Intelligent power distribution load prediction and adaptive scheduling method

Abstract

The invention provides an intelligent power distribution load prediction and adaptive scheduling method, and aims to improve the response capability and operation efficiency of a power distribution system to load fluctuation. According to the method, historical load data, real-time meteorological information and user power consumption behavior data of a target power distribution area are collected, a multi-dimensional input data set is established, a short-term load prediction model is constructed based on the data set, and prediction of future load levels of multiple feeder lines and key nodes is achieved. According to a prediction result, potential voltage deviation, load imbalance and equipment overload risks are identified, and a dynamic scheduling control strategy is generated by adopting a rolling horizon optimization method in combination with a power grid topological structure and operation constraints; the strategy comprises control measures such as transformer tap position adjustment and capacitor bank switching; through issuing a scheduling strategy to an equipment control unit, active adjustment is completed before load change, voltage stabilization, load balancing and energy optimization are realized, and intelligent and adaptive operation guarantee is provided for a power distribution system.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power dispatching, and in particular to an intelligent power distribution load prediction and adaptive dispatching method. Background Art

[0002] At present, the dispatching of distribution systems mainly relies on static rules or experience-based control strategies. Equipment adjustments are usually made based on real-time load monitoring results or preset threshold conditions, such as transformer tap adjustment, capacitor switching, and feeder load reconstruction. In some advanced systems, short-term load forecasting technology has also been introduced, using historical load data for trend estimation to provide a certain auxiliary reference for dispatching. However, this type of forecasting is mostly based on a single data source, has low spatial resolution, and fails to form a closed-loop linkage with equipment dispatching. In addition, existing dispatching strategies often use fixed time intervals for adjustments, which makes it difficult to respond quickly to sudden load changes.

[0003] With the increase in the access to distributed energy, the increasing complexity of user load behavior, and the increasing impact of meteorological conditions on grid operation, existing technologies are unable to cope with high-frequency and large-scale load fluctuations. Specifically, the current distribution system has a lag in voltage control, load balancing, and energy storage scheduling, which makes the system prone to voltage over-limit, equipment overload, or uneven load distribution during peak hours, affecting the safety and economy of power supply. In addition, the lack of a dynamic response mechanism to user-side behavior also prevents demand-side resources from effectively participating in system regulation, further limiting the flexibility and accuracy of scheduling.

[0004] Therefore, there is an urgent need for an intelligent scheduling method that can integrate multi-source data, realize prediction-driven and have adaptive capabilities, so as to improve the active response capability and operation optimization level of the distribution system under load fluctuation scenarios. Summary of the invention

[0005] The present application provides an intelligent power distribution load prediction and adaptive scheduling method to improve the responsiveness and operating efficiency of the power distribution system to load fluctuations.

[0006] The present application provides an intelligent power distribution load prediction and adaptive scheduling method, including: Obtain historical load data, real-time meteorological information, and user electricity consumption behavior data in the target distribution area, and establish a comprehensive load forecast input data set; Based on the comprehensive load forecast input data set, a short-term load forecast model is constructed to forecast the load levels of multiple feeders and key nodes in a preset future period, and obtain the forecast load time series at the node level and feeder level; According to the predicted load time sequence, determining the voltage deviation, load imbalance or equipment overload risk that may occur at each node and feeder within the predicted period; Based on the grid topology, electrical parameter constraints and load forecast timing, a rolling time domain optimization method is used to generate a dispatch control strategy, which includes the early adjustment of transformer tap positions, the advance switching plan of capacitor banks, the reclosing sequence and load distribution logic of feeder tie switches, the charging and discharging time and power setting of distributed energy storage devices, or the demand response execution instructions for the user side; The dispatching control strategy is sent to the corresponding device control unit, and voltage control, load reconstruction, energy regulation or user load response is performed before the predicted time period according to the dispatching control strategy.

