An edge-computing-based power system load prediction method

Abstract

This invention discloses a power system load forecasting method based on edge computing, belonging to the field of power load forecasting technology. Specifically, it includes: edge nodes first collecting local real-time operating state sequences, extracting local state codes through a local mapping layer, and then uploading them to a cloud center; the cloud center aggregates the codes of all edge nodes, combining the spatial coordinate connections of each node and real-time switching states to dynamically reconstruct a global topology feature map; subsequently, this feature map is input into a cloud-based graph neural network for feature aggregation, outputting customized features for the target edge nodes with spatial coupling awareness; next, the customized features are input into a cloud-based hypernetwork model to generate node network parameter tensors adapted to the current transient topology and distributed to the target edge nodes to update their local lightweight prediction networks; finally, the target edge nodes input the current state sequence into the updated prediction network for forward inference, outputting energy flow state predictions and limit exceedance risk assessment values.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and specifically to a power system load forecasting method based on edge computing. Background Technology

[0002] The in-depth development of smart grids and the large-scale integration of distributed microgrids have led to highly dynamic and complex changes in the operating status and electricity consumption behavior of power systems. This makes accurate load forecasting a core foundation for power system dispatching and planning, demand-side response, and energy optimization. In recent years, deep learning technology, with its superior nonlinear fitting and multi-dimensional time series mining capabilities, has been widely applied in the field of load forecasting. Simultaneously, the popularization of IoT technology has prompted the power system's computing architecture to gradually evolve from traditional centralized cloud processing to a cloud-edge collaborative edge computing architecture. Under this new computing paradigm, a large number of edge nodes deployed in distribution substations or on the user side can collect basic electrical characteristic data of their respective regions in real time. By processing massive amounts of high-frequency measurement status sequences locally, this not only effectively alleviates the transmission bandwidth pressure on the backbone communication network but also significantly reduces communication latency in data interaction, providing solid technical support for building a highly real-time and highly reliable distributed load forecasting network.

[0003] However, existing edge computing-based load forecasting solutions still face multi-dimensional technical bottlenecks in practical engineering applications. Edge terminals at different levels exhibit significant heterogeneity in geographical environment and load composition, making it difficult for traditional, single, global forecasting models to accurately fit the localized load variation patterns of each node. Simultaneously, edge computing devices face strict physical constraints in terms of computing power, storage space, and energy consumption, making it impossible to directly deploy and run computationally intensive, large-scale deep learning forecasting models. Existing lightweight networks or conventional distributed solutions, while adapting to limited edge resources, often severely sacrifice the ability to extract features from complex nonlinear power load sequences, leading to a significant decrease in the prediction accuracy and model generalization performance of local nodes. Summary of the Invention

[0004] The purpose of this invention is to provide a power system load forecasting method based on edge computing, thereby solving the problems in the background technology: The objective of this invention can be achieved through the following technical solutions: A power system load forecasting method based on edge computing includes the following steps: S1. Edge nodes collect local real-time operating status sequences, extract local status codes through the local mapping layer, and upload them to the cloud center. S2. The cloud center collects the local status codes of all edge nodes and dynamically reconstructs the global topology feature map based on the spatial coordinates of each node and the real-time uploaded physical switch status. S3. Input the dynamically reconstructed global topology feature map into the cloud graph neural network for feature aggregation calculation, and output customized features of the target edge nodes with the latest spatial coupling perception capabilities. S4. Input the node-customized features into the cloud-based hypernetwork model for matrix mapping calculation to generate a target edge node network parameter tensor that adapts to the current transient topology. S5. The cloud center sends the network parameter tensor to the target edge node and dynamically loads and updates it into the lightweight prediction network on the edge. S6. The target edge node inputs the current state sequence into the lightweight prediction network loaded with the latest parameters for forward inference, and outputs the energy flow state prediction and limit risk assessment value of the edge node.

[0005] As a further aspect of the present invention: In step S1, the process of the edge node collecting the local real-time operating status sequence, extracting the local status code through the local mapping layer, and uploading it to the cloud center is as follows: Edge nodes collect real-time voltage amplitude, current phase, and physical switch status of their respective areas. The acquired basic electrical characteristic data are then arranged and spliced ​​in chronological order of sampling time to form a continuous local real-time operating status sequence. Edge nodes input the local real-time running state sequence into the local mapping layer, and use the neural network structure inside the local mapping layer to extract the nonlinear correlation features inside the sequence, and generate the local state code through matrix linear transformation mapping calculation; Edge nodes acquire the generated local status code, attach the corresponding node's location coordinates and unique device identification code to the code data, encapsulate it into a transmission message according to the standard communication protocol format, and upload the message to the cloud center through a dedicated communication link.

[0006] As a further aspect of the present invention: In step S2, the cloud center collects the local state codes of all edge nodes, and dynamically reconstructs the global topology feature map based on the spatial coordinate connections of each node and the real-time uploaded physical switch status. The cloud center receives all messages uploaded by edge nodes, performs data parsing on the messages, and extracts the local status code, node spatial coordinates and physical switch status of the corresponding edge nodes. The cloud center maps the spatial coordinates of nodes to the node elements of the graph structure, and constructs the initial edges that reflect the adjacency relationship of the physical structure based on the inherent spatial coordinate connection relationship between the node elements. The cloud center reads the extracted physical switch status, performs connectivity determination on the initial edges based on the closed and open attributes of the physical switch status, and removes the initial edges corresponding to the open attribute to generate a dynamic topology; The cloud center integrates all local state codes into the node elements of the dynamic topology, completes the splicing of node feature data and dynamic connectivity attributes, and dynamically reconstructs a global topology feature map.

[0007] As a further aspect of the present invention: in step S3, the process of inputting the dynamically reconstructed global topology feature map into a cloud graph neural network for feature aggregation calculation, and outputting customized features of the target edge nodes with the latest spatial coupling perception capability, is as follows: The cloud center inputs the dynamically reconstructed global topology feature map into the cloud graph neural network. Based on the dynamic connectivity properties inside the global topology feature map, the cloud graph neural network identifies and extracts the node feature data of all adjacent nodes directly connected to the target edge node. The cloud-based graph neural network uses an internally defined weight matrix to perform linear mapping calculations on the extracted node feature data of each adjacent node, generating corresponding transmitted feature vectors, and then transmitting the transmitted feature vectors along the connected edge structure to the target edge node. The cloud-based graph neural network calls the feature aggregation function to receive all the transmitted feature vectors that are gathered to the target edge node, performs pooling summation on all the received transmitted feature vectors, and generates a neighborhood aggregation feature matrix that integrates the surrounding physical topology information. The cloud-based graph neural network concatenates the generated neighborhood aggregation feature matrix with the original node feature data of the target edge nodes, inputs the concatenated data into a nonlinear activation function for update calculation, and outputs customized features of the target edge nodes with the latest spatial coupling perception capabilities.

