Power network resource intelligent management method and system based on 6G and deep learning

Abstract

The invention relates to the technical field of power network resource management, and discloses a power network resource intelligent management method and system based on 6G and deep learning. The method comprises the following steps: acquiring node data through a 6G technology to obtain a topological mapping table, preprocessing the data to obtain a feature tensor, training a model cluster by adopting a space-time diagram convolution and long-short-term memory network hybrid architecture, analyzing a network state based on a model, generating a scheduling scheme, and issuing a control instruction through 6G network fragmentation to realize intelligent management of a power network. According to the invention, through the combination of the 6G communication positioning integration technology and the space-time diagram convolutional network and long-short term memory network mixed architecture, the real-time data acquisition precision and intelligent analysis capability of the power network are improved. Based on a collaborative optimization mechanism of multi-source data fusion and a deep learning model cluster, the comprehensiveness of power grid state evaluation and the accuracy of resource scheduling are enhanced.

Description

Technical Field

[0001] This application relates to the technical field of power grid resource management, and particularly to an intelligent management method and system for power grid resources based on 6G and deep learning. Background Art

[0002] At present, the management of power network resources mainly relies on traditional SCADA systems and power dispatching automation systems for monitoring and control. These systems collect operation data of power equipment through wired optical fibers or low-speed wireless networks, and use classical power flow calculation and economic dispatching algorithms for resource allocation. In terms of communication, 4G / 5G mobile communication technologies, power line carrier communication, and narrowband Internet of Things technologies are mainly adopted to provide limited data transmission capabilities. The positioning technology mainly relies on the GPS global positioning system or base station positioning technology, and the accuracy is generally in the meter range. At the decision-making level, traditional systems mainly adopt rule-based expert systems or simple statistical prediction models, rely on manual experience to construct decision rules, and lack the support of advanced intelligent algorithms such as deep learning. These systems can meet the basic needs in a relatively stable power grid environment and ensure the safe and stable operation of the power system.

[0003] However, with the continuous increase in the proportion of renewable energy, the large-scale access of distributed energy, and the enhanced volatility of electricity load, the traditional power network resource management methods are facing severe challenges. First, traditional communication technologies have problems such as limited bandwidth, high latency, and insufficient connection capacity, making it difficult to support the real-time data transmission of a large number of power equipment. Second, the accuracy of traditional positioning technologies is insufficient and cannot meet the requirements of refined resource management. Third, traditional data processing methods lack the ability to effectively utilize a large amount of heterogeneous data, and the data value is not fully mined. In addition, the traditional decision-making model has limited prediction ability for the power grid state in complex scenarios and is difficult to cope with the uncertainties brought by the high proportion of renewable energy access. Finally, traditional resource scheduling algorithms are difficult to achieve multi-objective optimization and it is difficult to balance economy, reliability, and environmental protection. These deficiencies seriously restrict the improvement of the power grid resource utilization efficiency and increase the operation risks of the power grid. Summary of the Invention

[0004] This application provides an intelligent management method and system for power grid resources based on 6G and deep learning, which improves the real-time data collection accuracy and intelligent analysis ability of the power network through the combination of 6G communication and positioning integration technology and the hybrid architecture of spatio-temporal graph convolutional network and long short-term memory network. Based on the collaborative optimization mechanism of multi-source data fusion and deep learning model clusters, the comprehensiveness of power grid state assessment and the accuracy of resource scheduling are enhanced.

[0005] In a first aspect, the present application provides an intelligent management method for power grid resources based on 6G and deep learning. The intelligent management method for power grid resources based on 6G and deep learning includes: collecting power network node data through 6G communication base stations and multi-source positioning sensors to obtain a power network topology physical coordinate mapping table, a real-time power parameter data stream, and communication network status information; performing data preprocessing on the power network topology physical coordinate mapping table, the real-time power parameter data stream, and the communication network status information to obtain a power data feature tensor; training a deep learning model cluster based on the power data feature tensor through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network to obtain a load prediction model, a state evaluation model, and a resource optimization model; collecting real-time power grid resource data, and performing power grid state analysis on the real-time power grid resource data according to the load prediction model, the state evaluation model, and the resource optimization model to obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction; calculating a resource scheduling plan according to the power grid state evaluation report, the fault risk distribution map, and the warning instruction to obtain a power generation equipment output plan, a charge-discharge strategy for an energy storage system, and a load switching plan; and sending control instructions through 6G network slicing according to the power generation equipment output plan, the charge-discharge strategy for the energy storage system, and the load switching plan to obtain power network operation state data and a performance evaluation report.

[0006] In a second aspect, the present application provides an intelligent management system for power grid resources based on 6G and deep learning. The intelligent management system for power grid resources based on 6G and deep learning includes: A collection module for collecting power network node data through 6G communication base stations and multi-source positioning sensors to obtain a power network topology physical coordinate mapping table, a real-time power parameter data stream, and communication network status information; A processing module for performing data preprocessing on the power network topology physical coordinate mapping table, the real-time power parameter data stream, and the communication network status information to obtain a power data feature tensor; A training module for training a deep learning model cluster based on the power data feature tensor through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network to obtain a load prediction model, a state evaluation model, and a resource optimization model; An analysis module for collecting real-time power grid resource data, and performing power grid state analysis on the real-time power grid resource data according to the load prediction model, the state evaluation model, and the resource optimization model to obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction; A calculation module for calculating a resource scheduling plan according to the power grid state evaluation report, the fault risk distribution map, and the warning instruction to obtain a power generation equipment output plan, a charge-discharge strategy for an energy storage system, and a load switching plan; A control module, configured to issue control instructions through 6G network slicing according to the power generation equipment output plan, the charge and discharge strategy of the energy storage system, and the load switching plan, and obtain the operation state data and performance evaluation report of the power grid.

[0007] The third aspect of the present invention provides a computer device, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor invokes the instructions in the memory to cause the computer device to execute the above-mentioned intelligent management method for power grid resources based on 6G and deep learning.

[0008] The fourth aspect of the present invention provides a computer-readable storage medium, in which instructions are stored, and when it runs on a computer, it causes the computer to execute the above-mentioned intelligent management method for power grid resources based on 6G and deep learning.

[0009] In the technical solution provided by this application, power grid node data is collected through 6G communication base stations and multi-source positioning sensors to obtain a power grid topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information, solving the problems of insufficient data collection accuracy and poor real-time performance in traditional technologies, and providing a high-quality data foundation for subsequent intelligent analysis. The technical feature of preprocessing the collected data to form a power data feature tensor effectively removes noise and redundancy in the original data, improves data quality, and enables the deep learning model to more accurately extract the key features of the power grid. The application of the hybrid architecture of the spatio-temporal graph convolutional neural network and the long short-term memory network fully considers the spatial topology structure and time series characteristics of the power grid. Compared with the traditional single-structure neural network, this hybrid architecture can simultaneously capture the spatial dependence relationship between power nodes and the time pattern of load changes, and the trained load prediction model, state evaluation model, and resource optimization model have higher accuracy and stronger generalization ability. The technical feature of collecting real-time resource data of the power grid and using the above models for power grid state analysis realizes the early identification and risk assessment of abnormal power grid states. Compared with the traditional threshold-based monitoring method, this deep learning-based state analysis can discover complex non-linear relationships and potential risk patterns, greatly improving the accuracy of power grid safety early warning. The technical feature of calculating a resource scheduling plan according to the power grid state evaluation result realizes the coordinated optimization of power generation, energy storage, and load. Compared with the traditional single-resource optimization, this multi-resource coordinated scheduling greatly improves the system flexibility and resource utilization efficiency. The technical feature of issuing control instructions through 6G network slicing ensures the ultra-reliable and low-latency transmission of key control instructions, solves the congestion and delay problems that may occur in traditional communication networks in emergency situations, and enables control instructions to be executed accurately and in a timely manner. Description of the Drawings

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0011] Figure 1 It is a schematic diagram of an embodiment of the intelligent management method for power grid resources based on 6G and deep learning in the embodiments of the present application; Figure 2 It is a schematic diagram of an embodiment of the intelligent management system for power grid resources based on 6G and deep learning in the embodiments of the present application; Figure 3 It is a schematic block diagram of the structure of a computer device in the embodiments of the present invention. Specific embodiments

[0012] The embodiments of the present application provide an intelligent management method and system for power grid resources based on 6G and deep learning. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned accompanying drawings of the present application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0013] For easy understanding, the following describes the specific process of the embodiments of the present application. Please refer to Figure 1 , an embodiment of the intelligent management method for power grid resources based on 6G and deep learning in the embodiments of the present application includes: Step S101: Collect power network node data through 6G communication base stations and multi-source positioning sensors to obtain a power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information; Step S102: Perform data preprocessing on the power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information to obtain a power data feature tensor; Step S103: According to the power data feature tensor, train a deep learning model cluster through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network to obtain a load prediction model, a state evaluation model, and a resource optimization model; Step S104: Collect real-time resource data of the power grid, and perform power grid state analysis on the real-time resource data of the power grid according to the load forecasting model, state evaluation model, and resource optimization model to obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction; Step S105: Calculate a resource scheduling plan based on the power grid state evaluation report, the fault risk distribution map, and the warning instruction to obtain a power generation equipment output plan, a charge and discharge strategy of the energy storage system, and a load switching plan; Step S106: According to the power generation equipment output plan, the charge and discharge strategy of the energy storage system, and the load switching plan, send control instructions in slices through the 6G network to obtain power network operation state data and a performance evaluation report.