[0007] The beneficial effects of the technical solution provided by this application include: (1) By integrating multi-source information such as historical load, real-time weather and user behavior, short-term forecasting can be carried out to grasp the load trend in the future period in advance, so that the system can take active control measures before load anomalies occur, avoiding the operational risks caused by passive response. (2) Based on the prediction results, the possible operational risks are judged, and combined with the distribution network structure and equipment constraints, an optimized dispatching strategy with strong operability is generated, which helps to stabilize node voltage, alleviate feeder congestion, and improve the operational safety and reliability of the system. (3) The dispatching strategy covers a variety of resources such as transformers, capacitors, feeder interconnection switches, energy storage devices and user load response, supports multi-objective and cross-level comprehensive control, and significantly improves the refinement level of resource allocation and energy management of the distribution system. (4) Through rolling optimization and dynamic execution mechanism, the dispatching strategy is continuously matched with real-time load changes, the energy loss caused by peak-to-valley difference is reduced, and the load balancing and energy efficiency management capabilities of the system are enhanced, thereby providing solid support for the construction of smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flow chart of an intelligent power distribution load prediction and adaptive scheduling method provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0009] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0010] The first embodiment of the present application provides a method for intelligent power distribution load prediction and adaptive scheduling. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 The first embodiment of the present application provides a method for intelligent power distribution load prediction and adaptive scheduling, which is described in detail.

[0011] Step S101: Acquire historical load data, real-time meteorological information and user power consumption behavior data in the target distribution area, and establish a comprehensive load forecast input data set.

[0012] Step S101 involves the acquisition and integration of historical load data, real-time meteorological information, and user electricity consumption behavior data in the target distribution area, aiming to establish a multi-dimensional, time-series continuous, and semantically complete comprehensive load forecast input data set, providing a high-quality input basis for subsequent short-term load forecasting and optimal scheduling.

[0013] First, for the acquisition of historical load data, the system retrieves the load data recorded by distributed measurement devices (such as smart meters, distribution terminal units DTU or feeder monitoring units FTU) from each key feeder and node in the target distribution area. The load data includes at least active power, reactive power, voltage, current and its corresponding timestamp. In order to ensure the spatial accuracy of the data, each set of data needs to be bound to the corresponding physical location coding information, such as feeder number, node number or geographic coordinates. The acquired data should cover a continuous period in the recent period (such as the past 7 days, 30 days or 90 days) to reflect the load change pattern under different climates and user behaviors. In the data preprocessing stage, the system cleans the historical data, including eliminating invalid records generated during communication failures, equipment anomalies or power outages, and using interpolation, moving average or repair algorithms based on time series models to correct missing or abnormally fluctuating data points to construct a historical load sample set with good time series continuity.

[0014] Secondly, real-time meteorological information can be obtained by accessing the national meteorological platform, third-party meteorological service interfaces, or locally deployed environmental monitoring devices. The collected meteorological parameters should at least include ambient temperature, relative humidity, wind speed, wind direction, atmospheric pressure, sunshine intensity, and weather conditions (such as sunny, cloudy, rainy, snowy, etc.). Since the power supply area of ​​the distribution network may span multiple microclimate sub-areas, in order to ensure the geographical consistency between meteorological data and load forecasting targets, the system implements spatial interpolation processing on meteorological data, and uses geographic statistical methods such as Kriging interpolation and inverse distance weighting to map sparse meteorological observation point data to the service range of each feeder or node in the distribution network, thereby obtaining a set of meteorological factors with spatial labels. During the processing process, meteorological change trend indicators (such as the temperature change slope in the next three hours) can also be introduced as auxiliary features of load forecasting to enhance the model's perception of load fluctuations under sudden weather events.

[0015] To obtain the user's electricity consumption behavior data, the system accesses the smart meter, building energy management system, home energy management system (HEMS) or commercial customer energy consumption collection platform to obtain the user's high-frequency electricity consumption records, including electricity consumption every 15 minutes or every hour, voltage level, electrical appliance operation status, demand data, etc. To achieve behavioral modeling, the system divides users into multiple categories, such as residential users, commercial users, small industrial users, etc., and then extracts typical load characteristics based on the time pattern in historical data, including daytime peak and valley periods, differences between weekdays and holidays, load curve stability, and the degree of response to electricity price changes. Clustering algorithms (such as K-means, DBSCAN, etc.) are used to form user portraits with similar behavioral characteristics. In addition, external variables can be introduced, such as whether the user participates in demand response, the installation of photovoltaic power generation or energy storage devices, etc., to enhance the semantic level of input data.

