Bus load forecasting method and device based on multi-dimensional demand coordination of power system

Abstract

The application provides a bus load prediction method and device based on power system multi-dimensional demand coordination, and relates to the technical field of power systems.The method comprises the following steps: collecting power system multi-dimensional demand data, and constructing a time-space correlation model of multi-dimensional demand and bus load; correlatively analyzing the collected multi-dimensional demand data, and outputting a multi-dimensional demand coordination coefficient; taking the multi-dimensional demand coordination coefficient as a long short-term memory network input correction term of a bus load time series prediction model, adjusting the weight distribution of each input feature, performing bus load prediction, and obtaining a bus load prediction result.The application can solve the technical problem that the bus load prediction in the prior art cannot accurately reflect the real influence of demand changes under complex operating conditions, and can accurately capture the influence of power source side output fluctuation, load side response characteristics and power grid side topology adjustment on the bus load, and improve the support capability of prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a method and equipment for bus load forecasting based on multi-dimensional demand coordination of power systems. Background Technology

[0002] With the large-scale integration of new energy sources, the increasing diversification of power load structures, and the growing complexity of power grid operation, bus load forecasting, as a crucial foundation for power system planning, dispatching, and safe operation, directly impacts the efficiency of power grid resource allocation and operational safety. Especially in operating environments characterized by a high proportion of new energy sources, electric vehicles, flexible loads, and frequent topology adjustments, bus load is no longer solely determined by historical load evolution but is simultaneously influenced by a combination of factors, including uncertainties in power generation output, load response behavior, and changes in grid structure.

[0003] Currently, most existing bus load forecasting technologies are based on traditional time series analysis methods or single deep learning models. They typically focus on modeling historical load data or a small number of exogenous variables, lacking a systematic characterization of the interactions between multi-dimensional demands from the power generation side, load side, and grid side. In particular, they struggle to reflect the objective laws governing the propagation of different demand factors along spatial paths within the grid topology and their time-lag effects. In practical applications, these forecasting methods often simply concatenate various demand characteristics and directly input them into the forecasting model, failing to establish a spatiotemporal correlation mechanism between demand changes and bus load. This results in the model's inability to identify the transmission paths, intensity, and time delay characteristics of demand impacts. Consequently, when renewable energy fluctuations increase, demand responses become more frequent, or the topology undergoes adjustments, the forecast results lack sensitivity to changes in operating conditions.

[0004] In summary, existing technologies suffer from a lack of a unified modeling and coordination analysis mechanism for the spatiotemporal coupling relationship between multidimensional power system demand and bus load. This results in bus load forecasting failing to accurately reflect the true impact of demand changes under complex operating conditions, further affecting the foresight of power grid dispatching decisions and operational safety. Summary of the Invention

[0005] The purpose of this application is to provide a method and equipment for bus load forecasting based on multi-dimensional demand coordination of power systems, in order to solve the technical problem that the lack of a unified modeling and coordination analysis mechanism for the spatiotemporal coupling relationship between multi-dimensional demand of power systems and bus load in the existing technology makes it difficult for bus load forecasting to accurately reflect the real impact of demand changes under complex operating conditions, and further affects the foresight of power grid dispatching decisions and operational safety.

[0006] In view of the above problems, this application provides a method and equipment for bus load forecasting based on multi-dimensional demand coordination of power systems.

[0007] Firstly, this application provides a bus load forecasting method based on multi-dimensional demand coordination in a power system. This method is implemented using a bus load forecasting device based on multi-dimensional demand coordination in a power system. The method includes: collecting multi-dimensional demand data from the power source side, load side, and grid side of the power system; constructing a spatiotemporal correlation model between multi-dimensional demand and bus load using a graph neural network to reflect the influence weights of each dimension of demand on the bus load and the spatiotemporal transmission rules; performing correlation analysis on the collected multi-dimensional demand data through the spatiotemporal correlation model to output multi-dimensional demand coordination coefficients; using the multi-dimensional demand coordination coefficients as input correction terms in the long short-term memory network of the bus load time series forecasting model; and performing bus load forecasting by adjusting the weight allocation of each input feature to obtain the bus load forecasting result.

[0008] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: power source side data including the fluctuation rate of new energy output, day-ahead forecast deviation distribution and inverter control mode; load side data including the adjustable capacity of flexible load, demand response event status and thermal inertia parameters of temperature-controlled load; and grid side data including real-time topology, equipment maintenance plan and safety constraint margin.

[0009] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: using sliding window statistics on power source side data to extract ultra-short-term volatility as an uncertainty quantification index to describe the uncertainty index of the power grid's demand for reserve capacity and rapid adjustment; constructing a demand response event map on load side data to mark the response delay characteristics of interruptible loads and the spatiotemporal transferability attributes of electric vehicle charging, thereby obtaining flexible adjustment demand characteristics oriented towards dispatch instructions; and identifying power flow transfer paths and weak links under fault conditions by correlating and analyzing protection device action signals and topology change information on power grid side data, thereby extracting the proactive defensive security demand characteristics of the power grid for load distribution and reconfiguration.

[0010] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: constructing a dynamic graph neural network based on the real-time topology, wherein the bus is used as a node and the transmission line is used as an edge, and the multi-dimensional demand data is input into the dynamic graph neural network as node features; based on the dynamic graph neural network, the spatial influence relationship and time transmission lag characteristics of each dimension of demand on the bus load are learned through a message passing mechanism based on the historical sample set, and a multi-dimensional demand coordination coefficient including spatiotemporal attention weights, dimensional gating coefficients and prediction confidence is generated to construct the spatiotemporal correlation model.

[0011] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: iteratively updating feature representations among neighboring nodes based on the message passing mechanism of a dynamic graph neural network, learning the spatial coupling strength of demand in each dimension to the bus load, and capturing the lag relationship of spatial coupling strength in the time dimension through temporal relationships, identifying the time delay characteristics of the impact of power supply output changes, load demand response, and grid topology adjustments on the bus load. Specifically, the electrical coupling strength and power interaction direction of adjacent buses under the current topology are quantified to generate the spatiotemporal attention weights, which reflect the spatial transmission impact of grid topology on each bus load; the input contribution ratios of regulating power supply uncertainty characteristics, load elastic regulation characteristics, and grid topology constraint characteristics are determined respectively to generate the dimensional gating coefficients, which reflect the dominant differences in demand in each dimension under different operating conditions; and the reliability range of the output forecast results is quantified based on historical forecast error distribution and current data quality assessment to generate the forecast confidence level.