[0014] It can be understood that the execution subject of this application can be a power grid resource intelligent management system based on 6G and deep learning, or a terminal or a server. Specifically, it is not limited here. In this embodiment of the application, the server is used as the execution subject for illustration.

[0015] Specifically, 6G communication base stations and multi-source positioning sensors are deployed at each node of the power grid to collect node data. Specifically, 6G communication modules installed at substations, key points of transmission lines, and distribution equipment measure the signal strength and transmission delay between nodes to generate a node communication connection matrix; intelligent power sensors collect voltage values, current values, active power, reactive power, and frequency values per second to form a real-time power parameter data stream; GPS receivers and base station-assisted positioning systems collect equipment positions and obtain centimeter-level spatial coordinates after Kalman filter processing; the node communication matrix is matched point-to-point with the equipment spatial coordinates to construct a power network topology physical coordinate mapping table; measure the signal-to-noise ratio, bit error rate, and available bandwidth of the 6G communication link to generate communication network state information. Subsequently, the collected data is preprocessed. The real-time power parameter data stream is decomposed by wavelet transform into high-frequency noise components and low-frequency trend components to form a power signal component matrix; the autoencoder detects and repairs outliers and missing values in the matrix to generate a cleaned power parameter matrix; a node adjacency relationship graph is constructed according to the topology physical coordinate mapping table, and the connection weight values between nodes are calculated; the signal-to-noise ratio, bit error rate, and bandwidth data in the communication network state information are normalized to form a communication state feature vector; the cleaned power parameter matrix is associated and fused with the node adjacency relationship graph and the communication state feature vector to construct a multi-dimensional association matrix; the matrix is tensor decomposed by the Tucker decomposition algorithm, and the rank of the time-dimensional core tensor is set to 24, the rank of the space-dimensional core tensor is 30% of the number of network nodes, and the rank of the physical characteristic-dimensional core tensor is 8 to obtain a power data feature tensor.

[0016] Based on the power data feature tensor, train a cluster of deep learning models. Divide the feature tensor into a training set and a validation set along the time dimension, establish a power network node relationship graph, and construct a model training data set; the spatio-temporal graph convolutional network processes this data set, extracts the spatial dependence features between power nodes, and generates a spatial correlation representation of power nodes; the long short-term memory network receives this representation, captures the temporal variation law of power load, and obtains the time correlation features; based on the time correlation features and historical load data, construct a load prediction model; based on the spatial correlation representation of power nodes and equipment status parameters, train a Bayesian neural network to quantify the uncertainty of the power grid state and form a state evaluation model; use the prediction results of the load prediction model and the evaluation results of the state evaluation model to train a deep reinforcement learning network, optimize the resource allocation strategy, and obtain a resource optimization model. Collect real-time resource data of the power grid and conduct power grid state analysis. The distributed sensor network collects the real-time operating parameters of each node, including load level, equipment temperature, transmission line impedance, and phase angle difference; the load prediction model calculates the predicted values of the load distribution for the next 24 hours and generates a load prediction result map; the state evaluation model performs Bayesian probability inference on the real-time resource data and the load prediction results, calculates the stability margin index and fault probability value of each node, and constructs a power grid stability evaluation matrix; conduct threshold analysis on this matrix, identify high-risk nodes, and generate a risk level classification table; map the risk level classification table onto the power network topology map to form a fault risk distribution map; according to the fault risk distribution map and the risk propagation evaluation results, generate a power grid state evaluation report and a warning instruction.

[0017] Calculate the resource scheduling plan according to the power grid state evaluation results. Extract the system key stability indicators and load prediction data from the power grid state evaluation report to construct a set of resource scheduling constraint conditions; according to the high-risk areas in the fault risk distribution map, calculate the risk coefficient and emergency power supply demand of each area to form a power supply priority table for risk areas; based on the fault type information and time estimate in the warning instruction, formulate a power grid scheduling time table and generate a scheduling cycle matrix; calculate the optimal output allocation ratio for various types of power generation equipment to generate a power generation equipment output plan; calculate the power gap and surplus situation, allocate charge and discharge periods and power values for the energy storage system to form an energy storage system charge and discharge strategy; determine the load shedding nodes and switching order to generate a load switching plan.

[0018] The control instructions are sent and the feedback data is obtained through 6G network slicing. The scheduling scheme is converted into a standardized control instruction format to generate instruction data packets; the instruction data packets are marked with priorities and divided into three levels: critical control instructions, regular scheduling instructions, and auxiliary management instructions; the communication resources are dynamically allocated through 6G network slicing technology to create different types of network slices; the distributed consensus algorithm is used to synchronize the control instructions among nodes to ensure the time synchronization and logical consistency of multi-node control; the synchronized control instruction set is sent to the control execution unit through their respective network slices, and the execution feedback data is collected to form the power network operation status data; the real-time operation parameters of the power network are compared and analyzed in multiple dimensions, the parameter deviation values and fluctuation curves are recorded, and the execution situation of the control instructions is correlated to form a performance evaluation report.

[0019] For example, when abnormal voltage fluctuations are detected at a certain power network node, through the spatio-temporal correlation analysis in the power data feature tensor, the state evaluation model calculates that the failure probability of this node is 78% and it is marked as a high-risk area. Based on this, the resource scheduling scheme configures the adjacent energy storage system to the discharge mode with a power setting of 2MW, and at the same time prepares 5MW of load that can be shed as an emergency measure. These control instructions are sent through the 6G ultra-reliable low-latency network slice to ensure that the instruction transmission latency is less than 5 milliseconds, effectively preventing the spread of power grid faults.

[0020] In the embodiments of the present application, 6G communication base stations and multi-source positioning sensors are used to collect power network node data, obtaining a power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information, which solves the problems of insufficient data collection accuracy and poor real-time performance in traditional technologies and provides a high-quality data foundation for subsequent intelligent analysis. The technical feature of preprocessing the collected data to form a power data feature tensor effectively removes noise and redundancy in the original data, improves data quality, and enables the deep learning model to more accurately extract the key features of the power network. The application of the hybrid architecture of the spatio-temporal graph convolutional neural network and the long short-term memory network fully considers the spatial topology structure and time series characteristics of the power network. Compared with the traditional single-structure neural network, this hybrid architecture can simultaneously capture the spatial dependence relationship between power nodes and the time pattern of load changes, and the trained load prediction model, state evaluation model, and resource optimization model have higher accuracy and stronger generalization ability. The technical feature of collecting real-time resource data of the power grid and using the above models to analyze the grid state realizes the early identification and risk assessment of abnormal grid states. Compared with the traditional threshold-based monitoring method, this deep learning-based state analysis can discover complex non-linear relationships and potential risk patterns, greatly improving the accuracy of grid safety early warning. The technical feature of calculating the resource scheduling scheme according to the grid state evaluation results realizes the coordinated optimization of power generation, energy storage, and load. Compared with traditional single-resource optimization, this multi-resource coordinated scheduling greatly improves the system flexibility and resource utilization efficiency. The technical feature of sending control instructions through 6G network slicing ensures the ultra-reliable and low-latency transmission of critical control instructions, solves the congestion and delay problems that may occur in traditional communication networks in emergency situations, and enables control instructions to be executed accurately and in a timely manner.

[0021] In a specific embodiment, the process of executing step S101 may specifically include the following steps: (1) Deploy 6G communication modules at key points of each substation, transmission line, and distribution equipment in the power network, measure the signal strength and transmission delay between each node, and generate a node communication connection matrix; (2) Collect the voltage value, current value, active power, reactive power, and frequency value of each node in the power network through intelligent power sensors every second to form a real-time power parameter data stream; (3) Use a GPS receiver and a base station-assisted positioning system to collect the accurate position of the equipment, and process the original positioning data using the Kalman filtering algorithm to obtain centimeter-level equipment spatial coordinates; (4) Perform point-to-point matching on the link state data in the node communication connection matrix and the equipment spatial coordinates to construct a power network topology physical coordinate mapping table including physical distance and link quality; (5) By measuring the signal-to-noise ratio, bit error rate, and available bandwidth of the 6G communication link, record the real-time state parameters of each communication channel to generate communication network status information; (6) Perform timestamp marking and alignment on the data points in the real-time power parameter data stream, eliminate the sampling time differences of different devices, and complete the data synchronization process of the power network nodes.

