CN-122026354-A - Power distribution network topology state real-time sensing method based on federal deep neural network

CN122026354ACN 122026354 ACN122026354 ACN 122026354ACN-122026354-A

Abstract

The invention discloses a power distribution network topological state real-time perception method based on a federal deep neural network, which belongs to the technical field of power distribution network operation monitoring and comprises the steps of collecting local measurement data of each power distribution network area, preprocessing to obtain an input data matrix, constructing and training a local deep neural network model based on the input data matrix to obtain weights of the local deep neural network models of each area, aggregating the weights of the local deep neural network models of each area by using a federal learning mechanism to obtain a global model, calling the global model to perform online reasoning and merging on the preprocessed local real-time measurement data, outputting a full-network topological state diagram reflecting the full-network connection relation, performing consistency verification on the full-network topological state diagram through a physical constraint and anomaly detection mechanism to obtain a final topological perception result, and identifying anomaly topological changes. The method effectively improves accuracy and instantaneity of sensing the topological state of the power distribution network while guaranteeing the data privacy.

Inventors

ZHOU SUYANG
ZHOU AIHUA
PENG LIN
FAN JILI
GAO MINGYANG
LIU MEIZHAO

Assignees

东南大学
中国电力科学研究院有限公司
国网江苏省电力有限公司信息通信分公司

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. The method for sensing the topological state of the power distribution network in real time based on the federal deep neural network is characterized by comprising the following steps of: Collecting local measurement data of each power distribution network area, and preprocessing to obtain an input data matrix; Based on the input data matrix, constructing and training a local deep neural network model to obtain the weight of the local deep neural network model of each region; the weight of the local deep neural network model of each region is aggregated by utilizing a federal learning mechanism to obtain a global model; invoking the global model to perform online reasoning and merging on the preprocessed local real-time measurement data, and outputting a full-network topology state diagram reflecting the full-network connection relation; And carrying out consistency check on the full-network topology state diagram through a physical constraint and anomaly detection mechanism to obtain a final topology sensing result, and identifying abnormal topology changes.
2. The federal deep neural network-based real-time perception method of topological state of power distribution network according to claim 1, wherein the local measurement data comprises: node electrical quantity information including node voltage amplitude, voltage phase angle, active power injection and reactive power injection; The branch electric quantity information comprises branch active power flow, reactive power flow and switch states.
3. The method for real-time sensing of topology state of power distribution network based on federal deep neural network according to claim 1, wherein the local deep neural network model comprises an input layer, a node feature coding layer, a graph attention propagation layer and an output layer, the local power distribution network topology state is modeled and predicted in real time by feature mapping and representation learning of node and branch operation states, and the prediction process comprises: S21, constructing a graph structure to represent the topological structure of the power distribution network, wherein the node characteristics of the graph structure Including voltage amplitude, voltage phase angle, active and reactive injection power, edge features of the graph Including power flow and switching state; s22, embedding each node in the graph structure by using the node characteristic coding layer: Wherein, the Representing the initial embedding vector of the node, 、 Representing the input layer weights and offsets of the input layer, Representing an activation function; S23, carrying out weighted information propagation on neighbor nodes and edge features by using a graph attention propagation layer, calculating attention weights by using edge features, realizing self-adaptive weighting of key branches, and enabling a local deep neural network to automatically pay attention to nodes and lines which are most sensitive to topology changes; S24, stacking the graph attention propagation layers to capture local and middle-long distance topological relations of the topological state of the power distribution network, and utilizing the combination of an attention mechanism and residual connection to prevent deep network degradation; S25, mapping the nodes into predicted topological states by using an output layer: Wherein, the Representing a predicted probability of switch closure of the switch, Representing the output layer weight matrix, Representing node characteristics of the last layer of the attention spreading layer.
4. The method for sensing the topological state of the power distribution network in real time based on the federal deep neural network according to claim 3, wherein when a local deep neural network model is trained, the output value of the deep neural network is optimized by using a two-class cross entropy loss function and a regularization term; the two classification cross entropy loss functions are: Wherein, the Indicating the actual state of the switch, Representing a predicted probability of switch closure of the switch, A corresponding two-class cross entropy loss for the local deep neural network of the kth local region, Representing an edge set of a graph structure corresponding to a kth local region; The regularization term is: Wherein, the The regularization coefficient is represented as a function of the regularization coefficient, In order to regularize the term(s), As a set of neighbor nodes for node i, To annotate the force propagation layer ith node feature, The j-th node feature of the force propagation layer is annotated.
5. The method for sensing the topological state of the power distribution network in real time based on the federal deep neural network according to claim 1, wherein the step of obtaining the global model by aggregation is as follows: Weights of the local deep neural network model obtained through training Uploading to the federal server, the uploaded weights Does not contain any raw measurement data to protect data privacy; The federal server receives the weight of each local deep neural network model Thereafter, a global model is generated by weighted averaging: Wherein, the The network weight of the global model, K is the total number of local depth networks, Representing a region Is used for the data amount of the (a), Indicating the total data volume of all the regions, Representing the number of iteration rounds; And carrying out online iterative updating on the global model and the local deep neural network model, and issuing the updated global model to each region for the next round of local training or real-time reasoning.
6. The method for sensing the topology state of the power distribution network in real time based on the federal deep neural network according to claim 5, wherein the process of outputting the topology state diagram of the whole network comprises the following steps: Using deep neural networks in global models For the preprocessed local real-time measurement data Real-time reasoning is carried out to obtain the deep neural network in the region k Predictive value for current topology state of branch i-j : S42, combining the predicted values of the topological states of all the areas to form a full-network topological state diagram: Wherein, the Representing a full network topology state diagram.
7. The method for sensing the topology state of a power distribution network in real time based on a federal deep neural network according to claim 1, wherein the step of checking consistency comprises: S51, performing physical consistency check on the full-network topology state diagram, and marking or correcting the prediction result which does not meet the constraint: Wherein, the For the node voltage magnitude value, Representing the rated voltage value of the distribution network, Indicating the active power of the branch circuit, Representing the maximum active power flow allowed by the branch; S52, comparing the current predicted topological state with the historical stable state, and judging whether the topological state of a certain branch or switch has abnormal change or not: Wherein, the An abnormality detection flag is set to 1, an abnormality is set to 0, a normal is set to 0, Is a threshold, represents the allowable range of topology state changes, Representing the predicted value of the neural network for the current topology state of branch i-j in region k, Representing the reference topology state recorded by the branch in the history of operation, Indicating a function, the condition is established, output 1 represents abnormality, otherwise output 0 represents normal.
8. A federal deep neural network based real-time perception system of power distribution network topology, performing the method of any of claims 1-7, comprising: The data acquisition processing module acquires local measurement data of each power distribution network area and performs preprocessing to obtain an input data matrix; The local model building training module is used for building and training a local deep neural network model based on the input data matrix to obtain the weight of the local deep neural network model of each region; The global model aggregation module is used for aggregating the weights of the local deep neural network models of all the areas by utilizing a federal learning mechanism to obtain a global model; The reasoning merging module is used for calling the global model to perform online reasoning and merging on the preprocessed local real-time measurement data and outputting a full-network topology state diagram reflecting the full-network connection relation; And the verification module is used for carrying out consistency verification on the full-network topology state diagram through a physical constraint and anomaly detection mechanism to obtain a final topology sensing result and identifying abnormal topology change.
9. A computer storage medium storing a readable program, wherein the program, when executed, is capable of instructing a computing device to perform the method of fly-ad hoc network routing based on a neural network and mobile awareness as claimed in any one of claims 1 to 7.
10. An electronic device is characterized by comprising a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete communication with each other through the communication bus; The memory is configured to store at least one executable instruction that causes the processor to perform operations corresponding to the load prediction method according to any one of claims 1 to 7.