[0012] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: using the spatial attention weight as the adjustment threshold of the forget gate, reducing the retention ratio of historical load memory states when a grid-side topology reconfiguration event is detected, and enhancing the response sensitivity to real-time topology changes; using the dimensional gating coefficient as the weighting coefficient of the input gate, increasing the input weight of the power supply side uncertainty characteristics when the power supply side renewable energy output fluctuation rate exceeds a preset threshold, and reducing the interference of historical concurrent load patterns; and selecting the time step for the multi-dimensional demand characteristics input to the long short-term memory network based on the time delay characteristics, selecting the characteristic data corresponding to the historical moment according to the different delay characteristics of the impact of power supply side output changes, load side demand response, and grid-side topology adjustments on the bus load, and compensating for the spatiotemporal transmission lag.

[0013] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of power systems further includes: selecting and aligning the input features based on time delay characteristics, mapping demand feature data with different delay characteristics to a unified time section, and obtaining a spatiotemporally decoupled synchronous input feature vector; updating the internal memory unit state based on the synchronous input feature vector, combined with the corrected forget gate and input gate output, to obtain a comprehensive state containing topology awareness and demand coordination information; selectively extracting information from the memory unit state through the output gate to generate a hidden state vector at the current moment, wherein the hidden state vector includes multi-dimensional demand spatiotemporal correlation features corrected by the coordination coefficient; inputting the hidden state vector into a fully connected decoding layer, and outputting the load forecast value of the target bus at the next moment through linear mapping and nonlinear activation processing, and fusing the prediction confidence coefficient with the historical prediction error statistical distribution to generate the standard deviation interval of the load point prediction value, and outputting the bus load forecast result containing the point prediction value and the confidence interval.

[0014] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: introducing power grid physical constraints during the training process of the long short-term memory network, constructing a power flow residual calculation layer, wherein the power flow residual calculation layer is constructed based on Kirchhoff's current law, comparing the bus load forecasting results with the power balance calculated based on the power grid topology; adding the power flow residual as a regularization term to the loss function, adjusting the long short-term memory network parameters through backpropagation, so that the output bus load forecasting value satisfies the power balance constraints of the entire network, and avoiding conflicts between the single bus forecasting results and the physical laws of the power grid.

[0015] Preferably, the bus load forecasting method based on multi-dimensional demand coordination of the power system further includes: dividing the multi-dimensional demand data into fast-changing components and slow-changing components, wherein the fast-changing components include high-frequency dynamic features extracted by encoding through a one-dimensional convolutional neural network, such as second- to minute-level fluctuations in renewable energy output, instantaneous adjustment of demand response, and fault transient processes; and the slow-changing components include low-frequency trend features extracted by multi-scale average pooling encoding, such as hourly temperature accumulation trends, day-ahead power generation plan curves, and equipment maintenance durations. The high-frequency dynamic features and low-frequency trend features are then concatenated to obtain multi-scale features, which are then input into the dynamic graph neural network.

[0016] Secondly, this application also provides a bus load forecasting device based on multi-dimensional demand coordination of a power system, used to execute the bus load forecasting method based on multi-dimensional demand coordination of a power system as described in the first aspect, including: a spatiotemporal correlation model construction module, used to collect multi-dimensional demand data from the power source side, load side, and grid side of the power system, and to construct a spatiotemporal correlation model between multi-dimensional demand and bus load using a graph neural network, used to reflect the influence weight of each dimension of demand on the bus load and the spatiotemporal transmission rules; a multi-dimensional demand coordination coefficient output module, used to perform correlation analysis on the collected multi-dimensional demand data through the spatiotemporal correlation model, and output multi-dimensional demand coordination coefficients; and a bus load forecasting result acquisition module, used to use the multi-dimensional demand coordination coefficients as input correction terms of the long short-term memory network of the bus load time series forecasting model, and to perform bus load forecasting by adjusting the weight allocation of each input feature, thereby obtaining the bus load forecasting result.

[0017] The technical solution provided in this application has at least the following technical effects or advantages: by achieving the technical goal of modeling and dynamically predicting the spatiotemporal correlation of bus load based on multi-dimensional demand coordination, it can accurately capture the impact of power supply side output fluctuations, load side response characteristics and grid side topology adjustments on bus load, thereby improving prediction accuracy, response real-time performance and support for grid operation and scheduling decisions.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the bus load forecasting method based on multi-dimensional demand coordination of the power system proposed in this application.

[0021] Figure 2 This is a schematic diagram of the bus load forecasting device based on multi-dimensional demand coordination of the power system, as described in this application.

[0022] Figure labeling: Spatiotemporal correlation model construction module 1, multi-dimensional demand coordination coefficient output module 2, bus load prediction result acquisition module 3. Detailed Implementation

[0023] This application provides a method and equipment for bus load forecasting based on multi-dimensional demand coordination in power systems. It addresses the technical problem in existing technologies where the lack of a unified modeling and coordination mechanism for the spatiotemporal coupling relationship between multi-dimensional power system demand and bus load leads to inaccurate bus load forecasting that fails to reflect the true impact of demand changes under complex operating conditions, further affecting the foresight and operational safety of power grid dispatching decisions. The application achieves the technical goal of spatiotemporal correlation modeling and dynamic forecasting of bus load based on multi-dimensional demand coordination, enabling precise capture of the impact of power source output fluctuations, load response characteristics, and grid topology adjustments on bus load, thereby improving forecast accuracy, real-time response, and support for power grid operation and dispatching decisions.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a bus load forecasting method based on multi-dimensional demand coordination of power systems, which is applied to bus load forecasting equipment based on multi-dimensional demand coordination of power systems. Specifically, it includes the following steps: Collect multi-dimensional demand data from the power source side, load side, and grid side of the power system, and use graph neural networks to construct a spatiotemporal correlation model between multi-dimensional demand and bus load to reflect the influence weight of each dimension of demand on bus load and the spatiotemporal transmission rules.

[0026] Furthermore, this application also includes: power supply side data including the volatility of new energy output, day-ahead forecast deviation distribution and inverter control mode; load side data including the adjustable capacity of flexible loads, demand response event status and thermal inertia parameters of temperature-controlled loads; grid side data including real-time topology, equipment maintenance plans and safety constraint margins.