[0022] Specifically, deploy 6G communication modules at key positions in the power network. These modules are installed at substations, key points of transmission lines, and distribution equipment to form a grid coverage. Each communication module is equipped with millimeter-wave and terahertz wave transceivers, capable of achieving ultra-high-speed data transmission. The communication modules measure the signal strength between each other. Specifically, they send detection signals and record the received power, and at the same time calculate the round-trip time of the signal to obtain the transmission delay. Convert the received power into a link quality index through the signal strength attenuation model to generate a connection status score for each pair of nodes. These scores are organized in matrix form, where the matrix elements represent the connection quality between nodes, thus forming a node communication connection matrix.

[0023] At the same time, intelligent power sensors perform high-frequency data acquisition at each power node. These sensors include devices such as intelligent electricity meters, current transformers, and voltage transformers, equipped with high-precision analog-to-digital converters, and obtain power parameters at a sampling rate per second. The collected data includes line voltage values (unit: volts), current values (unit: amperes), active power (unit: watts), reactive power (unit: vars), and frequency values (unit: hertz). These data are timestamp-marked and packed into a data stream format, and are transmitted to the data processing center in real time through 6G communication modules to form a real-time power parameter data stream.

[0024] For the precise positioning of the device location, multi-source positioning technology is adopted. Each power device is installed with a GPS receiver to receive satellite positioning signals to obtain initial position data. At the same time, utilize the positioning signals transmitted by nearby 6G base stations to calculate the distance between the device and the base station through the time difference of arrival algorithm. The original positioning data contains noise and errors, so the Kalman filtering algorithm is applied for processing. This algorithm consists of two stages: prediction and correction. In the prediction stage, the current state is predicted based on the previous state and the system model; in the correction stage, the prediction is corrected by combining the observation data. The Kalman filter recursively estimates the true position of the device, filters out random fluctuations, and finally precisely locates to the centimeter level.

[0025] After obtaining the node communication connection matrix and the device spatial coordinates, point-to-point matching processing needs to be carried out. First, assign a unique identifier to each communication node, and at the same time create a corresponding index for each device location coordinate. Associate the two data sets through the identifier, and integrate the communication quality data and the physical distance information. Calculate the Euclidean distance between any two points to represent the physical interval between nodes. At the same time, retain the link communication quality indicators, including the bandwidth capacity and the stability score. These data are comprehensively organized into a power network topology physical coordinate mapping table, which includes node identifiers, geographical coordinates, a list of connected nodes, and the corresponding link quality indicators.

[0026] In order to obtain the communication network status information, real-time monitoring of the 6G communication links is required. A dedicated network analyzer continuously measures the key parameters of each link: the signal-to-noise ratio (in decibels) reflects the anti-interference ability; the bit error rate represents the proportion of error bits in a unit bit; the available bandwidth (in megabits per second) represents the data transmission capacity. These measured values are processed by statistical methods to calculate the average value, peak value, and standard deviation, and generate a status description of each communication channel. These descriptions are organized into structured data to form the complete communication network status information. Perform time synchronization processing on the collected real-time power parameter data stream. Due to the slight differences in the sampling clocks of different devices, the sampling times are not exactly the same. First, add an accurate timestamp to each data point to mark the sampling moment. Then set a unified time reference, usually choosing Coordinated Universal Time (UTC) as the reference. Use the linear interpolation method to align the data at different time points to a unified time axis to eliminate the sampling time difference. The data after alignment processing is reorganized into a data stream in a unified format to complete the power network node data synchronization processing.

[0027] Taking a certain substation as an example, the 6G communication module deployed measures the signal strength with an adjacent line node as -70 dBm and the transmission delay as 2 milliseconds. Accordingly, record the connection quality as 0.85 (with a full score of 1) in the node communication connection matrix. At the same time, the intelligent power sensor collects the voltage of the main transformer of this substation as 220 kV, the current as 500 A, and the active power as 105 MW. The original position data obtained by the GPS receiver shows that the latitude deviation reaches 1.2 meters. After Kalman filtering, the positioning accuracy is improved to 8 centimeters. Integrate this information into the power network topology physical coordinate mapping table, and record the physical distance and communication quality between this substation and adjacent nodes. The communication link measurement shows that the signal-to-noise ratio is 21 dB, the bit error rate is 10^-6, and the available bandwidth is 1.2 Gbps. When it is found that there is a 15-millisecond deviation in the sampling time between different devices, through timestamp alignment processing, the data is unified to the reference time point to ensure the data synchronization for subsequent analysis.

[0028] In a specific embodiment, the process of executing step S102 may specifically include the following steps: (1) Perform wavelet transform decomposition on the real-time power parameter data stream to separate the high-frequency noise components and low-frequency trend components in the power signal, forming a power signal component matrix; (2) Use an autoencoder to detect and repair outliers and missing values in the power signal component matrix, generating a cleaned power parameter matrix; (3) Construct a node adjacency relationship graph according to the power network topology physical coordinate mapping table and calculate the connection weight values between nodes; (4) Normalize the signal-to-noise ratio, bit error rate, and bandwidth data in the communication network status information to form a communication status feature vector; (5) Correlate and fuse the cleaned power parameter matrix with the node adjacency relationship graph and the communication status feature vector to construct a multi-dimensional correlation matrix; (6) Perform tensor decomposition on the multi-dimensional correlation matrix through the Tucker decomposition algorithm. Set the core tensor rank of the time dimension to 24, the core tensor rank of the space dimension to 30% of the number of network nodes, and the core tensor rank of the physical property dimension to 8, and perform multi-linear feature extraction to obtain a power data feature tensor.

[0029] Specifically, when performing wavelet transform decomposition on the real-time power parameter data stream, the discrete wavelet transform (DWT) technology is adopted. Wavelet transform has the ability of multi-resolution analysis and decomposes the signal through low-pass and high-pass filters. The specific operation is to perform multi-level decomposition on power time series signals such as voltage, current, and power through the Daubechies wavelet basis (db4), usually performing 5-level decomposition. Each level of decomposition divides the signal into approximation coefficients (low-frequency part) and detail coefficients (high-frequency part). The low-frequency part reflects the overall trend of the signal and contains the basic change patterns of the power load; the high-frequency part contains short-term fluctuations, measurement noise, and interference information. The coefficients after decomposition are organized in matrix form, with each row representing a power parameter and each column representing different frequency components, forming a power signal component matrix.

[0030] Use an autoencoder to detect outliers and repair missing values in the power signal component matrix. An autoencoder is a neural network structure composed of an encoder and a decoder. The encoder compresses the input data into a low-dimensional latent space, and the decoder attempts to reconstruct the original data from the latent space. Normal data points can be reconstructed well, while outliers usually have a large reconstruction error. The specific operation is to segment the power signal component matrix by time window and input it into the autoencoder, and calculate the reconstruction error. If the reconstruction error of a data point exceeds a set threshold (usually set to 3 times the standard deviation of the average reconstruction error), it is marked as an outlier. For missing values, a method combining forward filling and values generated by the autoencoder is used for repair. For outliers, they are replaced with the reconstructed values of the autoencoder. The processed data is reorganized into a matrix form to obtain a cleaned power parameter matrix. Constructing a node adjacency relationship graph according to the power network topology physical coordinate mapping table is a key step. The power network can be regarded as a graph structure, where substations, key points of transmission lines, and distribution equipment are used as nodes, and the physical connection relationship is used as edges. The calculation of the connection weight value involves two factors: physical distance and communication quality, and can be expressed by the following formula: where, represents the connection weight value between node and node , and its value range is [0,1]; represents the physical Euclidean distance between nodes; is the distance attenuation parameter, which controls the influence degree of distance on the weight; represents the communication link quality, which is obtained by comprehensively evaluating the signal strength and delay; is the balance parameter, which controls the relative contribution of physical distance and communication quality to the weight, and its value range is [0,1]. A larger weight value indicates a stronger correlation between nodes. The calculated weight values are stored in the adjacency matrix to complete the construction of the node adjacency relationship graph.

[0031] Normalizing the communication network status information is a necessary step to ensure the comparability of different dimension indicators. The original communication status information includes parameters such as signal-to-noise ratio (unit: dB), bit error rate (dimensionless), and bandwidth (unit: bps), and the numerical ranges vary greatly. The Z-score standardization method is used to process each parameter: where, is the standardized parameter value, is the original parameter value, is the mean value of this parameter, is the standard deviation. After standardization, the mean of each parameter is 0 and the standard deviation is 1, and they are at the same order of magnitude. The standardized signal-to-noise ratio, bit error rate (taking negative values so that smaller values represent better performance), and bandwidth values are concatenated into a vector to form a communication state feature vector with a unified dimension.

[0032] The cleaned power parameter matrix is associated and fused with the node adjacency relationship graph and the communication state feature vector to construct a multi-dimensional association matrix. First, the topological structure information of the node adjacency relationship graph is combined with the power parameter matrix to construct a graph-enhanced power parameter representation. The specific operation is to perform graph convolution operations to weight and mix the power parameters of each node with the parameters of its neighbor nodes: Among them, is the enhanced power parameter representation, is the adjacency matrix, is the degree matrix (the diagonal elements are the number of edges of each node), is the cleaned power parameter matrix, is the learnable parameter matrix, is the non-linear activation function. Then, the communication state feature vector is used as an additional feature and concatenated with the graph-enhanced power parameter representation to form a multi-dimensional association matrix including time, space, and physical characteristics.