[0008] As a further aspect of the present invention: in step S4, the process of inputting the node-customized features into the cloud-based hypernetwork model for matrix mapping calculation to generate the target edge node network parameter tensor adapted to the current transient topology is as follows: The cloud center inputs the customized features of the target edge nodes into the basic mapping layer of the cloud hypernetwork model. The basic mapping layer performs nonlinear spatial transformation calculations on the customized features of the target edge nodes and outputs intermediate state vectors. The weight generation module of the cloud-based hypernetwork model receives the intermediate state vector and performs continuous matrix multiplication mapping operations on the intermediate state vector based on the network layer dimension parameters of the lightweight prediction network to generate a one-dimensional parameter sequence. The tensor reconstruction module of the cloud-based super network model obtains all one-dimensional parameter sequences, performs shape reshaping operations on the one-dimensional parameter sequences according to the neuron connection specifications inside the lightweight prediction network, and constructs a multi-dimensional weight matrix. The cloud-based hypernetwork model performs a concatenation and combination operation on all multidimensional weight matrices in hierarchical order to generate a set of network layer parameters, and finally outputs a target edge node network parameter tensor that adapts to the current transient topology.

[0009] As a further aspect of the present invention: the specific content of performing continuous matrix multiplication mapping operations on the intermediate state vector to generate a one-dimensional parameter sequence is as follows: The weight generation module extracts the network layer dimension parameters of the lightweight prediction network and constructs multiple concatenated linear transformation matrices with corresponding row and column dimensions based on these network layer dimension parameters. The weight generation module inputs the received intermediate state vector into the first linear transformation matrix for calculation, and passes the output result of each step to the next linear transformation matrix to perform continuous matrix multiplication mapping operations. The weight generation module extracts the final mapping features calculated from the terminal linear transformation matrix, performs a data flattening transformation operation on the final mapping features, and generates a one-dimensional parameter sequence by combining and arranging them according to the inherent order.

[0010] As a further aspect of the present invention: In step S5, the process by which the cloud center sends the network parameter tensor to the target edge node and dynamically loads and updates it into the lightweight prediction network locally at the edge is as follows: The cloud center obtains the network parameter tensor of the target edge node, performs serialization encoding on the network parameter tensor to convert it into a binary data stream, and encapsulates it into a communication message according to the basic communication protocol specification. The cloud center activates and establishes a downlink communication link, and sends the communication message along the downlink communication link to the target edge node. The target edge node receives the communication message through its listening port. The target edge node performs data parsing and verification comparison on the received communication messages. After confirming that the messages are correct, it performs deserialization decoding to extract and restore the network parameter tensor from the messages. The target edge node reads the restored network parameter tensor and, according to the network hierarchy, directly replaces and overwrites the tensor value into the memory address of the lightweight prediction network on the edge, performing dynamic loading and updating.

[0011] As a further aspect of the present invention: In step S6, the process by which the target edge node inputs its current state sequence into a lightweight prediction network loaded with the latest parameters for forward inference, and outputs the edge node's energy flow state prediction and limit exceedance risk assessment value, is as follows: The target edge node feeds the current state sequence into the lightweight prediction network that has completed parameter loading, and performs matrix multiplication and activation mapping forward inference operations sequentially along the internal layers of the network to generate hidden state feature vectors. The state output layer of the lightweight prediction network receives the hidden state feature vector, performs linear dimensionality reduction mapping calculation on it using the internal weight matrix, reconstructs and outputs the energy flow state prediction value of the target edge node; The target edge node reads the predicted energy flow state value, performs a one-way difference comparison operation with the preset safe operation boundary, and maps the deviation result of the obtained predicted value that approaches or exceeds the safe operation boundary as the over-limit risk assessment value of the target edge node.

[0012] The beneficial effects of this invention are: This invention effectively solves the problem of insufficient prediction accuracy caused by the heterogeneity of edge nodes, accurately adapting to the personalized load change patterns of each node. By extracting unique state codes through the local mapping layer of edge nodes and combining them with a dynamically reconstructed global topology feature map in the cloud, it achieves deep fusion of local personalized features and global spatial coupling features, breaking the adaptation limitations of traditional single global models. The feature aggregation calculation of the cloud-based graph neural network can accurately capture the correlation characteristics between edge nodes with different geographical environments and different loads. Furthermore, by generating customized network parameters for each node through a hypernetwork model, the lightweight prediction network for each edge node can fit its own load change patterns, significantly improving the prediction accuracy of local nodes and the model's generalization performance, effectively avoiding prediction bias caused by node heterogeneity.

[0013] This invention resolves the contradiction between edge-side resource constraints and feature extraction capabilities, achieving a dual improvement in both efficient resource utilization and predictive performance. It eliminates the need to deploy computationally intensive deep learning models at the edge, performing forward inference solely through a local lightweight prediction network. This significantly reduces the computational power, storage space, and energy consumption of edge devices, adapting to the strict physical constraints of the edge. Simultaneously, the cloud, through global feature aggregation and customized parameter generation, compensates for the feature extraction shortcomings of conventional lightweight networks and distributed solutions. This ensures that, while adapting to edge resources, the characteristics of complex nonlinear power load sequences are fully exploited. This guarantees real-time prediction while achieving accurate output of energy flow state prediction and limit-over-limit risk assessment, enhancing the reliability and practicality of engineering applications. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating a power system load forecasting method based on edge computing according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, this invention is a power system load forecasting method based on edge computing, comprising the following steps: S1. Edge nodes collect local real-time operating status sequences, extract local status codes through the local mapping layer, and upload them to the cloud center. S2. The cloud center collects the local status codes of all edge nodes and dynamically reconstructs the global topology feature map based on the spatial coordinates of each node and the real-time uploaded physical switch status. S3. Input the dynamically reconstructed global topology feature map into the cloud graph neural network for feature aggregation calculation, and output customized features of the target edge nodes with the latest spatial coupling perception capabilities. S4. Input the node-customized features into the cloud-based hypernetwork model for matrix mapping calculation to generate a target edge node network parameter tensor that adapts to the current transient topology. S5. The cloud center sends the network parameter tensor to the target edge node and dynamically loads and updates it into the lightweight prediction network on the edge. S6. The target edge node inputs the current state sequence into the lightweight prediction network loaded with the latest parameters for forward inference, and outputs the energy flow state prediction and limit risk assessment value of the edge node.