[0027] Furthermore, this application also includes: using sliding window statistics on power source side data to extract ultra-short-term volatility as an uncertainty quantification index to describe the uncertainty index of the grid's demand for reserve capacity and rapid adjustment; constructing a demand response event map on load side data to mark the response delay characteristics of interruptible loads and the spatiotemporal transferability attributes of electric vehicle charging, thereby obtaining flexible adjustment demand characteristics oriented towards dispatch instructions; and identifying power flow transfer paths and weak links under fault conditions by correlating protection device action signals and topology change information with grid side data, thereby extracting the grid's proactive defensive security demand characteristics for load distribution and reconfiguration.

[0028] Furthermore, this application also includes: constructing a dynamic graph neural network based on the real-time topology, wherein the bus is used as a node and the transmission line is used as an edge, and the multi-dimensional demand data is input into the dynamic graph neural network as node features; based on the dynamic graph neural network, learning the spatial influence relationship and time transmission lag characteristics of each dimension of demand on the bus load through a message passing mechanism according to the historical sample set, generating a multi-dimensional demand coordination coefficient including spatiotemporal attention weights, dimensional gating coefficients and prediction confidence, and constructing the spatiotemporal correlation model.

[0029] Furthermore, this application also includes: dividing the multi-dimensional demand data into fast-changing components and slow-changing components, wherein the fast-changing components include second- to minute-level fluctuations in new energy output, instantaneous adjustment of demand response, and transient fault processes, and high-frequency dynamic features extracted by encoding through a one-dimensional convolutional neural network; the slow-changing components include hourly temperature accumulation trends, day-ahead power generation plan curves, and equipment maintenance durations, and low-frequency trend features extracted by multi-scale average pooling encoding; the high-frequency dynamic features and low-frequency trend features are concatenated to obtain multi-scale features, which are then input into the dynamic graph neural network.

[0030] Furthermore, this application also includes: based on the message passing mechanism of a dynamic graph neural network, iteratively updating feature representations among neighboring nodes, learning the spatial coupling strength of demand in each dimension to the bus load, and capturing the lag relationship of spatial coupling strength in the time dimension through temporal relationships, identifying the time delay characteristics of the impact of power supply output changes, load demand response, and grid topology adjustments on the bus load. Specifically, it quantifies the electrical coupling strength and power interaction direction of adjacent buses under the current topology, generating the spatiotemporal attention weights to reflect the spatial transmission impact of grid topology on each bus load; it determines the input contribution ratios of regulating power supply uncertainty characteristics, load elastic regulation characteristics, and grid topology constraint characteristics, generating the dimensional gating coefficients to reflect the dominant differences in demand in each dimension under different operating conditions; and it quantifies the reliability range of the output prediction results based on historical prediction error distribution and current data quality assessment to generate the prediction confidence level.

[0031] Specifically, the power supply side data includes renewable energy output volatility, day-ahead forecast deviation distribution, and inverter control modes. Renewable energy output volatility characterizes the unstable power output characteristics of renewable energy sources such as wind and solar power during operation due to changes in meteorological conditions and the environment, reflecting the randomness and uncertainty of power output over time. Day-ahead forecast deviation distribution describes the statistical deviation between the day-ahead power generation forecast results and the actual output, and is used to characterize the systematic error characteristics of the forecast model under different operating periods and conditions. Inverter control modes indicate the control strategy configuration status of renewable energy grid-connected inverters during voltage support, reactive power regulation, and power point tracking, thereby reflecting the dynamic response capability of the power supply side to changes in the grid operating status. Load-side data includes the adjustable capacity of flexible loads, demand response event status, and thermal inertia parameters of temperature-controlled loads. The adjustable capacity of flexible loads characterizes the power range that the load can adjust without affecting the user's basic energy needs, reflecting the load's potential to participate in system regulation. The demand response event status describes the triggering, execution, or recovery phase of the load during demand response, reflecting the temporal relationship between load-side regulation behavior and dispatch commands. The thermal inertia parameters of temperature-controlled loads characterize the energy storage and release characteristics of temperature-controlled loads such as air conditioners and heat pumps during temperature changes, thus reflecting the lag of load power changes relative to control commands. The grid-side data includes real-time topology, equipment maintenance plans, and safety constraint margins. The real-time topology describes the network connections formed by buses, transmission lines, and switchgear under the current operating conditions, reflecting the transmission path and electrical coupling relationships of electrical energy in the grid. The equipment maintenance plan characterizes the outage, restriction, or restoration arrangements of grid equipment within a predetermined time range, revealing the variability of grid structure and transmission capacity over time. The safety constraint margin quantifies the operating margin retained by the grid under the conditions of stability, thermal stability, and voltage constraints, thereby reflecting the grid's safe carrying capacity in the face of load fluctuations and sudden disturbances.

[0032] A sliding window statistical approach is used for power source-side data to extract ultra-short-term volatility as an uncertainty quantification index, which is used to describe the uncertainty of the grid's demand for reserve capacity and rapid regulation. Specifically, the sliding window statistical approach is used to segment and roll the power output data of the power source in a continuous time series according to a preset time length to characterize the variation amplitude of renewable energy output within adjacent time windows. Ultra-short-term volatility is used to characterize the rapid fluctuation characteristics of power output on a time scale of seconds to minutes, thereby reflecting the degree of impact of random disturbances in renewable energy output on system balance. The uncertainty quantification index is used to transform the instability of power output changes into a calculable numerical feature, which is used to measure the potential demand level of the grid for reserve capacity allocation and rapid regulation resource deployment during operation.

[0033] Furthermore, a demand response event graph is constructed from load-side data to mark the response delay characteristics of interruptible loads and the spatiotemporal transferability of electric vehicle charging, thereby obtaining flexible adjustment demand characteristics oriented towards scheduling instructions. Specifically, the demand response event graph is used to describe the correlation, triggering conditions, and state evolution paths of different load units participating in the demand response process in a graph structure. The response delay characteristics of interruptible loads are used to characterize the time difference required for loads to complete power adjustment after receiving scheduling or control instructions, thereby reflecting the dynamic response capability of load-side adjustment behavior. The spatiotemporal transferability of electric vehicle charging is used to describe the feasibility characteristics of charging loads migrating or delaying between different time periods and different spatial nodes, thus forming flexible adjustment demand characteristics oriented towards scheduling instructions to support refined load control decisions.