[0016] After preprocessing and feature extraction, the above three types of data are uniformly formatted and fused into a structured comprehensive input data set. The data set is constructed into a sliding window structure of continuous time periods with time as the main axis. Each forecast period generates a composite sample containing load characteristics, meteorological characteristics and user behavior characteristics to drive the subsequent short-term load forecasting model. The system has the ability to dynamically update, that is, with the continuous access of real-time data, the data set can be updated at a fixed time step (for example, every 5 minutes or 15 minutes) to ensure that the forecast input is consistent with the current system status.

[0017] In summary, step S101 provides a complete path for data acquisition, cleaning, spatial matching, behavioral modeling and input structure construction, forming a load forecasting input data preparation mechanism that integrates multi-source heterogeneous information, and provides accurate, real-time and highly context-related data support for the intelligent distribution scheduling strategy of the present invention.

[0018] Furthermore, the acquisition of historical load data, real-time meteorological information and user electricity consumption behavior data in the target distribution area and the establishment of a comprehensive load forecast input data set include: Obtain historical data of active and reactive loads from multiple monitoring points within a time window, classify them by feeder and node structure partitions, and perform missing value completion, anomaly elimination, and trend stability analysis to form a basic load sample set with complete time series; Receive and integrate real-time meteorological information from multiple sources, including temperature, humidity, wind speed, sunshine intensity and weather type, map meteorological data to the coverage area of ​​each feeder through spatial interpolation algorithm, and construct a meteorological feature set with geographical annotations; Collect user-side electricity consumption behavior information, including periodic electricity consumption patterns, peak and valley electricity consumption response habits, electricity price sensitivity, and equipment operation logs, perform cluster modeling based on user categories, and generate a behavior label feature matrix; Based on the load sample set, meteorological feature set and behavior label feature matrix, a unified multi-dimensional input data structure is constructed, and the data set is dynamically updated according to the rolling forecast cycle to drive the subsequent short-term load forecasting model.

[0019] This embodiment constructs a multi-source fusion input data set with temporal continuity, spatial consistency and semantic identifiability through data cleaning, feature extraction, spatial mapping and behavior modeling, which is used to support the training and real-time deduction of the short-term load forecasting model.

[0020] First, in terms of load data processing, the system collects historical load data within a specified time window from multiple monitoring points in the target distribution area. The data should include two dimensions: active power (P) and reactive power (Q), and have a unified timestamp tag. To ensure the spatial analyzability of the data, the system partitions and classifies each monitoring point according to the location of the node in the feeder and network topology to which it belongs, thereby forming a preliminary load database for hierarchical management. On this basis, data cleaning operations are performed on the collected raw data. First, missing values ​​are identified and filled. The preferred completion methods include moving average based on time series, interpolation algorithms, or weighted mapping of adjacent node data. Secondly, the system uses statistical detection and machine learning models to identify outliers in the data, such as load jumps caused by sensor failure, communication packet loss, or human error, and removes or replaces them. Finally, trend stability analysis is performed on the processed load data to determine the fluctuation characteristics, periodicity, and abnormal mutation frequency of the data within the selected time window, ensuring that the constructed basic sample has a stable time series structure and can be effectively adopted by downstream models.

[0021] Secondly, in terms of meteorological information acquisition and processing, the system needs to synchronously receive real-time meteorological data from multiple sources, which can include national meteorological services, third-party commercial API interfaces, or locally deployed micro-meteorological stations. The meteorological parameters that need to be collected include but are not limited to ambient temperature, relative humidity, wind speed, sunshine intensity, and weather type (such as sunny, rainy, snowy, cloudy, etc.). Since the spatial distribution of meteorological observation points does not necessarily coincide with the distribution network nodes, the system uses geospatial interpolation algorithms (such as inverse distance weighting, Kriging interpolation, or Voronoi diagram methods) to map meteorological data to the service areas of each feeder and the periphery of key nodes to form a meteorological feature set with spatial identification. Each prediction unit (such as a feeder segment or transformer node) will be associated with a set of meteorological features that have been interpolated and smoothed to ensure that the prediction model input is geographically consistent.