[0018] In one embodiment of the present invention, the process of the edge node collecting local real-time operating status sequences, extracting local status codes through a local mapping layer, and uploading them to the cloud center in step S1 is as follows: Edge nodes are responsible for collecting real-time basic electrical characteristic data within their jurisdiction. This data includes real-time voltage amplitude, current phase, and the on / off status of physical switches at each node. To ensure data timeliness and accuracy, edge nodes are configured with high-precision sensor arrays for synchronous sampling, for example, setting the sampling frequency to 100 times per second, thus acquiring continuous physical quantity values ​​within extremely short time intervals. In a specific operating scenario, at a certain moment, the real-time voltage amplitude collected by the edge node might be 220 volts, corresponding to a current phase angle of 45 degrees, and the associated physical switch status might be read as closed (typically represented by the number 1 for closed and 0 for open). The edge node aligns these three heterogeneous data types—voltage amplitude, current phase, and switch status—acquired at the same sampling moment to form a multi-dimensional feature vector. Over time, the edge node arranges and horizontally concatenates these feature vectors collected at different timestamps, strictly following the chronological sampling order. Through this time-dependent arrangement and splicing method, the edge nodes successfully constructed a continuous local real-time operating status sequence that reflects the operating trend of power equipment in the region. This sequence is presented in a two-dimensional matrix format containing the time dimension, providing rich basic feature inputs for the prediction model.

[0019] After constructing a continuous local real-time running state sequence, the edge node inputs this data tensor into a local mapping layer deployed in local memory for deep feature mining. The core architecture of the local mapping layer includes a multi-layer feedforward neural network structure with pre-defined activation functions and weight parameters. When the local real-time running state sequence enters this mapping layer, neurons in each layer capture hidden nonlinear correlation features between different time steps within the sequence, such as the complex nonlinear mapping relationship between voltage fluctuations and current phase shifts. To achieve lightweight data representation while preserving core feature information, the local mapping layer, after extracting high-dimensional nonlinear features, calls its internal mapping module to perform matrix linear transformation mapping calculations. Specifically, this involves multiplying the extracted high-dimensional feature tensor with a pre-trained dimensionality-reduction weight matrix. Assuming the input high-dimensional feature has a dimension of 1000, after multiplication with a 1000×64 weight matrix, the result is directly compressed into a one-dimensional vector of length 64. The one-dimensional vector generated after nonlinear feature extraction and linear matrix mapping calculation is the final local state code. This code not only highly condenses the operating rules of the original electrical data, but also effectively reduces the computational load and data redundancy of edge nodes.

[0020] After successfully acquiring the generated local status code, the edge node needs to perform information enhancement and encapsulation processing on the encoded data to ensure that the cloud center can accurately identify the spatial source and device affiliation of the data. The edge node calls its local configuration file to read and extract its inherent geospatial coordinates and the unique device identification code burned into the device at the factory. For example, the extracted location coordinates are 120 degrees east longitude and 30 degrees north latitude, and the unique device identification code is a long string. The edge node uses the acquired location coordinate data and the unique device identification code as additional tags and directly concatenates them into the header area of ​​the local status code. After information concatenation, the edge node strictly follows the preset standard communication protocol format to package the fused data stream, along with necessary check bits and frame header and trailer information, into a standard format transmission message. After encapsulation, the edge node activates its underlying communication network card and uploads the transmission message to the cloud center via a highly reliable directional routing method through a pre-established mobile communication network or dedicated communication link. Upon receiving the message, the cloud center can parse the corresponding edge status information. This process ensures the security and stability of data transmission from the edge to the cloud.

[0021] In one embodiment of the present invention, in step S2, the cloud center collects the local state codes of all edge nodes, and dynamically reconstructs the global topology feature map based on the spatial coordinate connections of each node and the real-time uploaded physical switch status. During operation, the cloud center continuously monitors the communication port to receive real-time transmission messages uploaded by all edge nodes distributed across different geographical regions. In a specific application scenario, the cloud center concurrently receives communication messages from 100 different edge nodes within one second. To obtain the valid electrical characteristic information contained within the messages, the cloud center's parsing module performs deep data parsing operations on these encapsulated messages. The parsing process first strips away the communication frame header and checksum sequence to extract the core data payload, and then performs precise positioning and extraction according to a pre-agreed data format template. Through this separation and extraction mechanism, the cloud center can accurately separate the local state code representing the local operating characteristics from the messages of each edge node, such as a feature vector of length 64. It can also parse the node's spatial coordinates, such as longitude 120° and latitude 30°, and the current physical switch state attached to the node, such as using the number 1 to indicate closure and the number 0 to indicate openness. This precise data parsing and extraction process provides the most basic and core data support for the construction of the entire topology.

[0022] After obtaining the geographic location information of all edge nodes, the cloud center will activate the underlying graph construction engine to map the extracted node spatial coordinates into abstract node elements in the graph structure. Specifically, the cloud center will convert each physical device with independent latitude and longitude coordinates into a vertex in the virtual graph space. For example, the device corresponding to longitude 120 and latitude 30 will be mapped to node element number 1 in the graph structure, and the other coordinates will be mapped to node elements number 2 to 100. After completing the mapping and generation of node elements, the cloud center will further construct initial edges reflecting the adjacency relationships of the physical structure based on the inherent spatial coordinate connections of these node elements in the real physical world. The cloud center stores a set of static physical line survey data, which records the fixed power transmission cable connections between nodes. For example, there is a 15-kilometer-long physical cable between node 1 and node 5. Based on this inherent physical structure adjacency association information, the cloud center directly constructs initial edges between the corresponding graph node elements, and finally constructs 150 initial edges reflecting static connection relationships in the graph space containing 100 node elements.