[0033] Finally, the multi-dimensional association matrix is tensor decomposed by the Tucker decomposition algorithm. Tucker decomposition is a high-order singular value decomposition that decomposes a high-dimensional tensor into the product of a core tensor and multiple factor matrices: Among them, represents the original multi-dimensional association matrix, represents the core tensor, , , respectively represent the factor matrices in the dimensions of time, space, and physical characteristics, represents the tensor-matrix product along the v-th dimension. Considering the characteristics of the power network, the rank of the core tensor in the time dimension is set to 24 (corresponding to the 24-hour periodicity of a day), the rank of the core tensor in the space dimension is set to 30% of the number of network nodes (retaining the main spatial patterns), and the rank of the core tensor in the physical characteristics dimension is set to 8 (covering the key power parameter characteristics). By selecting appropriate ranks of the core tensor, data dimensionality reduction and feature extraction are achieved, and a compact and information-rich power data feature tensor is obtained.

[0034] In a specific embodiment, the process of executing step S103 may specifically include the following steps: (1) Divide the power data feature tensor into a training set and a validation set along the time dimension, establish a power network node relationship graph, and construct a model training data set; (2) Apply a spatio-temporal graph convolutional network to the model training data set, extract the spatial dependence features between power nodes, and generate a spatial correlation representation of power nodes; (3) Input the spatial correlation representation of power nodes into a long short-term memory network to capture the temporal variation law of power load and obtain temporal correlation features; (4) Construct a load prediction model based on the temporal correlation features and historical load data, optimize the loss function through the gradient descent algorithm, and adjust the model parameters until the prediction error converges; (5) Train a Bayesian neural network based on the spatial correlation representation of power nodes and equipment status parameters, quantify the uncertainty of the power grid state, and form a state evaluation model; (6) Use the prediction results of the load prediction model and the evaluation results of the state evaluation model to train a deep reinforcement learning network, optimize the resource allocation strategy through a reward and punishment mechanism, and obtain a resource optimization model.

[0035] Specifically, for the division of the power data feature tensor, this feature tensor is a three-dimensional data structure obtained after preprocessing power network data, including a time dimension (recording data at different time points), a spatial dimension (recording data at different nodes), and a feature dimension (recording different power parameters). The data division uses a sliding window method, taking the data for a continuous period as input to predict the data for the subsequent period. Usually, the input window value ranges from 7 to 30 days, and the prediction window value ranges from 1 to 7 days. The specific operation is to arrange the entire data in chronological order, taking the samples in the first 80% of the time period as the training set and the last 20% as the validation set. For example, for 90 days of historical data, the first 72 days of data enter the training set, and the last 18 days enter the validation set. This division method maintains the continuity of the time series and avoids the leakage of time information that may be caused by random sampling.

[0036] At the same time, based on the power network topology physical coordinate mapping table obtained in the preprocessing stage, the power network is represented as a graph structure, including a node set (substations, key points of transmission lines, distribution equipment, etc.) and an edge set (representing the physical connections between nodes). Each node contains a multi-dimensional attribute vector recording parameters such as the voltage, current, and power of this node; each edge contains a weight value representing the connection strength between nodes. The weight value of the edge comes from the connection weight value calculated in the aforementioned data preprocessing stage, comprehensively considering the physical distance and communication quality. The construction process of the training data set is to combine the node features in the power data feature tensor with the topological structure of the node relationship graph to form a complete data set with spatio-temporal association. Each training sample contains time series information and graph structure information, providing a basis for subsequent spatio-temporal joint modeling.

[0037] Apply the spatio-temporal graph convolutional network to the model training dataset, which is a deep learning architecture specifically designed to process graph-structured data with spatio-temporal characteristics. The network structure consists of multiple spatio-temporal convolutional blocks, each of which is composed of a spatial graph convolutional layer and a temporal convolutional layer. The spatial graph convolutional layer aggregates the feature information of each node and its neighboring nodes through the convolutional operation defined on the graph. Specifically, for each node, its first-order neighbor nodes are first determined through the adjacency matrix, and then the features of the node are weighted and aggregated with the features of its neighbor nodes, and the weights are determined by the connection strength. In this way, physically or electrically close nodes can influence each other to form a feature representation considering the spatial structure. Through multiple layers of graph convolutional operations, information can be transmitted over a larger range to capture the multi-order dependence relationships between nodes. The temporal convolutional layer then uses a one-dimensional convolutional sliding window to process the feature sequences of each node in the time dimension to capture short-term time patterns. The spatial graph convolution and the temporal convolution are alternated to extract and fuse the spatial dependence features and the temporal dependence features layer by layer, and finally generate the spatial correlation representation of power nodes, which is a feature representation containing the mutual influence relationships between nodes. Input the spatial correlation representation of power nodes into the long short-term memory network to further process the long-term dependence relationships in the time series. The long short-term memory network is a special recurrent neural network designed to solve the vanishing gradient problem of ordinary recurrent networks when dealing with long sequences. The core of the network is the memory unit, which contains three control units: the input gate, the forget gate, and the output gate. The input gate controls the degree to which the current input information enters the memory unit; the forget gate determines how much historical memory to retain; the output gate controls the influence of the memory unit state on the current output. The gating mechanism is implemented by the sigmoid function, and the output value is between 0 and 1, indicating the proportion of information passing through. Power load data usually exhibits complex time characteristics, such as short-term fluctuations, intra-day cycles, intra-week patterns, and seasonal variations. The long short-term memory network can learn and remember patterns at different time scales through the memory unit and the gating mechanism. The specific process is to input the spatial correlation representation of power nodes into the network in chronological order. The network gradually processes the sequence information, retains important historical information through the memory mechanism, and adjusts the importance of information at different time points through the gating mechanism, and finally outputs a feature representation containing rich time information, that is, the time correlation feature.

[0038] Construct a load forecasting model based on time correlation features and historical load data. The model adopts a multi-layer feedforward neural network structure. The input layer receives time correlation features and processed historical load data (such as statistical features of different time windows). The network contains multiple hidden layers, and each layer consists of a fully connected layer and a non-linear activation function. Commonly used activation functions include the rectified linear unit (ReLU), which can effectively alleviate the problem of gradient vanishing; the output layer usually does not use an activation function or uses the identity function to directly output the predicted value. The model training uses the gradient descent algorithm to optimize the mean squared error loss function. The training process includes four main steps: 1) Forward propagation to calculate the predicted value, passing the input data through each layer of the network to obtain the output predicted value; 2) Calculate the loss, comparing the predicted value with the actual value to calculate the mean squared error; 3) Backward propagation to calculate the gradient, calculating the parameter gradient layer by layer from the output layer to the input layer; 4) Parameter update, adjusting the network parameters according to the gradient direction. To improve the training efficiency, the mini-batch gradient descent method is usually adopted, randomly selecting a small batch of samples for training each time. The optimizer selects the Adam algorithm, combining the momentum method and the adaptive learning rate method to accelerate the convergence process. To prevent overfitting, regularization techniques are introduced, such as weight decay (adding a parameter norm penalty term to the loss function) and early stopping strategy (stopping training when the performance on the validation set no longer improves). The training process continues for multiple rounds of iteration, usually requiring hundreds of iterations until the prediction error on the validation set is stable or reaches a preset threshold, indicating that the model parameters have converged.

[0039] Train a Bayesian neural network based on the spatial correlation characterization of power nodes and device state parameters to quantify the uncertainty of the power grid state. The Bayesian neural network is a probabilistic neural network. Different from traditional deterministic neural networks, its parameters are probability distributions rather than single values. This feature enables it to not only give the prediction result but also quantify the uncertainty of the prediction. The training data includes the spatial correlation characterization of power nodes (reflecting the mutual influence relationship between nodes) and device state parameters (such as device temperature, vibration frequency, running time, maintenance records, etc.). Due to the complexity of the Bayesian neural network, it is usually not feasible to directly calculate the exact posterior distribution, so the variational inference method is adopted for approximation. Variational inference approximates the posterior distribution of parameters as a tractable simple distribution (such as a Gaussian distribution), and optimizes the approximation by minimizing the KL divergence between the two distributions. Specifically, the Monte Carlo dropout method can be used, randomly shutting down some neurons during training and prediction, which is equivalent to taking the average of multiple different networks, so as to obtain the distribution of the prediction rather than a single-point estimate. After training, the state evaluation model can not only give the device state evaluation result (such as the failure probability) but also provide the confidence interval of this evaluation, quantifying the uncertainty of the prediction. This enables decision-makers to distinguish high-certainty evaluations (which should be acted on immediately) from high-uncertainty evaluations (which may require further monitoring or inspection).