[0022] Third, in terms of user electricity consumption behavior modeling, the system needs to collect fine-grained electricity consumption data from the user side, and the data sources include smart meters, building automation systems, industrial control systems or user energy consumption management platforms. The collected content covers periodic electricity consumption patterns (such as daily electricity consumption time periods), peak and valley response behaviors (such as whether there is load reduction), sensitivity to electricity price fluctuations (such as whether there is the ability to automatically postpone non-critical loads when electricity prices change), and terminal equipment operation logs (such as the frequency and duration of turning on equipment such as air conditioners, water heaters, and charging piles). The system clusters behaviors according to the type of user (such as residential, commercial, light industrial, agricultural users, etc.), and uses algorithms such as K-means, DBSCAN or GMM to build a category model, further extracting user behavior labels that are representative in load forecasting, and forming a structured behavior label feature matrix to describe the behavioral driving force behind user load changes.

[0023] The behavior label feature matrix refers to a structured data set organized in the form of a two-dimensional array or tensor, which is used to quantitatively describe the key features of different categories of users in terms of electricity consumption behavior. Each row of the matrix corresponds to an individual user or user group, and each column corresponds to a behavior feature dimension, which is used to reflect the performance value or category label of the user on a certain behavior attribute.

[0024] The behavior feature columns in the behavior label feature matrix may include at least the following dimensions: Labels for periodic electricity consumption patterns, such as “dominated by daytime consumption”, “dominated by nighttime consumption”, “evenly distributed throughout the day”, or “significantly increased load on weekends”. The corresponding values ​​can be discrete categorical codes, such as 0, 1, 2, 3; The peak-valley response sensitivity index indicates the user's response to electricity price fluctuations or dispatch instructions. It can be a normalized floating point number between 0 and 1. The higher the value, the stronger the response capability. The electricity price elasticity coefficient is a parameter that describes the load change trend of users under different electricity prices. For example, it can be obtained by fitting a regression model of historical electricity prices and electricity consumption, reflecting the impact of electricity prices on user behavior. The equipment load constitutes a vector, which reflects the usage frequency of various types of high-energy-consuming equipment and their proportion of the total load, such as air conditioners, water heaters, charging piles, etc. The proportion of each type of equipment constitutes a dimension in the vector; The load fluctuation stability coefficient reflects the volatility of user load, such as the standard deviation or coefficient of variation calculated based on the historical load series, which is used to describe the difficulty of user load forecasting; and the Boolean label of whether to participate in the demand response plan, which is used to distinguish whether the user has the ability to actively adjust the load, and can take the value of 1 (participation) or 0 (non-participation).

[0025] After completing the data acquisition and feature extraction of the above three dimensions, the system performs multi-dimensional fusion of the load sample set, meteorological feature set and user behavior label matrix to construct an input data structure in a unified format. This structure is usually represented in the form of a tensor or table, and each sample contains multiple time steps (such as the past 24 hours), multiple feature dimensions (such as power, temperature, behavior labels, etc.) and spatial indexes (such as node numbers, feeder IDs). In order to adapt to the continuous changes in data during actual operation, the data structure is dynamically updated according to the rolling forecast cycle, for example, every 5 minutes, 15 minutes or 30 minutes, so that the prediction model always receives the latest and complete input information during operation.

[0026] In summary, the dataset construction method not only fully considers the spatial structure, electrical characteristics and user differences of the distribution system, but also ensures the high quality and timeliness of the input data through data processing and modeling. It provides a solid foundation for building a high-precision, interpretable and robust short-term load forecasting model, and is a key link in realizing the linkage between intelligent forecasting and scheduling.

[0027] Step S102: Based on the comprehensive load forecast input data set, a short-term load forecast model is constructed to forecast the load levels of multiple feeders and key nodes in a preset future period, and obtain the predicted load time series at the node level and feeder level.

[0028] Step S102 involves constructing a short-term load forecasting model based on a comprehensive load forecasting input data set to achieve accurate prediction of future load levels of multiple feeders and key nodes in the distribution system, and output load time series forecasting results with node-level and feeder-level resolution.

[0029] Before executing this step, you must first confirm that the comprehensive load forecast input data set has been constructed. This data set contains multi-dimensional feature information from multiple sources, including historical active and reactive load data, real-time meteorological factors such as temperature, humidity, wind speed, sunshine, etc., and time series characteristics of user power consumption behavior, such as load change patterns, response capabilities, power consumption categories, etc. All data are processed through time alignment, spatial mapping, feature normalization, etc. to form a unified time series sample, which constitutes the input of the prediction model.