[0023] After constructing the initial structure containing static physical connections, the cloud center calls the connectivity assessment module to read the real-time physical switch status data extracted in the previous analysis. The energy flow paths in the distribution network change drastically with switching actions; therefore, strict connectivity determination must be performed on each generated initial edge based on the closing and opening attributes of the physical switches. The cloud center traverses the 150 initial edges constructed above and checks the real-time switch status variables corresponding to the node elements at both ends of each edge. For example, when determining the initial edge between node element 1 and node element 5, if the physical switch status value controlling the line is read as 0, indicating a disconnection attribute, the cloud center determines that the initial edge does not have the ability to transmit energy and information at the current moment. Based on this clear determination, the graph construction engine immediately performs a pruning operation, accurately removing initial edges that contain disconnection attributes. After traversing and filtering the entire network, the cloud center may remove 20 edges in a disconnected state from the total 150 initial edges, thus using the remaining 130 effective connected edges to generate a dynamic topology that truly reflects the transient physical connectivity status of the current equipment.

[0024] After successfully acquiring the dynamic topology reflecting the current connectivity state, the cloud center needs to inject multi-dimensional operational features into this purely geometric graph structure. The cloud center strictly follows the mapping relationship between node numbers, fusing all previously extracted local state codes into the node elements of the dynamic topology. For example, the cloud center directly assigns the 64-dimensional local state code data containing implicit voltage and current features uploaded by node number 1 to the feature attribute dimension of node number 1 in the dynamic topology, while performing the same feature injection operation on all 100 node elements in the dynamic topology. Through this deep numerical fusion and assignment, the cloud center successfully completes the splicing of multi-dimensional node feature data with the feature matrix of dynamic connectivity attributes in the graph space. The final generated data structure not only includes a 100x100 adjacency matrix describing transient physical connections but also a 100x64 node feature matrix describing the real-time electrical characteristics of each node. The organic combination of these two matrices signifies that the cloud center has successfully dynamically reconstructed a global topology feature graph with complete spatiotemporal dimensional expression capabilities.

[0025] In one embodiment of the present invention, step S3, which involves inputting the dynamically reconstructed global topology feature map into a cloud-based graph neural network for feature aggregation calculation and outputting customized features of the target edge nodes with the latest spatial coupling perception capability, is as follows: After completing the graph structure construction steps described above, the cloud center will directly input the dynamically reconstructed global topology feature graph, containing 100 node elements and 130 effective connected edges, as a whole data stream into the cloud graph neural network deployed in the cloud server cluster. This cloud graph neural network has a powerful ability to process non-Euclidean spatial data. After the full graph data is loaded into memory, the core computing module of the network will first locate the target edge node that needs to be predicted for the current load. Let's assume we select edge node number 5 as the target edge node for the current calculation. In order to obtain the spatial state influence factors around the target edge node, the cloud graph neural network will strictly follow the internal structure of the global topology feature graph. The dynamic connectivity attribute is updated in real time for neighbor search. Specifically, it involves traversing outwards along the valid edges directly connected to node 5. Assuming that connectivity identification reveals that only nodes numbered 2, 8, and 15 have physically closed edge structures connected to node 5, the cloud graph neural network will determine these three nodes as direct neighbors of the target edge node. After determination, the feature extraction module of the cloud graph neural network will access the feature matrix storage area to accurately extract the node feature data of these three neighboring nodes. For example, the local state encoding vector of length 64 attached to each of these three neighboring nodes is completely extracted as the basic raw material for the feature interaction to be performed.

[0026] After successfully extracting the feature data of three directly adjacent nodes (numbered 2, 8, and 15), the cloud-based graph neural network performs preliminary feature transformation using a pre-set weight matrix trained through backpropagation to unify the feature space dimensions of different nodes and enhance feature representation capabilities. Specifically, the cloud-based graph neural network performs linear mapping calculations based on matrix multiplication for each extracted 64-dimensional feature vector of adjacent nodes. Assuming the current layer's weight matrix has a dimension of 64 x 128, the initial 64-dimensional feature vectors of the three adjacent nodes are respectively multiplied by... Multiplying these weight matrices expands the original low-dimensional feature space to a higher-dimensional latent feature space, successfully generating three corresponding transitive feature vectors, each with a dimension of 128. These three upgraded transitive feature vectors encode deep fusion information about the operating status of adjacent devices and the topological location of nodes. After generation, the graph message passing engine of the cloud-based graph neural network is activated. This engine guides the three 128-dimensional transitive feature vectors to move precisely along the previously identified connected edge structure, that is, along the virtual topological channels with physical connections, accurately transmitting the vector data containing neighborhood state information to the memory area of ​​the processing unit where the target edge node numbered 5 is located.

[0027] As the information transmission process concludes, the target edge node numbered 5 at the receiving end has already gathered information from its surrounding environment in its corresponding graph structure memory pool. At this point, the cloud-based graph neural network calls a specially designed feature aggregation function to receive all the transmitted feature vectors gathered at the target edge node. In this specific scenario, what is received are the three 128-dimensional transmitted feature vectors transmitted from nodes numbered 2, 8, and 15. To integrate the neighbor state information scattered in different physical directions into a global-local environment description, the cloud-based graph neural network uses the feature aggregation function to perform pooling summation on all received transmitted feature vectors. The specific computational details involve the summation of these three 128-dimensional vectors... For each corresponding value, a scalar addition operation is performed, that is, the first element of the three vectors is added together to form the first element of the new vector, the second element is added together to form the second element of the new vector, and so on until all 128 dimensions are accumulated and summed. This pooling summation operation can effectively capture and retain the total energy flow influence of each node in the neighborhood on the target node, avoiding information omission. After this element-by-element summation operation, the cloud graph neural network successfully generates a brand-new one-dimensional 128-dimensional vector structure. This generated vector is essentially a neighborhood aggregation feature matrix that integrates the surrounding physical topology information and the electrical characteristic state of the neighbors. It represents the local micro-grid operating environment where the target edge node is located at this moment.