[0040] Using the prediction results of the load prediction model and the evaluation results of the state assessment model, train a deep reinforcement learning network to optimize the resource allocation strategy. Deep reinforcement learning combines deep learning and reinforcement learning, and learns the optimal strategy through the interaction between the agent and the environment. First, define the state space, which includes information such as the load prediction results (future load distribution), state assessment results (equipment status and its uncertainty), and current resource status (such as the output of power generation equipment, the charge and discharge status of energy storage systems, and the situation of adjustable loads). Then, define the action space, which includes possible operations such as power generation equipment scheduling instructions (such as increasing or decreasing output), energy storage system control instructions (such as charging, discharging, maintaining), and load switching operations (such as peak shaving and valley filling). Designing the reward function is a key step, comprehensively considering multiple objectives such as economic cost (fuel consumption, power loss), system reliability (voltage stability, frequency stability), and environmental friendliness (carbon emissions). Adopt a dual network architecture, including a policy network and a value network. The policy network inputs the current state and outputs the probability distribution of different actions to guide the agent to select actions; the value network evaluates the long-term value of the state or state-action pair to assist in policy optimization. The training process uses deep Q-learning or policy gradient methods to accumulate experience through interaction with the simulated environment. Use an experience replay buffer to store historical interaction data (state, action, reward, next state), break the temporal correlation between samples, and improve training stability. To balance exploration and exploitation, adopt an epsilon-greedy strategy, and select random actions instead of the current optimal actions with a certain probability to explore possible better strategies. After sufficient training, the resource optimization model can generate the optimal resource allocation strategy according to the current system state, balancing multiple optimization objectives.

[0041] Taking a regional power grid as an example, this grid includes 50 main nodes (including 5 large substations, 15 medium substations, and 30 distribution substations) and the transmission lines connecting them. The historical data includes 3 months of continuous monitoring records. First, divide the data in chronological order, and use the data of the first 72 days as the training set and the data of the last 18 days as the validation set. Adopt the sliding window method, input 7 consecutive days of data each time, and predict the load distribution for the next 1 day. The connection weights between the nodes in the node relationship graph are calculated according to the physical distance and communication quality. For example, the connection weight between two substation nodes that are 2 kilometers apart but have good communication quality is 0.82, while the connection weight between nodes that are 10 kilometers apart and have average communication quality is only 0.45.

[0042] The spatio-temporal graph convolutional network is applied to training data to extract spatial dependence features. The network consists of 3 spatio-temporal convolutional blocks. In each block, the graph convolutional layer aggregates information from neighboring nodes, and the temporal convolutional layer captures short-term fluctuations. Through network processing, it is found that the load change in a certain area shows an obvious correlation with adjacent areas. For example, when the load in the industrial area increases, it is usually accompanied by a slight decrease in the load in the surrounding residential areas. These extracted spatial correlation features are input into a long short-term memory network, which has 2 layers with 100 units in each layer. The gating mechanism is used to capture complex temporal patterns: load peaks occur at 7-9 am and 6-8 pm on weekdays, and the load distribution is more flat on weekends; there is a lag correlation between the air-conditioning load in summer and the temperature. The load starts to increase significantly 1-2 hours after the temperature rises. The load prediction model uses a 4-layer feedforward neural network with the number of hidden layer units being [256, 128, 64], and the ReLU activation function is used. During the training process, the initial learning rate is set to 0.001 and decays by 10% every 50 iterations. The model converges after 300 iterations on the training set, and the mean square error drops from the initial 5.6 MW² to 0.8 MW². The error on the validation set is 1.2 MW², indicating that the model has good generalization ability. The state evaluation model uses a Bayesian neural network. By analyzing state parameters such as transformer temperature and vibration frequency, it is detected that there is an abnormality in Transformer No. 3, and the fault probability is given as 0.72, and the confidence interval is [0.57, 0.87], indicating a relatively high risk and medium certainty in the prediction. The resource optimization model is based on deep reinforcement learning. The state space dimension is 128, and the action space includes control instructions for 5 generator sets, 3 energy storage systems, and 10 adjustable loads. The reward function comprehensively considers power loss, equipment service life, and carbon emissions, with weights of 0.5, 0.3, and 0.2 respectively. The model is trained using a double deep Q network and converges after 100,000 environment interactions. The finally generated resource scheduling strategy includes: according to the predicted peak loads (8 am and 7 pm), adjust the output of Generator Sets No. 1 and No. 2, increasing them by 20 MW and 15 MW respectively; arrange the energy storage system to charge during the low load period (2-4 am) and discharge during the peak period, with the power set to 10 MW; at the same time, transfer the load of Transformer No. 3 to reduce its load rate from 85% to 65% and reduce the overload risk. This comprehensive scheduling plan not only meets the load demand but also reduces the equipment failure risk, while considering economy and environmental protection.

[0043] In a specific embodiment, the process of executing step S104 may specifically include the following steps: (1) Collect the real-time operation parameters of each node of the power grid through a distributed sensor network, including load level, equipment temperature, transmission line impedance, and phase angle difference, to form the real-time resource data of the power grid; (2) Input the real-time resource data of the power grid into the load prediction model, calculate the predicted values of the load distribution for the next 24 hours, and generate a load prediction result map; (3) Based on the real-time resource data of the power grid and the load prediction result atlas, perform Bayesian probability inference through the state evaluation model, calculate the stability margin index and fault probability value of each node, and construct a power grid stability evaluation matrix; (4) Conduct threshold analysis on the power grid stability evaluation matrix, identify high-risk nodes and potential fault points, and generate a risk level classification table; (5) Map the risk level classification table onto the power network topology diagram, mark the risk level and the influence scope, and form a fault risk distribution diagram; (6) Generate a power grid state evaluation report and a warning instruction including risk description, impact assessment, and handling suggestions based on the fault risk distribution diagram and the risk propagation evaluation result.

[0044] Specifically, the real-time operation parameters of each node of the power grid are collected through a distributed sensor network. These distributed sensors include intelligent power sensors, temperature sensors, vibration sensors, phasor measurement units, etc., which are respectively deployed at substations, key points of transmission lines, and distribution equipment. The data collected by the sensors include load level (indicating the magnitude of the electricity load, unit: megawatt), equipment temperature (indicating the working temperature of equipment such as transformers and circuit breakers, unit: degree Celsius), transmission line impedance (indicating the resistance and reactance values of the line, unit: ohm), and phase angle difference (indicating the voltage phase difference between different nodes of the network, unit: degree). These data are transmitted to the data processing center in real time through a 6G communication network, and the sampling frequency is usually at the second level or the minute level, forming the real-time resource data of the power grid with high spatio-temporal resolution.

[0045] Input the collected real-time resource data of the power grid into the trained load prediction model to calculate the predicted load distribution values for the next 24 hours. The load prediction model is obtained in the aforementioned deep learning model training stage and consists of a hybrid architecture of a spatio-temporal graph convolutional network and a long short-term memory network. The input data is first preprocessed, including missing value filling, outlier processing, and feature standardization. The preprocessed data is organized in the format required by the model, including historical load sequences, meteorological data (such as temperature, humidity), time features (such as hour, day of the week, holiday flag), etc. The model generates the predicted load distribution values for the next 24 hours hour by hour, including the electricity load of each node and the total load of the entire network. The prediction results are organized in two dimensions of time and space, forming a load prediction result atlas, visually showing the future load distribution, including peak hours, valley hours, and spatial distribution characteristics.

[0046] Based on the real-time resource data of the power grid and the load forecasting result atlas, Bayesian probability inference is carried out through the state evaluation model to calculate the stability margin index and failure probability value of each node. Bayesian probability inference is a statistical inference method based on Bayes' theorem, which can handle uncertainty and fuse multi-source information. In this solution, the state evaluation model is implemented using a Bayesian neural network to evaluate the state of each node. The calculation formula for the stability margin index (SII, Stability Index Indicator) is: where, represents the stability margin index of node , with a value range of [0, 1]. The larger the value, the higher the stability margin; represents the maximum carrying capacity of node ; represents the current load level of node ; represents the future load value of node predicted by the load forecasting model; represents the current load value of node ; represents the health state coefficient of node , which is calculated from parameters such as equipment temperature and vibration frequency.

[0047] The failure probability value of the node is calculated through a Bayesian network: where, represents the posterior probability of node failing under the condition of given observed data; represents the likelihood probability of observing data when node fails; represents the prior probability of node failing, which is set based on historical statistical data; represents the marginal probability of observed data . The calculated stability margin indices and failure probability values of each node are organized in matrix form to form the power grid stability evaluation matrix. The rows of the matrix represent different nodes, and the columns represent different evaluation indicators (such as stability margin index, failure probability value, temperature abnormality degree, etc.). This matrix representation is convenient for subsequent threshold analysis and risk level division.