[0030] During the model building phase, the system needs to select a forecasting algorithm suitable for multivariate time series modeling based on the timeliness and resolution requirements of the load forecast. Preferably, the long short-term memory network (LSTM) or gated recurrent unit (GRU) in the deep learning model can be used. Such models can effectively capture the nonlinear dynamic change trend and mutation behavior of historical loads and are suitable for processing seasonal, sudden or periodic load data. If the model calculation complexity and interpretability are considered, traditional statistical regression methods such as ARIMA or support vector regression (SVR) can also be used as a baseline comparison solution.

[0031] In the specific implementation, the input of each model includes data windows of multiple time steps, and each time step contains information of different feature dimensions. The model predicts the load level of each feeder and key node in the next several time steps based on the continuous feature sequence in the historical window. To improve the prediction accuracy, the system can introduce an attention mechanism to dynamically adjust the model's attention weights on different features, or adopt an integrated prediction strategy that integrates the results of multiple models to generate the final prediction output through weighted average or voting mechanism.

[0032] The model training phase is usually performed based on historical data, using a sliding window strategy to build a training sample set, and using actual load data as a supervisory label to minimize the error between the forecast result and the true value, such as the mean square error (MSE) or mean absolute percentage error (MAPE). After the model is deployed, the system will continue to input the latest data in a rolling forecasting manner, update the model status in real time, and output new forecast results to ensure that the forecast results are timely and adaptable.

[0033] The forecast output needs to cover different spatial levels, including the total load curve forecast at the feeder level and the local load forecast at the key nodes. The forecast duration can be set to the next 15 minutes, 30 minutes, 1 hour or longer time period according to the actual application. The forecast result format is a time series data with a timestamp. The output structure includes load value, confidence interval, trend direction and mutation warning information.

[0034] It is worth emphasizing that the prediction model in this step is not statically constructed and used for a long time, but a model that is dynamically updated in accordance with changes in the distribution network operating environment. The system supports periodic fine-tuning or incremental learning of the model based on real-time feedback data, thereby ensuring the accuracy and robustness of the model in the long term.

[0035] In summary, step S102 realizes active perception of the future load state of the distribution system by constructing a high-precision, updateable short-term load forecasting model, providing high-resolution, high-confidence data support for subsequent risk identification and scheduling strategy formulation, and is the key foundation for the present invention to achieve intelligent and forward-looking regulation.

[0036] Furthermore, the short-term load forecasting model includes a topology encoding submodule, a feature fusion submodule, a time series prediction submodule and an output reconstruction submodule; The topology encoding submodule takes the grid topology structure of the target distribution area as input, uses graph neural network to construct node vector embedding representation with electrical attributes, and encodes the connection relationship, impedance parameters, equipment constraints and physical position relationship of each feeder and node in the topology into a low-dimensional continuous vector as an explicit expression of the impact of the topology structure on load forecasting. The feature fusion submodule takes the topological coding results, historical load time series, meteorological features after spatial interpolation and behavior label matrix as joint inputs, adopts a multi-layer perception fusion mechanism, extracts key features based on the attention mechanism and feature gating mechanism, eliminates scale differences and semantic conflicts between data, and outputs a multi-dimensional dynamic feature vector sequence of uniform length as a comprehensive driving factor for load changes; The time series prediction submodule takes the fused multi-dimensional dynamic feature vector sequence as input, and uses a bidirectional gated recurrent unit network enhanced by time series attention to predict the node-level and feeder-level loads step by time. By introducing an adaptive time window mechanism and an abnormal state dynamic correction structure, the model has the ability to identify short-term load mutation events and actively suppress prediction errors. The output reconstruction submodule maps the predicted value back to the physical number of each node or feeder according to the relationship between the timing prediction results and the actual configuration of the power grid nodes, and adds confidence indicators, upper and lower bounds of the prediction error, and typical power consumption scenario labels to provide multi-level prediction information support for subsequent scheduling strategy optimization and execution control.

[0037] The overall model includes a topological encoding submodule, a feature fusion submodule, a time series prediction submodule and an output reconstruction submodule. The modules are connected in sequence to form a prediction system with a clear structure and a closed logic loop.