[0028] After obtaining the neighborhood aggregation feature matrix reflecting the influence of the surrounding environment, the cloud graph neural network needs to combine this external environmental influence with the original operating state of the target node itself, thus performing a data concatenation and fusion operation. At this time, the cloud graph neural network will read the original node feature data of the target edge node numbered 5. Assuming that this original feature is consistent with the initial neighboring node and is also a 64-dimensional local state code, the network operation unit will directly concatenate the newly generated 128-dimensional neighborhood aggregation feature matrix with the original node feature data of this 64-dimensional target edge node on the axis of the feature dimension. After the concatenation operation, the two originally separate feature vectors are combined into a joint feature tensor of length 192 dimensions. This tensor contains both its own real-time state and... Having acquired spatial interaction information from surrounding nodes, the cloud-based graph neural network then inputs the concatenated data into an internally configured nonlinear activation function for update calculations. In this stage, a linear rectified activation function can be used to iterate through all elements of the 192-dimensional joint feature tensor, resetting negative values ​​less than zero to zero while retaining the original values ​​of elements greater than zero. This round of calculation, filtering, and numerical mapping by the nonlinear activation function not only removes feature redundancy but also makes the final output vector highly robust, ultimately successfully outputting customized features for target edge nodes with the latest spatial coupling perception capabilities. These customized features are a 192-dimensional high-order representation tensor.

[0029] In one embodiment of the present invention, step S4, the process of inputting the node-customized features into the cloud-based hypernetwork model for matrix mapping calculation to generate a target edge node network parameter tensor adapted to the current transient topology, is as follows: After acquiring the customized features of the target edge nodes with the latest spatial coupling perception capabilities, the cloud center activates the cloud hypernetwork model pre-deployed within the cloud computing nodes. The cloud center directly inputs the acquired customized features of the target edge nodes into the front-end structure of this cloud hypernetwork model, namely the basic mapping layer. Assuming the input customized features are a high-order representation tensor with a dimension of 256, the basic mapping layer typically consists of multiple layers of cascaded fully connected deep feedforward networks, with activation function elements such as modified linear units configured between each layer. Once this 256-dimensional customized feature tensor enters the basic mapping layer, the computing units at this level immediately take over the computation... According to the flow, the deep nonlinear spatial transformation calculation is performed on the customized features of the target edge node using the internally initialized transformation matrix. This spatial transformation calculation is essentially mapping the original local topological association features to a more open and expressive hidden parameter space. For example, after three matrix multiplications and nonlinear activation filtering operations, the basic mapping layer successfully expands and transforms the input 256-dimensional tensor, and finally outputs a stable intermediate state vector with a dimension of up to 1024 in the memory buffer. This generated intermediate state vector not only fully inherits the transient physical topological association information of the edge node, but also has the numerical basis to guide the generation of weights in the lightweight network through high-dimensional spatial expansion.

[0030] The weight generation module of the cloud-based hypernetwork model receives intermediate state vectors and, based on the network layer dimension parameters of the lightweight prediction network, performs continuous matrix multiplication mapping operations on the intermediate state vectors to generate a one-dimensional parameter sequence; specifically: Before starting specific numerical calculations, the weight generation module needs to perform structural analysis on the lightweight prediction network running at the edge. This module reads and extracts the dimensional parameters of each network layer within the lightweight prediction network. For example, in a specific configuration, a hidden layer of the target network contains 64 input nodes and 32 output nodes. Based on these specific network layer dimensional parameters, the weight generation module dynamically constructs multiple cascaded linear transformation matrices responsible for spatial mapping in memory. To achieve a smooth and high-dimensional transition from intermediate states to final parameters, three cascaded linear transformation matrices can be constructed. Assuming the dimension of the input intermediate state is 1024, the row and column dimensions of the first linear transformation matrix are automatically set to 1024 x 512; the row and column dimensions of the second linear transformation matrix are set to 512 x 256; and the third, which is also the final linear transformation matrix, needs to precisely match the overall capacity requirements of the target network parameters, and its internal structure is adjusted according to the final target parameter quantity. These transformation matrices with corresponding row and column dimensions are logically cascaded and combined sequentially, forming the physical computation channel for the supernetwork to perform complex feature parsing and reconstruction.

[0031] After the matrix structure is constructed, the weight generation module officially initiates the data flow calculation process. This module first inputs the high-dimensional intermediate state vector carrying spatial topological information received earlier into the newly established first linear transformation matrix. Assuming this input intermediate state vector is a one-dimensional tensor of dimension 1024, it undergoes a standard dot product with the first linear transformation matrix of dimension 1024 x 512. This mapping operation outputs a new transition feature vector of dimension 512. Next, the calculation engine, following a predetermined concatenated topological path, propagates the result of this calculation down level by level. This transition feature vector of dimension 512 is immediately fed into the next, second linear transformation matrix for computation. The second matrix has a dimension of 512 x 256, and through multiplication and fusion calculations, the feature dimension is compressed and refined to 256 dimensions. This method of accurately propagating the output result to the next linear transformation matrix ensures the continuity of feature transformation. By performing such continuous matrix multiplication mapping operations, the abstract topological patterns implicit in the original state vector are deconstructed layer by layer and gradually transformed into specific numerical distribution patterns that adapt to the weights of the neural network.

[0032] As the cascaded operation reaches the last processing node, the weight generation module needs to process and collect the data generated by the computing terminal. This computing unit extracts the final mapping feature calculated from the terminal linear transformation matrix from the computing memory. Considering the example network layer with 64 input nodes and 32 output nodes, the total number of target weight parameters to be generated is 2048; therefore, the final output structure of the terminal linear transformation matrix is ​​designed as a feature tensor containing 2048 floating-point values. To accommodate the specific format requirements of the parameter loading, the weight generation module must call the underlying processing function to perform a data flattening transformation operation on the final mapping feature. Specifically, regardless of whether the final mapping feature is presented in memory as a 2x1024 two-dimensional matrix or a multi-dimensional three-dimensional structure, the flattening operation will forcibly erase its multi-dimensional attributes, stretching all its values ​​into a single row or column of data. After the flattening transformation is completed, the data processing module will recombine these scattered values ​​according to the inherent order of continuous memory address allocation. After a precise recombination process, a one-dimensional parameter sequence with a fixed length of 2048 is successfully generated. Each value in this sequence corresponds precisely to the initial weight of a connection channel within the target lightweight prediction network.