[0048] Perform threshold analysis on the power grid stability assessment matrix to identify high-risk nodes and potential fault points. Threshold analysis includes single-index threshold judgment and multi-index comprehensive evaluation. Single-index threshold judgment sets the danger thresholds for each index. For example, if the stability margin index is lower than 0.3, it is judged as high risk; if the fault probability value is higher than 0.6, it is judged as high risk. Multi-index comprehensive evaluation considers the combined effects of multiple indexes and calculates the comprehensive risk score through weighted summation or fuzzy logic methods. According to the results of threshold analysis, the nodes are divided into three levels: high risk, medium risk, and low risk, and the risk sources (such as overload risk, equipment aging risk, short-term sudden load increase risk, etc.) are recorded in detail. This information is organized into a structured risk level classification table, which contains fields such as node identification, risk level, risk source, and risk index value.

[0049] Map the risk level classification table to the power network topology diagram to visually display the risk distribution. The power network topology diagram is a visual representation constructed based on the power network topology physical coordinate mapping table, where the nodes represent entities such as substations and distribution equipment, and the edges represent transmission lines. During the mapping process, nodes with different risk levels are marked with different colors: high-risk nodes are marked in red, medium-risk nodes are marked in yellow, and low-risk nodes are marked in green. At the same time, calculate the influence range of the risk, that is, analyze the downstream nodes and regions that may be affected by high-risk nodes, and determine it through topological connection relationships and power flow analysis. The influence range is marked with a dotted line or a shaded area on the topology diagram, supplemented by a quantitative description of the influence degree. The resulting fault risk distribution diagram visually displays the spatial distribution and potential propagation path of the power grid risk. According to the fault risk distribution diagram and the risk propagation assessment results, generate a power grid state assessment report and warning instructions. The power grid state assessment report is a structured document, which includes four main parts: an overview of the overall system state, a detailed analysis of high-risk nodes, an analysis of the risk propagation path, and countermeasure suggestions. The overview of the overall system state summarizes the risk distribution of the entire network, such as the number and proportion of high, medium, and low-risk nodes; the detailed analysis of high-risk nodes details the risk indicators, risk causes, and risk levels for each high-risk node; the analysis of the risk propagation path analyzes the possible diffusion paths and influence ranges of faults based on the power grid topology structure and power flow characteristics; the countermeasure suggestions provide specific treatment measures for different risks, such as equipment maintenance, load transfer, emergency power generation, etc. The warning instructions are a set of structured instruction sets, which are graded according to the urgency level and include information such as risk node identification, risk type, processing time limit, and countermeasure measures, facilitating the rapid implementation of risk control measures.

[0050] In a specific embodiment, the process of executing step S105 may specifically include the following steps: (1) Extract the system key stability indicators and load forecast data from the power grid state assessment report to construct a resource scheduling constraint condition set; (2)Calculate the risk coefficient and emergency power supply demand for each area based on the high-risk areas identified in the fault risk distribution map, and form a power supply priority table for risk areas; (3)Based on the fault type information and time estimate in the early warning instruction, formulate a power grid dispatching schedule for different time scales, and generate a dispatching cycle matrix; (4)According to the power supply priority table for risk areas and the dispatching cycle matrix, calculate the optimal output allocation ratio for various types of power generation equipment, and generate a power generation equipment output plan including the start-stop time, power curve and ramp rate of power generation; (5)Based on the power generation equipment output plan and the power grid load curve, calculate the power gap and surplus situation, allocate the charge and discharge time periods and power values for the energy storage system, and form an energy storage system charge and discharge strategy including the charge and discharge time, power magnitude and duration; (6)Based on the power supply priority table for risk areas and the stability assessment results of each node, determine the load-shedding nodes and the load-shedding sequence, and generate a load switching plan including the load switching time point, switching capacity and restoration sequence.

[0051] Specifically, extract the system key stability indicators and load forecasting data from the power grid state assessment report to construct a resource scheduling constraint condition set. The power grid state assessment report is a structured document, including information such as the overall system state overview, high-risk node analysis, risk propagation path and countermeasure suggestions. The extraction process uses natural language processing technology to identify the key indicators in the report, including the node stability margin index, fault probability value, voltage deviation rate and frequency stability degree, etc. At the same time, extract the load forecasting data, including the load forecasting values and fluctuation ranges of each node in the next 24 hours. These extracted data are standardized and then transformed into the constraint conditions for resource scheduling, including power generation capacity constraint, line transmission constraint, node voltage constraint and system frequency constraint, etc. These constraint conditions jointly constitute the resource scheduling constraint condition set, providing boundary conditions for subsequent optimal dispatching.

[0052] According to the high-risk areas identified in the fault risk distribution map, calculate the risk coefficient and emergency power supply demand for each area. The fault risk distribution map visually shows the spatial distribution of risks in the power grid, and marks nodes and areas with different risk levels through different colors. For each high-risk area, calculate its risk coefficient: Among them, represents the risk coefficient of area ; represents the node index within the area; represents the fault probability of node k; represents the fault probability threshold, usually taken as 0.6; represents node 's stability margin index; denotes the stability margin threshold, usually taken as 0.3; and are weight coefficients that adjust the relative importance of the failure probability and the stability margin in risk assessment; represents the criticality of node k, considering the importance and influence scope of the node.

[0053] Meanwhile, calculate the emergency power supply demand of each region: where, represents the emergency power supply demand of region ; represents the predicted load value of node k; represents the safety load threshold of node k; is an indicator function that takes the value of 1 when the stability margin index of node k is lower than the threshold, and 0 otherwise.

[0054] Based on the fault type information and time estimation in the early warning instruction, formulate the power grid scheduling time table at different time scales. The early warning instruction includes information such as the urgency level, fault type, estimated occurrence time, and duration. According to this information, the scheduling time is divided into multiple time scales: the emergency response time scale (minute level), the short-term scheduling time scale (hour level), and the medium-term scheduling time scale (day level). For each time scale, determine the start and end times, scheduling period, and scheduling step size of the scheduling. These information are organized into a scheduling period matrix, where the rows of the matrix represent different time scales, and the columns represent scheduling parameters (start and end times, period, step size). The scheduling period matrix provides a time frame for subsequent resource scheduling to ensure that corresponding scheduling operations are executed at the appropriate time.

[0055] According to the power supply priority table of the risk area and the scheduling period matrix, calculate the optimal output allocation ratio of various power generation equipment. For each power generation equipment, first consider its technical characteristics, including the maximum output, minimum output, ramp rate limit, start-stop time constraint, etc. The output allocation of the power generation equipment adopts the following formula: where, represents the output value of power generation equipment g at time point t; and represent the minimum and maximum outputs of power generation equipment g respectively; represents the power supply demand of region at time point t; represents the power supply priority of region ; represents the power supply accessibility of power generation equipment g to region , where h is an index variable representing all power generation equipment in the system.

[0056] At the same time, consider the ramp rate constraint of the power generation equipment: where represents the ramp rate limit of the power generation equipment g; represents the interval between adjacent scheduling time points.

[0057] According to the power generation equipment output plan and the power grid load curve, calculate the power gap and surplus situation, and allocate the charge and discharge time periods and power values for the energy storage system. Based on the power supply priority table of the risk area and the stability evaluation results of each node, determine the load shedding nodes and the load shedding sequence, and generate a load switching plan.

[0058] In a specific embodiment, the process of executing step S106 may specifically include the following steps: (1) Convert the power generation equipment output plan, the energy storage system charge and discharge strategy, and the load switching plan into a standardized control instruction format, and generate an instruction data packet; (2) Mark the priority of the instruction data packet, and divide it into three levels: critical control instructions, regular scheduling instructions, and auxiliary management instructions according to the urgency; (3) Dynamically allocate communication resources through the 6G network slicing technology, create an ultra-reliable low-latency network slice for critical control instructions, create a mobile broadband slice for regular scheduling instructions, and create a machine-type communication slice for auxiliary management instructions; (4) Synchronize the control instructions among nodes to ensure the time synchronization and logical consistency of multi-node control, and generate a synchronized control instruction set; (5) Send the synchronized control instruction set to each control execution unit through their respective network slices, and collect the execution feedback data to form power network operation status data including the measured parameter values, equipment status, and instruction execution results; (6) Based on the power network operation status data, conduct multi-dimensional comparison and analysis of the real-time operation parameters of the power network, record various parameter deviation values and fluctuation curves, and associate the corresponding control instruction execution situations to form a performance evaluation report including the power system operation quality evaluation, control instruction execution effect, and resource scheduling rationality.

[0059] Specifically, convert the power generation equipment output plan, energy storage system charge and discharge strategy, and load switching plan into a standardized control instruction format to generate an instruction data packet. This involves three types of data conversion: converting the power generation equipment output plan into a power setting instruction, which includes parameters such as the target power value, execution time point, and ramp rate limit; converting the energy storage system charge and discharge strategy into an energy storage control instruction, including information such as charge and discharge mode, power magnitude, and duration; converting the load switching plan into a load control instruction, including data such as cut-off / resume flag, target load quantity, and execution time point. The conversion uses a unified data structure, usually consisting of an instruction header (transmission time, receiving object, instruction type), an instruction body (specific parameter settings), and a checksum (to ensure data integrity). This standardized format helps different types of devices understand and execute control instructions, and also facilitates the transmission and management of instructions.