[0038] The topology encoding submodule takes the grid topology of the target distribution area as input. The topology consists of transformers, feeders, tie switches, nodes, and connection relationships, with electrical parameters of each line segment, such as impedance, conductance, capacitance, voltage level, and boundary parameters of adjustable devices. In order to effectively extract the constraints and coupling effects of topology on load evolution, this module is modeled based on graph neural networks (GNNs). In implementation, the grid structure is first represented as a graph, where nodes represent physical entities (such as distribution transformers, load nodes, and switch locations), and edges represent electrical connection relationships, with electrical properties of each edge. Subsequently, a graph convolutional neural network (GCN) or a graph attention network (GAT) is used to learn vector embedding for each node, so that the output low-dimensional node representation vector can capture both its local electrical characteristics and the global topological context. These node vectors will be used as topology-aware representations to participate in the subsequent feature fusion process.

[0039] The feature fusion submodule takes the node vector output by the topology encoding submodule as the starting point, and combines the historical load time series data, the meteorological feature data after spatial interpolation processing, and the user behavior label matrix. The historical load data provides the power curve information of each node or feeder in the past period of time. The meteorological feature data includes temperature, humidity, wind speed, sunshine intensity, etc., and has been mapped to the specific feeder or node service area through geographic interpolation methods (such as Kriging interpolation or inverse distance weighted method). The behavior label matrix is ​​used to describe the typical power consumption characteristics of each user or load node, such as periodic load, response capability, and electricity price sensitivity.

[0040] In terms of fusion methods, the feature fusion module uses a multi-layer perception mechanism to construct a multi-source joint representation. The specific implementation is to embed and encode each type of feature input, unify its dimensions and then splice it, and automatically learn the weight distribution of various features for the prediction target through the attention mechanism. At the same time, in order to avoid scale inconsistency, statistical distribution differences and information redundancy between different data sources, the module also introduces feature gating mechanisms, such as gating units or channel-wise attention mechanisms, to achieve feature selection and suppression functions. Finally, the fusion module outputs a multi-dimensional dynamic feature vector sequence of uniform length and uniform dimension as the direct input for time series prediction.

[0041] The goal of the time series prediction submodule is to use the fused dynamic feature sequence to predict the load level of each node or feeder in the future time period step by time. This module adopts a bidirectional gated recurrent unit network (Bi-GRU) structure, and combines the time series attention mechanism to enhance the model's ability to focus on key time segments. The model not only realizes the fusion of previous and next information in time, but also dynamically determines the historical depth required for prediction by introducing an adaptive time window mechanism. For example, when the load volatility is large or the weather changes frequently, the system automatically increases the input window length to improve the model's context grasping ability; while in the stable operation stage, the window length is shortened to improve computational efficiency. In addition, the module also integrates an abnormal state dynamic correction structure, which can adjust the feature disturbance or confidence weight of the sudden abnormalities detected in the input sequence (such as outliers caused by equipment jumps or extreme climate), thereby avoiding overfitting of the model to abnormal data and improving the overall prediction robustness. The final output is a future load time series prediction sequence at the node level and feeder level, covering multiple future time steps.

[0042] The output reconstruction submodule restores the above prediction results to structured outputs that can be recognized by the dispatcher. This module decodes the vector or sequence output by the model according to the physical number and mapping relationship of each prediction unit in the original topology of the power grid, and restores it to the physical level feeder ID, node number and load prediction value at each prediction time. In order to enhance the interpretability and practicality of the prediction results, this module also calculates the confidence level of each prediction value (such as based on Bayesian estimation or Monte Carlo Dropout), the upper and lower bounds of the error (such as 95% confidence interval), and automatically adds typical power consumption scene labels, such as "high-temperature load", "holiday cycle load", "weather mutation impact load" and other labels to indicate which typical mode the current prediction belongs to. This additional information helps the dispatch control module to make early warning judgments, priority settings and strategy redundancy design when generating control strategies in the future.

[0043] Through the collaborative work of the above submodules, the short-term load forecasting model realizes the whole process modeling link from structural modeling, multi-source fusion, dynamic prediction to executable output, fully combining the physical structure characteristics, load behavior laws and operating environment changes of the distribution network, and providing an accurate, explainable and controllable prediction basis for subsequent adaptive scheduling. The structure has good scalability and model migration capabilities, and can adapt to the needs of intelligent distribution scheduling in different regions and different network structures.