[0033] Once the one-dimensional parameter sequence calculation is complete, the tensor reconstruction module of the cloud-based hypernetwork model immediately intervenes in the data processing chain, retrieving all the newly generated one-dimensional parameter sequences from the shared memory area. However, the one-dimensional data structure cannot directly adapt to the multi-dimensional matrix multiplication requirements within the neural network. Therefore, the tensor reconstruction module must process these flattened values ​​according to the strictly defined neuron connection specifications within the lightweight prediction network at the edge. These neuron connection specifications detail the mapping topology between input and output nodes in different network layers. Based on this specification, the tensor reconstruction module calls the underlying tensor operation function library. The module performs precise shape reshaping on the read one-dimensional parameter sequence containing 2048 floating-point values. The reshaping process is essentially adjusting the dimensional properties of the tensor without changing the original numerical order, which means folding and transforming a long single-dimensional sequence. For example, the tensor reconstruction module successfully transforms the 2048-digit data sequence into a multi-dimensional weight matrix according to a two-dimensional arrangement of 64 rows and 32 columns. This newly constructed multi-dimensional weight matrix perfectly matches the parameter shape requirements of a specific layer of the target edge node prediction model, providing the core computing carrier and weight benchmark for the upcoming edge inference task.

[0034] Since a complete lightweight prediction network is typically composed of multiple cascaded computational layers of varying depths, the above steps are executed iteratively within the supernetwork to generate multiple sets of weight matrices that match the requirements of all layers. After all the required matrix data for all network layers has been generated, the cloud-based supernetwork model performs a high-dimensional concatenation and combination operation on all the obtained multidimensional weight matrices according to the strict hierarchical order of the lightweight prediction network from the input layer to the hidden layer and then to the output layer. For example, the 64x32 dimension matrix generated by the first layer, the 32x16 dimension matrix generated by the second layer, and the 16x1 dimension matrix generated by the third layer are packaged and integrated to generate a single matrix. A unified set of network layer parameters; this splicing and combination operation not only ensures the continuous distribution of model parameters in physical memory, but also ensures that there will be no hierarchical disorder errors during data loading; through this global parameter aggregation and format encapsulation, the cloud center finally successfully generates and outputs the target edge node network parameter tensor adapted to the current transient topology; this network parameter tensor is like an algorithm brain specially tailored for the target node in the current specific network connectivity state, completely avoiding the defects of traditional fixed parameter models that are difficult to adapt to drastic topology changes, and laying an extremely accurate and highly agile intelligent foundation for the local prediction tasks that the edge nodes will soon carry out.

[0035] In one embodiment of the present invention, the process in step S5 whereby the cloud center sends the network parameter tensor to the target edge node and dynamically loads and updates it into the lightweight prediction network at the edge is as follows: After generating the target edge node network parameter tensor adapted to the current transient topology, the cloud center needs to transmit it securely and efficiently to the corresponding physical device. The data scheduling module within the cloud center first retrieves the target edge node network parameter tensor from the high-frequency cache region, assuming this parameter tensor contains 2048 single-precision floating-point values. To adapt to network bandwidth limitations and ensure cross-platform compatibility, the cloud center calls the underlying encoding library to perform rigorous serialization encoding on the network parameter tensor, transforming it from a multi-dimensional, memory-resident data structure into a compact, continuous binary data stream. In this transformation step… In this process, each floating-point value is precisely mapped to a 32-bit binary code, effectively eliminating gaps and addressing overhead in the data structure. After serialization, the cloud center performs multi-level encapsulation on this exposed binary data stream according to the basic communication protocol specifications. Specifically, a control frame header containing the device's physical address and timestamp information is appended to the head of the binary stream, and a cyclic redundancy check code for integrity verification is appended to the tail. This standardized assembly forms the communication message to be sent out, which at this point has the ability to be independently addressed and reliably traversed in a wide area network or local area network.

[0036] After the communication message is constructed, the network control node in the cloud center immediately activates the pre-established downlink communication link. This downlink communication link can be built based on 5G mobile communication technology or a dedicated fiber optic line, providing extremely low transmission latency and extremely high bandwidth capacity. The routing module in the cloud center sends the encapsulated communication message along this downlink communication link to the designated target edge node using a high-priority scheduling strategy. During this transmission phase, the network routing device plans an optimal physical transmission route for the message based on the physical address identifier in the message frame header within the complex network topology. This system prevents data packets from getting stuck or lost at congested nodes. Meanwhile, the target edge node at the receiving end constantly polls and monitors the network interface. The target edge node opens specific communication listening ports, such as configuring and continuously listening for incoming data requests on port 8080, to be ready to receive instructions and parameters from the cloud. When the sent communication packet arrives at the physical network card, the underlying driver of the target edge node triggers a hardware interrupt, and uses direct memory access technology to quickly move the sent communication packet to the receive buffer in the operating memory, thus successfully completing the cross-level reception and storage task of the packet.

[0037] After receiving the transmitted communication message in its receive buffer, the target edge node must ensure the absolute security and accuracy of its core operational data. Therefore, its built-in parsing engine performs deep data parsing and rigorous checksum comparison on the received transmitted communication message. The parsing engine first strips away the network encapsulation layer outside the message, reads the 16-bit cyclic redundancy check (CRC) code at the end, and recalculates the checksum of the main binary stream using the same generator polynomial. The recalculated result is then compared bit by bit with the checksum carried in the message. If a discrepancy is found, the edge node will discard the message and send a retransmission request to the cloud. After confirming that the message is correct and its source is legitimate, the decoding module of the target edge node officially takes over the data stream and reverses the deserialization decoding operation corresponding to the cloud serialization. This decoding mechanism re-cuts the continuous binary data stream and converts it into floating-point format according to the pre-agreed byte alignment rules, accurately restoring the original data hierarchy and logical dimension, thereby extracting and restoring the network parameter tensor originally generated by the hypernetic network from the message without loss. For example, it perfectly restores the 8192-byte binary stream into the aforementioned combination of 2048 single-precision floating-point values, preparing the core material for the self-evolution of the edge prediction algorithm.

[0038] After the target edge node reads the reconstructed network parameter tensor from the memory stack, it enters the most critical stage of parameter implementation. The lightweight prediction network running inside the target edge node occupies a fixed, contiguous physical address space in memory, and the weight variables of each computational layer are bound to unique memory pointers. The update program strictly follows the topological order of the network hierarchy, for example, from the bottom-most feature input layer to the hidden mapping layer and then to the top-most risk assessment output layer, traversing the tensor values ​​sequentially. During the traversal, the edge node calls the underlying memory write function, bypassing the conventional file storage and application restart process, to write the tensor values. The values ​​are directly replaced and overwritten into the memory address of the lightweight prediction network at the edge. Assuming that the starting memory address corresponding to a certain hidden layer weight matrix is ​​0x1000 in hexadecimal, the update program will use pointer offset calculation to write the newly extracted tensor values ​​one by one into the memory block area starting from this address, forcibly overwriting the old parameters of the previous cycle. Through this direct address overwriting mechanism based on memory level, the target edge node realizes dynamic loading and updating within a millisecond-level time window, enabling the prediction model to instantly switch to the parameter state with the latest topology awareness capability without interrupting the real-time energy flow monitoring task.