[0060] Mark the generated instruction data packet with priorities and divide them into three levels according to the urgency. Critical control instructions involve core operations for the safe and stable operation of the system, such as emergency power generation adjustment to deal with sudden failures and emergency load shedding. Such instructions have the highest priority and require immediate execution without delay. Routine scheduling instructions include normal operation instructions such as daily power balance adjustment and economic optimization scheduling, which have medium priority and allow short delays but need to be completed according to the plan. Auxiliary management instructions involve non-critical operations such as status monitoring, data collection, and parameter fine-tuning, and have relatively low priority with less strict requirements for the execution time. The priority marking uses a binary flag method, adding a priority field to the instruction header and using 2-bit binary codes to represent the three levels (00 for auxiliary management, 01 for routine scheduling, and 10 for critical control). Through priority marking, critical instructions can be guaranteed to be transmitted and executed first when network resources are limited.

[0061] Dynamically allocate communication resources through 6G network slicing technology to create dedicated network slices for instructions with different priorities. Network slicing is a core technology of 6G networks. By using virtualization technology, multiple logically independent end-to-end networks are created on the same physical network, and each slice has unique quality-of-service characteristics. Create an ultra-reliable low-latency network slice for critical control instructions. This slice has a transmission latency of less than 1 millisecond and uses dedicated spectrum resources and redundant transmission paths to ensure reliable transmission of instructions even in emergency situations. Create an enhanced mobile broadband slice for routine scheduling instructions. This slice provides high bandwidth and medium latency (5 - 10 milliseconds) to meet the transmission requirements of a large amount of scheduling data. Create a massive machine-type communication slice for auxiliary management instructions. This slice supports a large number of device connections. Although the bandwidth of a single connection is low, it can effectively handle a large amount of non-critical data exchange. The creation of network slices uses software-defined network technology, and the network resources are dynamically allocated by a central controller, and the slice parameters are adjusted according to real-time communication requirements.

[0062] The distributed consensus algorithm is used to synchronize control instructions among nodes, ensuring the time synchronization and logical consistency of multi-node control. In a power network, the control operations of multiple nodes usually need to be coordinated, such as the joint dispatching of multiple generators and the collaborative load transfer of multiple distribution areas. The distributed consensus algorithm is based on the two-phase commit protocol. First, a pre-execution message is sent to all executing nodes to collect the readiness status of each node. When it is confirmed that all nodes are ready, a formal execution message is sent to trigger synchronous execution. To solve the problem of time asynchrony caused by network latency differences, the Precision Time Protocol (PTP) is used to achieve clock synchronization at the microsecond level, and all nodes execute instructions based on the same time reference. Through this mechanism, a synchronized control instruction set containing global timestamps and execution sequence numbers is generated to ensure the coordination of complex operations.

[0063] The synchronized control instruction set is sent to each control execution unit through their respective network slices, and execution feedback data is collected. The sending process combines point-to-point and multicast methods. Key control instructions usually use point-to-point direct transmission to ensure reliable delivery, while conventional scheduling instructions and auxiliary management instructions can use multicast to improve transmission efficiency. After the instructions are sent, each execution unit returns confirmation information, including the instruction reception status, execution progress, and execution result. During the execution process, the control execution unit continuously collects device operation data, including measured values of actual output power, voltage and current values, device temperature and other parameters, as well as device operation status (such as normal, alarm, fault, etc.) information. This data, together with the instruction execution result (success, partial success or failure), is uploaded to the data processing center through the 6G network to form power network operation status data. The data is stored in a time series database for subsequent time series analysis and performance evaluation.

[0064] Based on the power network operation status data, multi-dimensional comparison and analysis of the real-time operation parameters of the power network are carried out. First, parameter deviation analysis is performed to calculate the difference between the actual value and the instruction set value, such as power generation deviation, voltage deviation, etc., to evaluate the control accuracy. Then, time response analysis is carried out to measure the time from instruction issuance to execution completion to evaluate the control timeliness. Next, stability analysis is carried out to record the parameter fluctuation curve and calculate the standard deviation and the maximum fluctuation amplitude to evaluate the operation stability. Finally, correlation analysis is carried out to correlate the parameter changes with the execution of control instructions to identify causal relationships. Through these analyses, a performance evaluation report including the operation quality evaluation of the power system, the execution effect of control instructions, and the rationality of resource scheduling is formed. This report not only summarizes the current control effect but also provides data support for the optimization of subsequent scheduling strategies.

[0065] The above describes the intelligent management method of power grid resources based on 6G and deep learning in the embodiments of the present application. Next, the intelligent management system of power grid resources based on 6G and deep learning in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the intelligent management system of power grid resources based on 6G and deep learning in the embodiments of the present application includes: An acquisition module, configured to acquire power network node data through a 6G communication base station and multi-source positioning sensors, and obtain a power network topology physical coordinate mapping table, a real-time power parameter data stream, and communication network status information; A processing module, configured to perform data preprocessing on the power network topology physical coordinate mapping table, the real-time power parameter data stream, and the communication network status information to obtain a power data feature tensor; A training module, configured to train a deep learning model cluster through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network according to the power data feature tensor, and obtain a load prediction model, a state evaluation model, and a resource optimization model; An analysis module, configured to collect real-time power grid resource data, and perform power grid state analysis on the real-time power grid resource data according to the load prediction model, the state evaluation model, and the resource optimization model, and obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction; A calculation module, configured to calculate a resource scheduling plan according to the power grid state evaluation report, the fault risk distribution map, and the warning instruction, and obtain a power generation equipment output plan, a charge and discharge strategy of an energy storage system, and a load switching plan; A control module, configured to issue control instructions through 6G network slicing according to the power generation equipment output plan, the charge and discharge strategy of the energy storage system, and the load switching plan, and obtain power network operation state data and a performance evaluation report.

[0066] Through the collaborative cooperation of the above-mentioned various components, the 6G communication base station and multi-source positioning sensors are used to collect power network node data, obtaining the power network topology physical coordinate mapping table, real-time power parameter data stream and communication network status information, solving the problems of insufficient data collection accuracy and poor real-time performance in traditional technologies, and providing a high-quality data basis for subsequent intelligent analysis. The technical feature of preprocessing the collected data to form a power data feature tensor effectively removes the noise and redundancy in the original data, improves the data quality, and enables the deep learning model to more accurately extract the key features of the power network. The application of the hybrid architecture of the spatio-temporal graph convolutional neural network and the long short-term memory network fully considers the spatial topology structure and time series characteristics of the power network. Compared with the traditional single-structure neural network, this hybrid architecture can simultaneously capture the spatial dependence relationship between power nodes and the time pattern of load changes, and the trained load prediction model, state evaluation model and resource optimization model have higher accuracy and stronger generalization ability. The technical feature of collecting real-time resource data of the power grid and using the above models for power grid state analysis realizes the early identification and risk assessment of abnormal power grid states. Compared with the traditional threshold-based monitoring method, this deep learning-based state analysis can discover complex non-linear relationships and potential risk patterns, greatly improving the accuracy of power grid security early warning. The technical feature of calculating the resource scheduling scheme according to the power grid state evaluation results realizes the collaborative optimization of power generation, energy storage and load. Compared with the traditional single-resource optimization, this multi-resource collaborative scheduling greatly improves the system flexibility and resource utilization efficiency. The technical feature of sending control instructions through 6G network slicing ensures the ultra-reliable and low-latency transmission of key control instructions, solves the congestion and delay problems that may occur in traditional communication networks in case of emergencies, and enables the control instructions to be executed accurately and in a timely manner.

[0067] Referring to Figure 3 , an embodiment of the present invention further provides a computer device, which may be a server, and its internal structure may be as Figure 3 shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0068] Those skilled in the art can understand that Figure 3The structure shown is only a block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0069] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0070] Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0071] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0072] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0073] As described above, the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. An intelligent management method for power grid resources based on 6G and deep learning, characterized in that, The intelligent management method of power grid resources based on 6G and deep learning includes: Collect power network node data through 6G communication base stations and multi-source positioning sensors to obtain a power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information; Perform data preprocessing on the power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information to obtain a power data feature tensor; According to the power data feature tensor, train a deep learning model cluster through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network to obtain a load prediction model, a state evaluation model, and a resource optimization model; Collect real-time power grid resource data, and perform power grid state analysis on the real-time power grid resource data according to the load prediction model, state evaluation model, and resource optimization model to obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction; Calculate a resource scheduling plan according to the power grid state evaluation report, fault risk distribution map, and warning instruction to obtain a power generation equipment output plan, an energy storage system charge and discharge strategy, and a load switching plan; According to the power generation equipment output plan, energy storage system charge and discharge strategy, and load switching plan, send control instructions through 6G network slicing to obtain power network operation status data and a performance evaluation report.

