure; Extracting features from a node set in a graph structure of the distribution network to obtain node features corresponding to each node in the node set, and weighting the graph structure of the distribution network based on the node features to obtain a weighted graph structure of the distribution network; Based on a pre-trained regional anomaly detection model, anomaly detection is performed on the weighted graph structure of the distribution network to obtain anomaly detection labels of each sub-region in the target region, wherein the anomaly detection labels include regional anomaly labels and regional normal labels; The abnormality root source is located for the sub-region whose abnormality detection label is the regional abnormality label in the target region, so as to obtain the abnormality root source corresponding to the sub-region.

2. The method for detecting abnormal power consumption based on artificial intelligence according to claim 1, characterized in that: The graph structure of the distribution network includes a node set and an edge set, wherein the node set includes a device node set and a user node set, the device node corresponds to the power equipment in the distribution network, the user node corresponds to the power users in the distribution network, and the edge corresponds to the power line between the nodes.

3. The method for detecting abnormal power consumption based on artificial intelligence according to claim 2, characterized in that: Performing feature extraction on a node set in a graph structure of the distribution network to obtain a node feature corresponding to each node in the node set includes: Performing feature extraction on a user node set in a graph structure of the distribution network to obtain a user node feature corresponding to each user node in the user node set; Feature extraction is performed on the device node set in the graph structure of the distribution network to obtain device node features corresponding to each device node in the device node set, wherein the device node features include the voltage level of the power equipment, the load of the power equipment, the current state of the power equipment and the switching state of the power equipment.

4. The method for detecting abnormal power consumption based on artificial intelligence according to claim 3 is characterized in that: Performing feature extraction on a user node set in the graph structure of the distribution network to obtain a user node feature corresponding to each user node in the user node set includes: The power consumption variation coefficient of each user node in the user node set in the graph structure of the distribution network is extracted to obtain the power consumption variation coefficient corresponding to each user node in the user node set, wherein the expression of the power consumption variation coefficient is as follows: Among them, u i represents the power consumption change coefficient corresponding to the i-th user node in the user node set, Q d represents the user's electricity consumption data on the dth day, and x represents the days before and after the electricity consumption data statistics; The power consumption reduction coefficient of each user node in the user node set in the graph structure of the distribution network is extracted to obtain the power consumption reduction coefficient corresponding to each user node in the user node set, wherein the expression of the power consumption reduction coefficient is as follows: Among them, M i It represents the power consumption reduction coefficient corresponding to the i-th user node in the user node set. D(d) indicates whether there is a power consumption reduction on the d-th day. When the power consumption reduction coefficient on the d-th day is less than that on the i-1-th day, D(d) takes the value of 1, otherwise D(d) takes the value of 0. The power consumption difference coefficient between each user node and other user nodes in the user node set in the graph structure of the distribution network is extracted to obtain the power consumption difference coefficient between each user node and other user nodes in the user node set, wherein the expression of the power consumption difference coefficient is as follows: Among them, α i (j, d) represents the difference coefficient of power consumption between the i-th user node and the j-th user node in the user node set, represents the power consumption in the hth hour of the dth day, and r represents the number of user nodes with similar power consumption patterns.

5. The method for detecting abnormal power consumption based on artificial intelligence according to claim 4, characterized in that: The regional anomaly detection model includes an embedding layer, a graph attention layer, a propagation aggregation layer and a prediction layer, wherein the embedding layer is used to embed node features in the weighted graph structure of the distribution network to obtain a node feature embedding matrix, the graph attention layer is used to calculate the attention weights between each node in the weighted graph structure of the distribution network through an attention mechanism, and the first-order information of the weighted graph structure of the distribution network is captured through a weighted representation based on the attention weights between each node in the weighted graph structure of the distribution network, and the first-order information of the weighted graph structure of the distribution network is fused through a fusion gate to obtain a fused embedding vector corresponding to the weighted graph structure of the distribution network, the propagation aggregation layer is used to use a graph neural network to perform message transmission on the fused embedding vector corresponding to the weighted graph structure of the distribution network to obtain the representations of each layer corresponding to the weighted graph structure of the distribution network, and the prediction layer is used to connect the representations of each layer with a single vector and then obtain the prediction result by inner product.