[0039] In one embodiment of the present invention, in step S6, the process by which the target edge node inputs its current state sequence into a lightweight prediction network loaded with the latest parameters for forward inference, and outputs the edge node's energy flow state prediction and limit exceedance risk assessment value is as follows: After confirming that the network parameter tensor in the memory address has been loaded, the target edge node immediately starts its local lightweight prediction network to perform forward inference. At this time, the data reading module of the target edge node extracts the current state sequence from the cache queue. Assume this sequence is a time-series feature matrix containing the past 10 sampling times with a total dimension of 64. The target edge node precisely feeds this 64-dimensional current state sequence into the input layer of the lightweight prediction network, which has completed parameter loading. The data flow then enters the deep computation stage within the network. Inside the network, the computational units strictly execute intensive matrix multiplication and activation mapping forward inference operations sequentially along the network's internal hierarchy. Specifically… The 64-dimensional input features are first multiplied by the latest weight matrix of size 64 by 128 in the first hidden layer, which expands the original low-dimensional physical features to a 128-dimensional high-order space. Then, these 128 calculation results are fed into a non-linear activation function for truncation and filtering, for example, all negative values ​​less than 0 are forced to be set to 0 to filter out invalid interference signals. Through this continuous and alternating linear transformation and non-linear activation filtering between multiple network layers, the original electrical features are fully deconstructed and refined. Finally, a hidden state feature vector with a fixed dimension of 256 is successfully generated in the deepest feature extraction layer. This vector highly condenses the inherent non-linear laws of energy flow evolution.

[0040] With the successful completion of the deep feature extraction stage, the data flow of the forward inference officially enters the final processing stage of the lightweight prediction network. The state output layer of the lightweight prediction network immediately receives the 256-dimensional hidden state feature vector generated in the above steps. Unlike the hidden layers, which are mainly responsible for nonlinear exploration of the high-dimensional feature space, the core task of this state output layer is to reproject these abstract high-dimensional features back into the space of directly observable physical quantities. Therefore, the state output layer uses its own internal dedicated output weight matrix to perform precise linear dimensionality reduction mapping calculations on this vector. Assuming that the weight matrix stored inside this output layer has a size of 256 rows and 1 column, the computation unit projects these 256-dimensional hidden state features into the... The quantity is then subjected to a standard vector inner product operation with the weight matrix. In this purely linear operation, the feature weights carried by the 256 hidden nodes are fused and summed according to a specific ratio, ultimately collapsing directly into a single scalar value. The entire process does not introduce any nonlinear activation functions with truncation or compression properties to ensure that the predicted value has complete physical dimensions and a continuous gradient. After this deep numerical integration and linear reconstruction operation, the state output layer is successfully reconstructed and finally outputs the predicted energy flow state value of the target edge node at a specific time in the future to an external register. For example, it accurately predicts that the real-time current value carried by the physical device in the next 5 minutes will be 115.5 amperes.

[0041] After the core processing unit of the target edge node reads the newly generated energy flow state prediction value from the register, it needs to quantitatively evaluate the operational health status of the equipment. Pre-defined safe operating boundaries, established by electrical experts, are pre-programmed into the edge node's memory; for example, the upper limit of the safe current for this type of switchgear is set at 120.0 amps. The target edge node then performs a rigorous one-way difference comparison operation between the read prediction value (e.g., 115.5 amps as mentioned above) and the pre-defined safe operating boundary. Specifically, the current prediction value is subtracted from the pre-defined safe boundary value. In this example, the resulting one-way difference is 4.5 amps, indicating that the equipment has a 4.5 amp margin before overload failure. Space; if the predicted value reaches 125.0 amperes, the resulting one-way difference will be negative, indicating a serious substantial exceedance of the limit; after obtaining this difference result, the calculation engine will input the deviation result representing the degree to which the predicted value approaches or exceeds the safe operating boundary into a special risk mapping function; this function, according to a specific exponential decay or linear mapping rule, transforms the absolute physical difference into a relative risk index. For example, the above 4.5 ampere margin is mapped and output as a risk score of 85. The higher the score, the more dangerous it is. Finally, the accurate mapping output is the exceedance risk assessment value of the target edge node, providing reliable data reference for the automatic cut-off protection device on site.

[0042] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A power system load forecasting method based on edge computing, characterized in that, Includes the following steps: S1. Edge nodes collect local real-time operating status sequences, extract local status codes through the local mapping layer, and upload them to the cloud center. S2. The cloud center collects the local status codes of all edge nodes and dynamically reconstructs the global topology feature map based on the spatial coordinates of each node and the real-time uploaded physical switch status. S3. Input the dynamically reconstructed global topology feature map into the cloud graph neural network for feature aggregation calculation, and output customized features of the target edge nodes with the latest spatial coupling perception capabilities. S4. Input the node-customized features into the cloud-based hypernetwork model for matrix mapping calculation to generate a target edge node network parameter tensor that adapts to the current transient topology. S5. The cloud center sends the network parameter tensor to the target edge node and dynamically loads and updates it into the lightweight prediction network on the edge. S6. The target edge node inputs the current state sequence into the lightweight prediction network loaded with the latest parameters for forward inference, and outputs the energy flow state prediction and limit risk assessment value of the edge node.

2. The power system load forecasting method based on edge computing according to claim 1, characterized in that, In step S1, the process of the edge node collecting local real-time operating status sequences, extracting local status codes through the local mapping layer, and uploading them to the cloud center is as follows: Edge nodes collect real-time voltage amplitude, current phase, and physical switch status of their respective areas. The acquired basic electrical characteristic data are then arranged and spliced ​​in chronological order of sampling time to form a continuous local real-time operating status sequence. Edge nodes input the local real-time running state sequence into the local mapping layer, and use the neural network structure inside the local mapping layer to extract the nonlinear correlation features inside the sequence, and generate the local state code through matrix linear transformation mapping calculation; Edge nodes acquire the generated local status code, attach the corresponding node's location coordinates and unique device identification code to the code data, encapsulate it into a transmission message according to the standard communication protocol format, and upload the message to the cloud center through a dedicated communication link.