2. The intelligent management method of power grid resources based on 6G and deep learning according to claim 1, characterized in that, The process of collecting power network node data through 6G communication base stations and multi-source positioning sensors to obtain a power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information includes: Deploy 6G communication modules at key points of substations, transmission lines, and distribution equipment in the power network, measure the signal strength and transmission delay between nodes, and generate a node communication connection matrix; Collect the voltage value, current value, active power, reactive power, and frequency value of each node in the power network through intelligent power sensors every second to form a real-time power parameter data stream; Use a GPS receiver and a base station-assisted positioning system to collect the accurate position of the equipment, and process the original positioning data using the Kalman filter algorithm to obtain centimeter-level equipment spatial coordinates; Perform point-to-point matching on the link state data in the node communication connection matrix and the equipment spatial coordinates to construct a power network topology physical coordinate mapping table including physical distance and link quality; Measure the signal-to-noise ratio, bit error rate, and available bandwidth of the 6G communication link, record the real-time state parameters of each communication channel, and generate communication network status information; Mark and align the timestamps of the data points in the real-time power parameter data stream to eliminate the sampling time difference of different devices and complete the synchronization processing of power network node data.

3. The intelligent management method of power grid resources based on 6G and deep learning according to claim 1, characterized in that The process of performing data preprocessing on the power network topology physical coordinate mapping table, real-time power parameter data streams, and communication network status information to obtain a power data feature tensor includes: Perform wavelet transform decomposition on the real-time power parameter data stream to separate the high-frequency noise component and low-frequency trend component in the power signal, and form a power signal component matrix; Use an autoencoder to detect and repair outliers and missing values in the power signal component matrix to generate a cleaned power parameter matrix; Construct a node adjacency relationship graph based on the power network topology physical coordinate mapping table, and calculate the connection weight values between nodes; Normalize the signal-to-noise ratio, bit error rate, and bandwidth data in the communication network state information to form a communication state feature vector; Associate and fuse the cleaned power parameter matrix with the node adjacency relationship graph and the communication state feature vector to construct a multi-dimensional association matrix; Perform tensor decomposition on the multi-dimensional association matrix through the Tucker decomposition algorithm. Set the core tensor rank of the time dimension to 24, the core tensor rank of the space dimension to 30% of the number of network nodes, and the core tensor rank of the physical property dimension to 8, and perform multi-linear feature extraction to obtain the power data feature tensor.

4. The intelligent management method of power grid resources based on 6G and deep learning according to claim 1, characterized in that, According to the power data feature tensor, train a deep learning model cluster through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network to obtain a load prediction model, a state evaluation model, and a resource optimization model, including: Divide the power data feature tensor into a training set and a validation set according to the time dimension, establish a power network node relationship graph, and construct a model training data set; Apply a spatio-temporal graph convolutional network to the model training data set to extract the spatial dependence features between power nodes and generate a spatial correlation representation of power nodes; Input the spatial correlation representation of power nodes into a long short-term memory network to capture the temporal variation law of power load and obtain temporal correlation features; Construct a load prediction model based on the temporal correlation features and historical load data, optimize the loss function through the gradient descent algorithm, and adjust the model parameters until the prediction error converges; Train a Bayesian neural network based on the spatial correlation representation of power nodes and equipment state parameters to quantify the uncertainty of the power grid state and form a state evaluation model; Use the prediction results of the load prediction model and the evaluation results of the state evaluation model to train a deep reinforcement learning network, and optimize the resource allocation strategy through a reward and punishment mechanism to obtain a resource optimization model.

5. The intelligent management method of power grid resources based on 6G and deep learning according to claim 1, wherein, Collect real-time resource data of the power grid, and perform power grid state analysis on the real-time resource data of the power grid according to the load prediction model, the state evaluation model, and the resource optimization model to obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction, including: Collect the real-time operation parameters of each node of the power grid through a distributed sensor network, including load level, equipment temperature, transmission line impedance, and phase angle difference, to form real-time resource data of the power grid; Input the real-time resource data of the power grid into the load prediction model, calculate the predicted values of the load distribution for the next 24 hours, and generate a load prediction result map; According to the real-time resource data of the power grid and the load prediction result map, perform Bayesian probability inference through the state evaluation model, calculate the stability margin index and fault probability value of each node, and construct a power grid stability evaluation matrix; Perform threshold analysis on the power grid stability evaluation matrix to identify high-risk nodes and potential fault points, and generate a risk level classification table; Map the risk level classification table to the power network topology map, mark the risk level and the influence range, and form a fault risk distribution map; Generate a power grid status assessment report and early warning instructions containing risk descriptions, impact assessments, and handling suggestions based on the fault risk distribution map and risk propagation assessment results.

6. The intelligent management method of power grid resources based on 6G and deep learning according to claim 1, characterized in that, Calculate a resource scheduling plan based on the power grid status assessment report, fault risk distribution map, and early warning instructions to obtain a power generation equipment output plan, energy storage system charge and discharge strategy, and load switching plan, including: Extract system key stability indicators and load prediction data from the power grid status assessment report to construct a resource scheduling constraint condition set; Calculate the risk coefficient and emergency power supply demand of each region according to the high-risk regions identified in the fault risk distribution map to form a power supply priority table for risk regions; Based on the fault type information and time estimate in the early warning instructions, formulate a power grid scheduling time table with different time scales to generate a scheduling cycle matrix; According to the power supply priority table for risk regions and the scheduling cycle matrix, calculate the optimal output allocation ratio for various power generation equipment to generate a power generation equipment output plan containing power generation start and stop times, power curves, and ramp rates; Based on the power generation equipment output plan and the power grid load curve, calculate the power gap and surplus conditions, allocate charge and discharge time periods and power values for the energy storage system to form an energy storage system charge and discharge strategy containing charge and discharge times, power magnitudes, and durations; Based on the power supply priority table for risk regions and the stability assessment results of each node, determine the load shedding nodes and load shedding order to generate a load switching plan containing load switching time points, switching capacities, and restoration orders; 7. The intelligent management method of power grid resources based on 6G and deep learning according to claim 1, characterized in that According to the power generation equipment output plan, energy storage system charge and discharge strategy, and load switching plan, send control instructions through 6G network slicing to obtain power network operation status data and performance assessment reports, including: Convert the power generation equipment output plan, energy storage system charge and discharge strategy, and load switching plan into a standardized control instruction format to generate an instruction data packet; Mark the priority of the instruction data packet and divide it into three levels: critical control instructions, regular scheduling instructions, and auxiliary management instructions according to the urgency; Dynamically allocate communication resources through 6G network slicing technology to create an ultra-reliable low-latency network slice for the critical control instructions, a mobile broadband slice for the regular scheduling instructions, and a machine-type communication slice for the auxiliary management instructions; Synchronize the control instructions for nodes to ensure the time synchronization and logical consistency of multi-node control to generate a synchronized control instruction set; Send the synchronized control instruction set to each control execution unit through their respective network slices and collect execution feedback data to form power network operation status data containing measured parameter values, equipment status, and instruction execution results; Based on the power network operation status data, conduct multi-dimensional comparison and analysis of the real-time operation parameters of the power network, record various parameter deviation values and fluctuation curves, and associate the corresponding control instruction execution situations to form a performance assessment report containing power system operation quality evaluation, control instruction execution effect, and resource scheduling rationality.

8. An intelligent management system for power grid resources based on 6G and deep learning, which is used to implement the intelligent management method for power grid resources based on 6G and deep learning according to any one of claims 1-7, characterized in that, The intelligent power grid resource management system based on 6G and deep learning includes: The acquisition module is used to collect power network node data through 6G communication base stations and multi-source positioning sensors, and obtain a power network topology physical coordinate mapping table, a real-time power parameter data stream, and communication network status information; The processing module is used to perform data preprocessing on the power network topology physical coordinate mapping table, the real-time power parameter data stream, and the communication network status information to obtain a power data feature tensor; The training module is used to train a deep learning model cluster according to the power data feature tensor through a hybrid architecture of a spatio-temporal graph convolutional neural network and a long short-term memory network, and obtain a load prediction model, a state evaluation model, and a resource optimization model; The analysis module is used to collect real-time power grid resource data, and perform power grid state analysis on the real-time power grid resource data according to the load prediction model, the state evaluation model, and the resource optimization model, and obtain a power grid state evaluation report, a fault risk distribution map, and a warning instruction; The calculation module is used to calculate a resource scheduling plan according to the power grid state evaluation report, the fault risk distribution map, and the warning instruction, and obtain a power generation equipment output plan, a charge and discharge strategy of an energy storage system, and a load switching plan; The control module is used to issue control instructions through 6G network slicing according to the power generation equipment output plan, the charge and discharge strategy of the energy storage system, and the load switching plan, and obtain power network operation status data and a performance evaluation report.

9. A computer device, characterized in that, It includes a memory and a processor, and the memory stores a computer program that can be run on the processor. The processor, when executing the computer program, implements the 6G and deep learning-based intelligent management method for power grid resources according to any one of claims 1 to 7.

10. A computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the processor is caused to execute the 6G and deep learning-based intelligent management method for power grid resources according to any one of claims 1 to 7.

Images