[0034] Simultaneously, by correlating the action signals of protection devices with topology change information, the power flow transfer paths and weak links under fault conditions are identified using grid-side data. This allows for the extraction of proactive defensive safety requirements of the power grid in response to load redistribution and reconfiguration. Specifically, protection device action signals reflect the activation, disconnection, or reconfiguration behavior of relay protection and automation devices under abnormal operating conditions; topology change information describes the adjustment of network connection relationships after changes in the status of buses, lines, and switches; correlative analysis uncovers the inherent logical relationship between protection device actions and topology changes to identify the redistribution paths of electrical energy in the network when a fault occurs; power flow transfer paths characterize the process of power transfer from the original transmission channels to other channels under fault or maintenance conditions; and weak links indicate network locations prone to overload, exceeding limits, or stability degradation during power flow reconfiguration, thereby extracting the proactive defensive safety requirements of the power grid during load redistribution and structural reconfiguration.

[0035] A dynamic graph neural network is constructed based on the real-time topology, where the bus is used as a node and the transmission line as an edge. The multi-dimensional demand data is divided into fast-changing components and slow-changing components. The fast-changing components are used to characterize the operational characteristics that change significantly within a short time scale. The fast-changing components include the second- to minute-level fluctuations in renewable energy output, instantaneous adjustment of demand response, and fault transient processes. The renewable energy output fluctuations are used to describe the rapid power change characteristics of power sources such as wind power and photovoltaics caused by environmental disturbances within a very short time scale. The instantaneous adjustment of demand response is used to characterize the immediate power adjustment behavior of the load after receiving dispatch or control signals. The fault transient process is used to reflect the rapid evolution characteristics of the power, voltage, and power flow state of the power grid in the early stage of an abnormal event. One-dimensional convolutional neural network encoding is used to perform local perception and feature extraction on continuous data in the time series dimension. High-frequency change patterns are extracted through convolution kernel sliding operations, thereby forming high-frequency dynamic features that can reflect rapid dynamic characteristics. Slowly varying components are used to characterize operational features that exhibit stable or gradual changes over a longer time scale. These components include hourly temperature cumulative trends, day-ahead power generation plan curves, and equipment maintenance durations. Temperature cumulative trends describe the gradual changes and cumulative effects of ambient temperature over a continuous period. Day-ahead power generation plan curves characterize the trajectory of power generation output arrangements made by the dispatching department during the day-ahead phase over time. Equipment maintenance durations depict the time span by which grid equipment affects network structure and transmission capacity during maintenance. Multi-scale average pooling encoding is used to smooth and aggregate the slowly varying data at different time window scales to extract representative low-frequency trend features, thereby reflecting the long-term evolution of the system's operating state.

[0036] Simultaneously, high-frequency dynamic features and low-frequency trend features are concatenated to obtain multi-scale features, which are then input into a dynamic graph neural network. Feature concatenation is used to combine features extracted from different time scales along the feature dimension to form a unified feature expression that simultaneously contains rapid change information and long-term trend information. Multi-scale features are used to comprehensively reflect the operating status of the power system at different time resolutions. The dynamic graph neural network is used to model and propagate the bus node features under the condition that the power grid topology changes over time, thereby realizing the collaborative learning of multi-scale operating features and network structure information.

[0037] Furthermore, based on the message passing mechanism of the dynamic graph neural network, the feature representation is iteratively updated among neighboring nodes to learn the spatial coupling strength of demand in various dimensions to the bus load. The spatial coupling strength is captured in the time dimension through temporal relationships, and the time delay characteristics of the impact of power supply output changes, load demand response, and grid topology adjustments on the bus load are identified. Among them, the electrical coupling strength and power interaction direction of adjacent buses under the current topology are quantified to generate spatiotemporal attention weights, which are used to reflect the spatial transmission impact of grid topology on each bus load. The message passing mechanism of the dynamic graph neural network is used to propagate node feature information along the edge in the graph structure and realize the aggregation of local information to global representation through multiple rounds of iteration. Neighbor nodes represent the set of buses directly connected by transmission lines in the power grid topology; feature representation describes the state vector formed by nodes under the influence of multi-dimensional demand; spatial coupling strength characterizes the degree of load impact caused by electrical connections between different buses; temporal relationship describes the order and dependency of feature evolution over time; time delay characteristic characterizes the time interval during which different types of demand changes affect bus load; electrical coupling strength quantifies the degree of electrical association between buses under the current topology; power interaction direction indicates the dominant direction of power transmission between adjacent buses; and spatiotemporal attention weight comprehensively reflects the importance distribution of spatial structure and temporal evolution on bus load prediction.

[0038] Furthermore, the input contribution ratios for regulating the uncertainty characteristics of the power supply side, the flexible regulation characteristics of the load side, and the topological constraint characteristics of the grid side are determined respectively, and dimensional gating coefficients are generated to reflect the dominant differences in demand of each dimension under different operating conditions. Based on the historical prediction error distribution and the current data quality assessment, the reliability interval of the output prediction results is quantified, and the prediction confidence is generated. Among them, the input contribution ratio is used to represent the relative weight of different demand characteristics in the model input stage; the uncertainty characteristics of the power supply side are used to describe the instability factors caused by the fluctuation of renewable energy output and prediction deviation; the flexible regulation characteristics of the load side are used to characterize the variable capacity of the load under demand response and flexible regulation conditions; the topological constraint characteristics of the grid side are used to reflect the limiting effect of network structure, operation mode and security constraints on the formation of power distribution; the dimensional gating coefficients are used to dynamically adjust the influence of different dimensional demand characteristics within the model; the historical prediction error distribution is used to statistically analyze the prediction deviation characteristics of the model in past operating samples; the data quality assessment is used to measure the credibility level of the current input data in terms of completeness, accuracy and timeliness; the reliability interval is used to describe the statistically significant fluctuation range of the prediction results; and the prediction confidence is used to quantify the degree to which the prediction results can be trusted. Based on a multi-dimensional demand coordination coefficient that includes spatiotemporal attention weights, dimensional gating coefficients, and prediction confidence, a spatiotemporal correlation model is constructed, thereby forming a multi-dimensional demand coordination coefficient that is used to construct a spatiotemporal correlation model between demand and bus load.