3. The power system load forecasting method based on edge computing according to claim 1, characterized in that, In step S2, the cloud center collects the local state codes of all edge nodes and dynamically reconstructs the global topology feature map based on the spatial coordinates of each node and the real-time uploaded physical switch status. The cloud center receives all messages uploaded by edge nodes, performs data parsing on the messages, and extracts the local status code, node spatial coordinates and physical switch status of the corresponding edge nodes. The cloud center maps the spatial coordinates of nodes to the node elements of the graph structure, and constructs the initial edges that reflect the adjacency relationship of the physical structure based on the inherent spatial coordinate connection relationship between the node elements. The cloud center reads the extracted physical switch status, performs connectivity determination on the initial edges based on the closed and open attributes of the physical switch status, and removes the initial edges corresponding to the open attribute to generate a dynamic topology; The cloud center integrates all local state codes into the node elements of the dynamic topology, completes the splicing of node feature data and dynamic connectivity attributes, and dynamically reconstructs a global topology feature map.

4. The power system load forecasting method based on edge computing according to claim 1, characterized in that, In step S3, the process of inputting the dynamically reconstructed global topology feature map into the cloud graph neural network for feature aggregation calculation and outputting customized features of the target edge nodes with the latest spatial coupling perception capability is as follows: The cloud center inputs the dynamically reconstructed global topology feature map into the cloud graph neural network. Based on the dynamic connectivity properties inside the global topology feature map, the cloud graph neural network identifies and extracts the node feature data of all adjacent nodes directly connected to the target edge node. The cloud-based graph neural network uses an internally defined weight matrix to perform linear mapping calculations on the extracted node feature data of each adjacent node, generating corresponding transmitted feature vectors, and then transmitting the transmitted feature vectors along the connected edge structure to the target edge node. The cloud-based graph neural network calls the feature aggregation function to receive all the transmitted feature vectors that are gathered to the target edge node, performs pooling summation on all the received transmitted feature vectors, and generates a neighborhood aggregation feature matrix that integrates the surrounding physical topology information. The cloud-based graph neural network concatenates the generated neighborhood aggregation feature matrix with the original node feature data of the target edge nodes, inputs the concatenated data into a nonlinear activation function for update calculation, and outputs customized features of the target edge nodes with the latest spatial coupling perception capabilities.

5. The power system load forecasting method based on edge computing according to claim 1, characterized in that, In step S4, the process of inputting the node-customized features into the cloud-based hypernetwork model for matrix mapping calculation to generate the target edge node network parameter tensor adapted to the current transient topology is as follows: The cloud center inputs the customized features of the target edge nodes into the basic mapping layer of the cloud hypernetwork model. The basic mapping layer performs nonlinear spatial transformation calculations on the customized features of the target edge nodes and outputs intermediate state vectors. The weight generation module of the cloud-based hypernetwork model receives the intermediate state vector and performs continuous matrix multiplication mapping operations on the intermediate state vector based on the network layer dimension parameters of the lightweight prediction network to generate a one-dimensional parameter sequence. The tensor reconstruction module of the cloud-based super network model obtains all one-dimensional parameter sequences, performs shape reshaping operations on the one-dimensional parameter sequences according to the neuron connection specifications inside the lightweight prediction network, and constructs a multi-dimensional weight matrix. The cloud-based hypernetwork model performs a concatenation and combination operation on all multidimensional weight matrices in hierarchical order to generate a set of network layer parameters, and finally outputs a target edge node network parameter tensor that adapts to the current transient topology.

6. The power system load forecasting method based on edge computing according to claim 5, characterized in that, The specific content of performing continuous matrix multiplication mapping operations on the intermediate state vector to generate a one-dimensional parameter sequence is as follows: The weight generation module extracts the network layer dimension parameters of the lightweight prediction network and constructs multiple concatenated linear transformation matrices with corresponding row and column dimensions based on these network layer dimension parameters. The weight generation module inputs the received intermediate state vector into the first linear transformation matrix for calculation, and passes the output result of each step to the next linear transformation matrix to perform continuous matrix multiplication mapping operations. The weight generation module extracts the final mapping features calculated from the final linear transformation matrix, performs a data flattening transformation operation on the final mapping features, and generates a one-dimensional parameter sequence by combining and arranging them according to their inherent order.

7. The power system load forecasting method based on edge computing according to claim 1, characterized in that, In step S5, the process by which the cloud center sends the network parameter tensor to the target edge node and dynamically loads and updates it into the lightweight prediction network locally at the edge is as follows: The cloud center obtains the network parameter tensor of the target edge node, performs serialization encoding on the network parameter tensor to convert it into a binary data stream, and encapsulates it into a communication message according to the basic communication protocol specification. The cloud center activates and establishes a downlink communication link, and sends the communication message along the downlink communication link to the target edge node. The target edge node receives the communication message through its listening port. The target edge node performs data parsing and verification comparison on the received communication messages. After confirming that the messages are correct, it performs deserialization decoding to extract and restore the network parameter tensor from the messages. The target edge node reads the restored network parameter tensor and, according to the network hierarchy, directly replaces and overwrites the tensor value into the memory address of the lightweight prediction network on the edge, performing dynamic loading and updating.

8. The power system load forecasting method based on edge computing according to claim 1, characterized in that, In step S6, the process by which the target edge node inputs its current state sequence into a lightweight prediction network loaded with the latest parameters for forward inference, and outputs the edge node's energy flow state prediction and limit exceedance risk assessment value, is as follows: The target edge node feeds the current state sequence into the lightweight prediction network that has completed parameter loading, and performs matrix multiplication and activation mapping forward inference operations sequentially along the internal layers of the network to generate hidden state feature vectors. The state output layer of the lightweight prediction network receives the hidden state feature vector, performs linear dimensionality reduction mapping calculation on it using the internal weight matrix, reconstructs and outputs the energy flow state prediction value of the target edge node; The target edge node reads the predicted energy flow state value, performs a one-way difference comparison operation with the preset safe operation boundary, and maps the deviation result of the obtained predicted value that approaches or exceeds the safe operation boundary as the over-limit risk assessment value of the target edge node.

Images