6. The method for detecting abnormal power consumption based on artificial intelligence according to claim 5, characterized in that: The expression of the attention weight is as follows: a=softmax(á); Among them, a represents the attention weight, softmax represents the softmmax activation function, á=(W r e u )·tanh(W r e i +e r ), W r Indicates that the node is represented by d e The dimensional node space is projected to d r dimensional relational space, W r e i +e r Indicates embedding the feature of the i-th node into the matrix e i The transformation matrix W through edge r r Perform a linear transformation and then add the relationship vector e of edge r r The offset of , tanh represents the hyperbolic tangent activation function; The expression of the first-order information of the weighted graph structure of the distribution network is as follows: in, Represents the first-order information of node i of category c.

7. The method for detecting abnormal power consumption based on artificial intelligence according to claim 6, characterized in that: The expression of the fused embedding vector is as follows: Among them, gi represents the fused embedding vector, σ represents the activation function, and W c and W k represents the learnable transformation parameters, Represents the first-order information of node i of category k; The expressions corresponding to each layer of the weighted graph structure of the distribution network are as follows: in, represents the representation of the l+1th layer in the weighted graph structure of the distribution network, LeakyReLU represents the LeakyReLU activation function, represents the fused embedding vector of the lth layer in the weighted graph structure of the distribution network, Represents the fused embedding vector after message passing at the lth layer in the weighted graph structure of the distribution network.

8. The method for detecting abnormal power consumption based on artificial intelligence according to claim 1, characterized in that: The abnormality root source is located for the sub-region whose abnormality detection label is the regional abnormality label in the target region to obtain the abnormality root source corresponding to the sub-region, including: Constructing a fault knowledge graph based on the graph structure of the distribution network corresponding to the sub-region whose abnormal detection label in the target region is the regional abnormal label and the preset fault record nodes; Based on a pre-trained graph neural network, information propagation and node representation are performed on the fault knowledge graph to update entity node representations of the fault knowledge graph; Calculating an anomaly score for each entity node in the fault knowledge graph based on the updated entity node representation of the fault knowledge graph; The fault source is located based on the anomaly score of each entity node in the fault knowledge graph to obtain the anomaly root cause corresponding to the sub-region.

9. The method for detecting abnormal power consumption based on artificial intelligence according to claim 8, characterized in that: The fault knowledge graph includes a set of entity nodes and a set of relationship edges, wherein the entity nodes include device nodes, user nodes and fault recording nodes, and the relationship edges include connection relationships, fault propagation relationships and state relationships. The device nodes correspond to the power equipment in the distribution network, the user nodes correspond to the electricity users in the distribution network, the fault recording nodes correspond to recording the fault information of the power equipment and the abnormal occurrence information of the electricity users, the connection relationships correspond to the power lines between the nodes, the fault propagation relationships correspond to the impact of the failure of the device nodes on the remaining entity nodes, and the state relationships correspond to the state connection relationships between the device nodes.

10. The method for detecting abnormal power consumption based on artificial intelligence according to claim 9, characterized in that: The entity node of the fault knowledge graph represents the updated expression as follows: in, Represents entity node v i In the feature representation of the k+1th layer, N(i) represents the entity node v i The set of neighbor entity nodes, c ij Represents the normalization factor, which is used to adjust the weight of the relationship edge, W (k) and b (k) Represents the learned weight matrix and bias term; The expression of the abnormality score of each entity node in the fault knowledge graph is as follows: Among them, score(v i ) represents the abnormality score of the entity node, Represents entity node v i In the last layer of feature representation, Represents entity node v i In the initial feature representation, |||| represents the Euclidean distance calculation function.

Images