[0039] The spatiotemporal correlation model is used to analyze and correlate the collected multi-dimensional demand data, and output the multi-dimensional demand coordination coefficient.

[0040] Specifically, a spatiotemporal correlation model is used to analyze the collected multi-dimensional demand data and output multi-dimensional demand coordination coefficients. The spatiotemporal correlation model is used to characterize the coupling relationship between multi-dimensional demand characteristics and bus load in the power system in terms of spatial structure and temporal evolution. The multi-dimensional demand data is used to describe different dimensions of operational demand information, such as power source side operational uncertainty, load side regulation behavior, and grid side structure and security constraints. The correlation analysis is used to calculate and infer the interaction relationship between different demand dimensions based on the model learning results, so as to identify the path, intensity, and sequence of the impact of each demand characteristic on the bus load. The multi-dimensional demand coordination coefficients are used to comprehensively express the relative dominance, coordination relationship, and regulation effect of each dimension of demand under the current operating conditions in a parametric form.

[0041] The multi-dimensional demand coordination coefficient is used as the input correction term of the long short-term memory network in the bus load time series prediction model. By adjusting the weight distribution of each input feature, the bus load is predicted, and the bus load prediction result is obtained.

[0042] Furthermore, this application also includes: using the spatial attention weight as the adjustment threshold of the forget gate, reducing the retention ratio of historical load memory states when a grid-side topology reconfiguration event is detected, and enhancing the response sensitivity to real-time topology changes; using the dimensional gating coefficient as the weighting coefficient of the input gate, increasing the input weight of the power supply side uncertainty characteristics when the power supply side renewable energy output fluctuation rate exceeds a preset threshold, and reducing the interference of historical concurrent load patterns; and, based on the time delay characteristics, selecting the time step for the multi-dimensional demand characteristics of the input long short-term memory network, and selecting the characteristic data corresponding to the historical moment according to the different delay characteristics of the power supply side output change, load side demand response, and grid-side topology adjustment on the bus load, and compensating for the spatiotemporal transmission lag.

[0043] Furthermore, this application also includes: based on time delay characteristics, performing time step selection and time alignment on the input features, mapping demand feature data with different delay characteristics to a unified time section, and obtaining a spatiotemporally decoupled synchronous input feature vector; based on the synchronous input feature vector, combining the corrected forget gate and input gate output, updating the internal memory unit state, and obtaining a comprehensive state containing topology awareness and demand coordination information; selectively extracting information from the memory unit state through the output gate to generate a hidden state vector at the current moment, the hidden state vector including multi-dimensional demand spatiotemporal correlation features corrected by the coordination coefficient; inputting the hidden state vector into a fully connected decoding layer, and through linear mapping and nonlinear activation processing, outputting the load prediction value of the target bus at the next moment, and fusing the prediction confidence coefficient with the historical prediction error statistical distribution to generate the standard deviation interval of the load point prediction value, and outputting the bus load prediction result containing the point prediction value and the confidence interval.

[0044] Furthermore, this application also includes: introducing power grid physical constraints during the training process of the long short-term memory network, constructing a power flow residual calculation layer, wherein the power flow residual calculation layer is constructed based on Kirchhoff's current law, comparing the difference between the bus load prediction results and the power balance calculated based on the power grid topology; adding the power flow residual as a regularization term to the loss function, adjusting the long short-term memory network parameters through backpropagation, so that the output bus load prediction value satisfies the power balance constraint of the entire network, and avoiding conflicts between the single bus prediction results and the physical laws of the power grid.

[0045] Specifically, spatial attention weights are used as the adjustment threshold for the forget gate. When a grid-side topology reconfiguration event is detected, the retention ratio of historical load memory states is reduced to enhance the sensitivity to real-time topology changes. Specifically, spatial attention weights quantify the impact of different bus nodes and their interconnections on load evolution under the current topology; the forget gate controls the retention rate of historical state information in the long short-term memory network at the current moment; the adjustment threshold limits the proportion of historical memory information participating in subsequent calculations; grid-side topology reconfiguration events describe changes in network connectivity due to fault isolation, line switching, or operational mode adjustments; and historical load memory states characterize the historical evolution features of bus loads stored within the model. By reducing the retention ratio of historical load memory states, the model reduces its dependence on load patterns under existing topology conditions, thereby improving its ability to perceive and respond to real-time topology changes.

[0046] Furthermore, the dimensionality gating coefficient is used as the weighting coefficient of the input gate. When the power output volatility of renewable energy sources exceeds a preset threshold, the input weight of the uncertainty features of the power supply side is increased to reduce the interference of historical load patterns. Specifically, the dimensionality gating coefficient is used to characterize the relative dominance of different demand dimensions in the influence of the bus load under the current operating conditions; the input gate is used to control the proportion of external feature information entering the internal state of the long short-term memory network; the weighting coefficient is used to assign weights to features from different sources; the renewable energy output volatility is used to reflect the drastic degree of power output change of renewable energy sources in a short time scale; and the preset threshold is used to define the judgment condition for renewable energy output to change from a stable state to a high uncertainty state. By increasing the input weight of the uncertainty features of the power supply side, the model pays more attention to real-time output changes when renewable energy fluctuations are significant, thereby suppressing the misleading influence of historical load patterns on the prediction results.

[0047] Meanwhile, based on the time delay characteristics, time step selection is performed on the multi-dimensional demand features input to the Long Short-Term Memory (LSTM) network. According to the different delay characteristics of the impact of power supply output changes, load demand response, and grid topology adjustments on the bus load, feature data corresponding to the historical time are selected to compensate for the spatiotemporal transmission lag. Among them, the time delay characteristics are used to describe the time interval from the occurrence of different demand changes to their significant impact on the bus load; the time step selection is used to determine the feature sampling points most relevant to the current prediction time in the historical sequence; the multi-dimensional demand features are used to characterize the comprehensive information of power supply operating status, load adjustment behavior, and grid structural changes; by selecting historical time feature data that matches the delay of various demand impacts, the compensation processing for the time-dimensional transmission lag of demand changes is realized, thereby improving the model's accuracy in depicting the dynamic evolution of the load.