Description

Power distribution network topology state real-time sensing method based on federal deep neural network Technical Field The invention belongs to the technical field of operation monitoring of power distribution networks, and particularly relates to a federal deep neural network-based real-time sensing method for the topology state of a power distribution network. Background Along with the large-scale access of the distributed power supply, the flexible load and the energy storage device in the power distribution network, the operation form of the power distribution network is gradually changed into a complex system with multiple power supplies, multiple nodes and multiple operation states simultaneously by a traditional unidirectional power supply structure. In the actual operation process, the power distribution network frequently changes the network topology structure through the modes of switching operation, fault isolation, reconstruction and the like so as to meet the requirements of power supply reliability, economy and safety. Therefore, accurately and real-timely acquiring the topology state of the power distribution network becomes an important foundation for key functions such as power distribution network scheduling control, state estimation, fault positioning, self-healing control and the like. However, due to factors such as uneven deployment, communication delay, data loss, complex running environment and the like of the measuring device, the topology state of the power distribution network is difficult to sense reliably in real time in a traditional manual entry or rule matching mode, and the concealment and uncertainty of the topology change are further increased particularly in a complex scene containing a large amount of distributed energy sources. In the prior art, for the topology identification and perception problems of a power distribution network, a state estimation method, a logic inference method or a centralized machine learning model based on measurement data is generally relied on. Most of the methods need to upload the original running data of each area to a master station for unified processing, so that higher requirements are put on communication bandwidth and computing resources, and data privacy and security risks are easily caused in practical application. In addition, the generalization capability and the robustness of the centralized model are obviously limited when the centralized model faces the problems of different regional distribution network structure differences, inconsistent operation characteristics, unbalanced data distribution and the like. Meanwhile, although the topology recognition accuracy is improved in part based on the deep learning method, the physical operation constraint of the power system is often ignored, so that the model output result is difficult to directly apply in engineering and even does not accord with the basic power law. Therefore, how to fully utilize the cooperative information of the multi-region operation data and combine the physical constraint of the power system to realize the high-precision and real-time sensing of the topology state of the power distribution network on the premise of protecting the data privacy is still a technical problem to be solved in the current intelligent operation field of the power distribution network. Disclosure of Invention Aiming at the defects of the prior art, the invention aims to provide a power distribution network topology state real-time sensing method based on a federal deep neural network, which solves the problems in the prior art. The aim of the invention can be achieved by the following technical scheme: a method for sensing the topological state of a power distribution network in real time based on a federal deep neural network comprises the following steps: Collecting local measurement data of each power distribution network area, and preprocessing to obtain an input data matrix; Based on the input data matrix, constructing and training a local deep neural network model to obtain the weight of the local deep neural network model of each region; the weight of the local deep neural network model of each region is aggregated by utilizing a federal learning mechanism to obtain a global model; invoking the global model to perform online reasoning and merging on the preprocessed local real-time measurement data, and outputting a full-network topology state diagram reflecting the full-network connection relation; And carrying out consistency check on the full-network topology state diagram through a physical constraint and anomaly detection mechanism to obtain a final topology sensing result, and identifying abnormal topology changes. Further, the local measurement data includes: node electrical quantity information including node voltage amplitude, voltage phase angle, active power injection and reactive power injection; The branch electric quantity information comprises branch active powe