[0048] Based on the time delay characteristics, time step selection and time alignment are performed on the input features to map demand feature data with different delay characteristics to a unified time profile, obtaining a synchronous input feature vector after spatiotemporal decoupling. Here, the time delay characteristics describe the different time intervals during which changes in power supply output, load response behavior, and grid-side structural adjustments affect the bus load; time step selection determines the feature sampling time most relevant to the prediction target in the historical sequence; time alignment performs time-series correction on demand features with different lags; the unified time profile represents the alignment status of multi-source demand features at the same prediction reference time; spatiotemporal decoupling separates the coupling effects of time lag and spatial transmission of demand features; and the synchronous input feature vector carries the multi-dimensional demand information after time alignment, serving as the input basis for subsequent model calculations.

[0049] Furthermore, based on the synchronous input feature vector, combined with the corrected forget gate and input gate output, the internal memory cell state is updated to obtain a comprehensive state containing topology awareness and demand coordination information. Specifically, the forget gate controls the proportion of historical state information retained at the current moment; the input gate controls the intensity of the synchronous input feature vector entering the memory cell; the corrected forget gate and input gate output reflect the moderating effect of spatial attention weights and dimensional gating coefficients on the memory update process; the internal memory cell state stores load evolution information across multiple time scales; topology awareness reflects the impact of grid structure changes on load distribution; and demand coordination information reflects the comprehensive effect of multi-dimensional demand under the current operating conditions.

[0050] Simultaneously, selective information extraction of the memory cell state is performed through the output gate to generate the hidden state vector at the current moment. The hidden state vector includes multi-dimensional spatiotemporal correlation features of demand after being corrected by the coordination coefficient. The output gate is used to control the proportion of effective information in the memory cell being transmitted to the external state; selective information extraction is used to highlight the state components that contribute more to the prediction target; the hidden state vector is used to represent the intermediate expression results formed by the model at the current moment; and the multi-dimensional spatiotemporal correlation features of demand are used to characterize the comprehensive influence pattern of demand features in spatial structure and temporal evolution.

[0051] Furthermore, the hidden state vector is input into the fully connected decoding layer. Through linear mapping and nonlinear activation processing, the load prediction value of the target bus at the next time step is output. The prediction confidence coefficient is fused with the historical prediction error statistical distribution to generate the standard deviation interval of the load point prediction value. The bus load prediction result containing the point prediction value and the confidence interval is output. Here, the fully connected decoding layer is used to realize the mapping of the hidden state vector to the prediction target space; the linear mapping is used to establish the functional relationship between state characteristics and load values; the nonlinear activation processing is used to enhance the model's ability to express complex load change patterns; the prediction confidence coefficient is used to quantify the reliability of the prediction result; the historical prediction error statistical distribution is used to describe the error characteristics of the model in previous samples; and the standard deviation interval is used to characterize the range of possible fluctuations in the prediction value.

[0052] Furthermore, power grid physical constraints are introduced during the training of the Long Short-Term Memory network to construct a power flow residual calculation layer. This layer is based on Kirchhoff's Current Law and is used to compare the differences between the bus load prediction results and the power balance calculated based on the power grid topology. Specifically, the power grid physical constraints are used to ensure that the prediction results conform to the basic laws of power system operation; Kirchhoff's Current Law is used to describe the conservation relationship of node currents; and the power flow residual is used to quantify the degree of deviation between the predicted load state and the physical equilibrium state.

[0053] Furthermore, the power flow residual is added as a regularization term to the loss function, and the parameters of the long short-term memory network are adjusted through backpropagation to ensure that the output bus load prediction values ​​meet the power balance constraints of the entire network, thus avoiding conflicts between the prediction results of a single bus and the physical laws of the power grid. Here, the regularization term is used to introduce physical consistency constraints during model training; the loss function is used to measure the deviation between the prediction results and the actual load; backpropagation is used to update the network parameters; and the power balance constraints of the entire network are used to ensure that the prediction results meet the power conservation requirements at the system level.

[0054] In summary, the bus load forecasting method based on multi-dimensional demand coordination of the power system provided in this application has the following technical effects: by achieving the technical goal of spatiotemporal correlation modeling and dynamic forecasting of bus load based on multi-dimensional demand coordination, it can accurately capture the impact of power supply side output fluctuations, load side response characteristics and grid side topology adjustments on bus load, thereby improving forecast accuracy, real-time response and support for grid operation and dispatching decisions.

[0055] Example 2: Based on the same inventive concept as the bus load forecasting method based on multi-dimensional demand coordination of the power system in the foregoing examples, this application also provides a bus load forecasting device based on multi-dimensional demand coordination of the power system. Please refer to the appendix. Figure 2The system includes: a spatiotemporal correlation model construction module 1, used to collect multi-dimensional demand data from the power source side, load side, and grid side of the power system, and to construct a spatiotemporal correlation model between multi-dimensional demand and bus load using a graph neural network to reflect the influence weight of each dimension of demand on the bus load and the spatiotemporal transmission rules; a multi-dimensional demand coordination coefficient output module 2, used to perform correlation analysis on the collected multi-dimensional demand data through the spatiotemporal correlation model and output multi-dimensional demand coordination coefficients; and a bus load prediction result acquisition module 3, used to use the multi-dimensional demand coordination coefficients as input correction terms of the long short-term memory network of the bus load time series prediction model, and to perform bus load prediction by adjusting the weight distribution of each input feature to obtain the bus load prediction result.

[0056] Furthermore, the bus load forecasting equipment based on multi-dimensional demand coordination of the power system is also used for: power source side data including the fluctuation rate of new energy output, day-ahead forecast deviation distribution and inverter control mode; load side data including the adjustable capacity of flexible load, demand response event status and thermal inertia parameters of temperature-controlled load; grid side data including real-time topology, equipment maintenance plan and safety constraint margin.

[0057] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used for: using sliding window statistics on power source side data to extract ultra-short-term volatility as an uncertainty quantification index to describe the uncertainty index of the power grid's demand for reserve capacity and rapid adjustment; constructing a demand response event map on load side data to mark the response delay characteristics of interruptible loads and the spatiotemporal transferability attributes of electric vehicle charging, thereby obtaining flexible adjustment demand characteristics oriented towards dispatch instructions; and identifying power flow transfer paths and weak links under fault conditions by correlating and analyzing protection device action signals and topology change information on power grid side data, thereby extracting the proactive defensive security demand characteristics of the power grid for load distribution and reconfiguration.

[0058] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used to: construct a dynamic graph neural network based on the real-time topology, wherein the bus is used as a node and the transmission line is used as an edge, and the multi-dimensional demand data is input into the dynamic graph neural network as node features; based on the dynamic graph neural network, the spatial influence relationship and time transmission lag characteristics of each dimension of demand on the bus load are learned through a message passing mechanism based on the historical sample set, and a multi-dimensional demand coordination coefficient including spatiotemporal attention weights, dimensional gating coefficients and prediction confidence is generated to construct the spatiotemporal correlation model.

[0059] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used for: iteratively updating feature representations among neighboring nodes based on the message passing mechanism of a dynamic graph neural network, learning the spatial coupling strength of demand in each dimension to the bus load, and capturing the lag relationship of spatial coupling strength in the time dimension through temporal relationships, identifying the time delay characteristics of the impact of power supply output changes, load demand response, and grid topology adjustments on the bus load. Specifically, it quantifies the electrical coupling strength and power interaction direction of adjacent buses under the current topology, generating the spatiotemporal attention weights to reflect the spatial transmission impact of grid topology on each bus load; determines the input contribution ratios of regulating power supply uncertainty characteristics, load elastic regulation characteristics, and grid topology constraint characteristics, generating the dimensional gating coefficients to reflect the dominant differences in demand in each dimension under different operating conditions; and quantifies the reliability range of the output forecast results based on historical forecast error distribution and current data quality assessment to generate the forecast confidence level.

[0060] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used to: use the spatial attention weight as the adjustment threshold of the forget gate, and when a grid-side topology reconfiguration event is detected, reduce the retention ratio of historical load memory states to enhance the response sensitivity to real-time topology changes; use the dimensional gating coefficient as the weighting coefficient of the input gate, and when the power generation fluctuation rate of new energy sources exceeds a preset threshold, increase the input weight of power generation uncertainty features to reduce the interference of historical concurrent load patterns; and based on the time delay characteristics, perform time step selection on the multi-dimensional demand features input to the long short-term memory network, and select feature data corresponding to the historical moment according to the different delay characteristics of the impact of power generation output changes, load-side demand response, and grid-side topology adjustments on the bus load, to compensate for the spatiotemporal transmission lag.

[0061] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used for: selecting and aligning the input features based on time delay characteristics, mapping demand feature data with different delay characteristics to a unified time section, and obtaining a synchronous input feature vector after spatiotemporal decoupling; updating the internal memory unit state based on the synchronous input feature vector, combined with the corrected forget gate and input gate output, to obtain a comprehensive state containing topology sensing and demand coordination information; selectively extracting information from the memory unit state through the output gate to generate a hidden state vector at the current moment, wherein the hidden state vector includes multi-dimensional demand spatiotemporal correlation features corrected by the coordination coefficient; inputting the hidden state vector into a fully connected decoding layer, and outputting the load forecast value of the target bus at the next moment through linear mapping and nonlinear activation processing, and fusing the prediction confidence coefficient with the historical prediction error statistical distribution to generate the standard deviation interval of the load point prediction value, and outputting the bus load forecast result containing the point prediction value and the confidence interval.

[0062] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used to: introduce power grid physical constraints during the training process of the long short-term memory network, construct a power flow residual calculation layer, which is constructed based on Kirchhoff's current law, compare the bus load forecasting results with the power balance calculated based on the power grid topology; add the power flow residual as a regularization term to the loss function, and adjust the long short-term memory network parameters through backpropagation so that the output bus load forecasting value meets the power balance constraints of the entire network, avoiding conflicts between the single bus forecasting results and the physical laws of the power grid.

[0063] Furthermore, the bus load forecasting device based on multi-dimensional demand coordination of the power system is also used to: divide the multi-dimensional demand data into fast-changing components and slow-changing components, wherein the fast-changing components include second- to minute-level fluctuations in new energy output, instantaneous adjustment of demand response, and fault transient processes, and high-frequency dynamic features extracted by one-dimensional convolutional neural network encoding; the slow-changing components include hourly temperature accumulation trends, day-ahead power generation plan curves, and equipment maintenance durations, and low-frequency trend features extracted by multi-scale average pooling encoding; the high-frequency dynamic features and low-frequency trend features are concatenated to obtain multi-scale features, which are then input into the dynamic graph neural network.

[0064] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The bus load forecasting method and specific examples based on multi-dimensional demand coordination of the power system in the foregoing embodiment 1 are also applicable to the bus load forecasting device based on multi-dimensional demand coordination of the power system in this embodiment. Through the foregoing detailed description of the bus load forecasting method based on multi-dimensional demand coordination of the power system, those skilled in the art can clearly understand the bus load forecasting device based on multi-dimensional demand coordination of the power system in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0066] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A bus load forecasting method based on multi-dimensional demand coordination in power systems, characterized in that, include: Collect multi-dimensional demand data from the power source side, load side, and grid side of the power system, and use graph neural networks to construct a spatiotemporal correlation model between multi-dimensional demand and bus load to reflect the influence weight of each dimension of demand on bus load and the spatiotemporal transmission rules. The spatiotemporal correlation model is used to perform correlation analysis on the collected multi-dimensional demand data and output the multi-dimensional demand coordination coefficient. The multi-dimensional demand coordination coefficient is used as the input correction term of the long short-term memory network in the bus load time series prediction model. By adjusting the weight distribution of each input feature, the bus load is predicted, and the bus load prediction result is obtained.

2. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 1, characterized in that, Power supply side data includes the volatility of renewable energy output, the distribution of day-ahead forecast deviations, and inverter control modes; Load-side data includes the adjustable capacity of flexible loads, demand response event status, and thermal inertia parameters of temperature-controlled loads; Grid-side data includes real-time topology, equipment maintenance plans, and safety constraint margins.

3. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 2, characterized in that, Collect multi-dimensional demand data from the power system's power source side, load side, and grid side, including: Sliding window statistics are used to analyze power supply side data and extract ultra-short-term volatility as an uncertainty quantification indicator to describe the uncertainty of the power grid's demand for reserve capacity and rapid regulation. A demand response event map is constructed from load-side data, marking the response delay characteristics of interruptible loads and the spatiotemporal transferability of electric vehicle charging, thereby obtaining the flexible adjustment demand characteristics oriented towards scheduling instructions. By correlating and analyzing the action signals of protection devices with topology change information, we can identify the power flow transfer paths and weak links under fault conditions, and extract the proactive defensive security requirements of the power grid for load distribution and reconfiguration.

4. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 3, characterized in that, A spatiotemporal correlation model between multidimensional demand and bus load is constructed using graph neural networks, including: A dynamic graph neural network is constructed based on the real-time topology, wherein the bus is used as a node and the transmission line is used as an edge, and the multi-dimensional demand data is input into the dynamic graph neural network as node features. Based on a dynamic graph neural network, the spatial impact relationship and time transmission lag characteristics of demand in various dimensions on bus load are learned through a message passing mechanism based on historical sample sets. This generates a multi-dimensional demand coordination coefficient that includes spatiotemporal attention weights, dimensional gating coefficients, and prediction confidence, and constructs the spatiotemporal correlation model.

5. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 4, characterized in that, Generate a multi-dimensional demand reconciliation coefficient that includes spatiotemporal attention weights, dimensional gating coefficients, and prediction confidence, including: Based on the message passing mechanism of dynamic graph neural network, the feature representation is iteratively updated between neighboring nodes to learn the spatial coupling strength of demand in various dimensions to the bus load. The spatial coupling strength is captured in the time dimension through temporal relationship, and the time delay characteristics of the impact of power supply output change, load demand response and grid topology adjustment on the bus load are identified. Among them, the electrical coupling strength and power interaction direction of adjacent buses under the current topology are quantified to generate the spatiotemporal attention weight, which is used to reflect the spatial transmission impact of grid topology on each bus load. The input contribution ratios of the uncertainty characteristics of the power supply side, the elastic regulation characteristics of the load side, and the topological constraint characteristics of the grid side are determined respectively, and the dimensional gating coefficients are generated to reflect the dominant differences in the demand of each dimension under different operating conditions. Based on the historical prediction error distribution and the current data quality assessment, the reliability range of the output prediction results is quantified, and the prediction confidence is generated.

6. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 5, characterized in that, The multi-dimensional demand coordination coefficient is used as an input correction term for the long short-term memory network of the bus load time series prediction model, including: The spatial attention weight is used as the adjustment threshold of the forget gate. When a grid-side topology reconfiguration event is detected, the retention ratio of historical load memory states is reduced, thereby enhancing the response sensitivity to real-time topology changes. The dimensional gating coefficient is used as the weighting coefficient of the input gate. When the power output fluctuation rate of new energy sources on the power supply side exceeds the preset threshold, the input weight of the uncertainty characteristics on the power supply side is increased, and the interference of historical load patterns is reduced. Based on the time delay characteristics, time steps are selected for the multi-dimensional demand characteristics of the input long short-term memory network. According to the different delay characteristics of the impact of power supply output changes, load demand response and grid topology adjustment on bus load, feature data corresponding to the historical moment are selected to compensate for the spatiotemporal transmission lag.

7. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 6, characterized in that, Obtain bus load forecast results, including: Based on the time delay characteristics, time step selection and time alignment are performed on the input features to map the required feature data with different delay characteristics to a unified time section, thereby obtaining a synchronous input feature vector after spatiotemporal decoupling. Based on the synchronous input feature vector, combined with the corrected forget gate and input gate output, the internal memory unit state is updated to obtain a comprehensive state containing topological awareness and demand coordination information. The output gate selectively extracts information from the state of the memory cell to generate a hidden state vector at the current moment. The hidden state vector includes multi-dimensional spatiotemporal correlation features of demand after being corrected by the coordination coefficient. The hidden state vector is input into the fully connected decoding layer. Through linear mapping and nonlinear activation processing, the load prediction value of the target bus at the next time moment is output. The prediction confidence coefficient is fused with the historical prediction error statistical distribution to generate the standard deviation interval of the load point prediction value. The bus load prediction result containing the point prediction value and the confidence interval is output.

8. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 7, characterized in that, Also includes: During the training process of the Long Short-Term Memory Network, power grid physical constraints are introduced to construct a power flow residual calculation layer. The power flow residual calculation layer is constructed based on Kirchhoff's Current Law. The difference between the bus load prediction results and the power balance calculated based on the power grid topology is compared. The power flow residual is added as a regularization term to the loss function, and the parameters of the long short-term memory network are adjusted through backpropagation so that the output bus load prediction value meets the power balance constraint of the whole network, thus avoiding the conflict between the single bus prediction result and the physical law of the power grid.

9. The bus load forecasting method based on multi-dimensional demand coordination of power systems according to claim 4, characterized in that, Before inputting the multi-dimensional demand data as node features into the dynamic graph neural network, the process further includes: The multi-dimensional demand data is divided into fast-changing components and slow-changing components. The fast-changing components include second- to minute-level fluctuations in new energy output, instantaneous adjustment of demand response, and transient fault processes, which are high-frequency dynamic features extracted by one-dimensional convolutional neural network encoding. The slow-changing components include hourly temperature accumulation trends, day-ahead power generation plan curves, and equipment maintenance durations, which are low-frequency trend features extracted by multi-scale average pooling encoding. The high-frequency dynamic features and low-frequency trend features are concatenated to obtain multi-scale features, which are then input into the dynamic graph neural network.

10. A bus load forecasting device based on multi-dimensional demand coordination of power systems, characterized in that, The steps for implementing the bus load forecasting method based on multi-dimensional demand coordination of power systems according to any one of claims 1 to 9 include: The spatiotemporal correlation model construction module is used to collect multi-dimensional demand data from the power source side, load side, and grid side of the power system. It uses graph neural networks to construct a spatiotemporal correlation model between multi-dimensional demand and bus load, which reflects the influence weight of each dimension of demand on the bus load and the spatiotemporal transmission rules. The multi-dimensional demand coordination coefficient output module is used to perform correlation analysis on the collected multi-dimensional demand data through the spatiotemporal correlation model and output the multi-dimensional demand coordination coefficient. The bus load forecasting result acquisition module is used to use the multi-dimensional demand coordination coefficient as the input correction term of the long short-term memory network of the bus load time series forecasting model, and to perform bus load forecasting by adjusting the weight distribution of each input feature, thereby obtaining the bus load forecasting result.

Images