[0044] The abnormal state dynamic correction structure in the present invention belongs to the compensation mechanism in the time series prediction submodule, which aims to identify and correct the prediction deviation caused by sudden meteorological changes, drastic fluctuations in user behavior or other external disturbances. The core of this structure is to make adjustable and interpretable secondary corrections based on the initial prediction results according to the current environmental characteristics, behavior change characteristics and node historical response characteristics, so as to enhance the robustness and accuracy of the model under complex operating conditions.

[0045] Furthermore, the abnormal state dynamic correction structure in the time series prediction submodule performs node-level dynamic adjustment on the initial prediction value through the following formula 1: ; in, Indicates the corrected Nodes at time The predicted load value is the initial predicted value The prediction result is obtained by combining multiple correction factors with weighted compensation. This prediction value will be used for subsequent grid dispatch judgment.

[0046] For the Nodes at time The initial prediction value of is output by the main model (such as the bidirectional GRU structure) without considering abnormal disturbance factors. This value provides the main trend of the prediction, but may deviate when the load fluctuates violently.

[0047] For Node The individual correction coefficient of is obtained by joint training based on the load fluctuation degree of the node in the historical prediction task and the model prediction confidence, and is defined as: ; in For Node The historical load The variance of For historical load The mean of For Node The average forecast confidence of is used to quantify the credibility of the forecast results given by the main model (i.e., short-term load forecasting model) for this node in the historical forecasting process. The purpose is to introduce uncertainty factors as part of the personalized correction coefficient. In other words, The smaller it is, the greater the uncertainty of the model's prediction of the node, and stronger post-processing correction is required; When it approaches 1, it means that the model is highly stable and reliable in predicting the node, and the correction coefficient should be appropriately reduced.

[0048] In implementation, the confidence can be obtained through a Bayesian neural network structure or Monte Carlo sampling dropout method (MC Dropout) to obtain a prediction distribution rather than a single prediction point value. The degree of discreteness of the prediction distribution can be used to express uncertainty, which is often measured by the "bandwidth of the prediction interval".

[0049] Assuming that the Bayesian method or MC Dropout mechanism is used, for nodes At the point in time conduct prediction sampling, we get: ; in, Indicates Nodes at time The prediction result of It is generated by using a model with Dropout or a Bayesian model with parameter sampling under the same input data.

[0050] The confidence band width for the set of predictions can be calculated, for example: ; The meaning of this formula is to express the uncertainty range of the prediction result at a certain node or time point. It is obtained by statistically analyzing multiple prediction values ​​to obtain a confidence interval width that covers most possible situations.

[0051] Specifically, for the node In time The prediction does not take a single value of the model output, but obtains a set of prediction values ​​through multiple sampling or model uncertainty modeling (such as MC Dropout, Bayesian neural network, etc.), which reflects the output distribution of the model under the current input. After sorting this set of prediction values, take the difference between the 97.5th percentile and the 2.5th percentile. This difference represents the fluctuation range of the prediction value at the 95% confidence level, that is, the width of the prediction interval.

[0052] Therefore, this width reflects the model's At the moment The stability of the prediction results. The narrower the interval, the more concentrated and stable the model prediction is, and the more confident it is; the wider the interval, the more divergent the prediction distribution is, the higher the uncertainty is, and the lower the model's confidence in the prediction point is. This width will eventually be used to construct the prediction confidence index of the node, further affecting the response strength of the anomaly correction mechanism.

[0053] In order to unify the dimensions and range of each node, it is also necessary to normalize them. For example, the normalized width is defined as: ; In this way, the nodes with more unstable predictions and wider intervals are The smaller it is; while the prediction is concentrated, the nodes with small fluctuations, Approaching 1.

[0054] Final That is, the average confidence of the node in several historical periods (such as the past week or an operating cycle): ; in Indicates the number of sampling time points. It is recommended to take at least (i.e. 15 minutes of data per day) to ensure statistical stability.

[0055] It represents the overall load variability index at the current moment, which is dynamically calculated based on the standard deviation of the network-wide forecast load sequence; it is defined as: ; in is the total number of nodes participating in the prediction in the system; It is Nodes at time The initial predicted load value is the original prediction result output by the short-term load forecasting model (such as the bidirectional gated recurrent neural network Bi-GRU) before the abnormal correction mechanism is applied.

[0056] It's time The average initial prediction value of all nodes. This parameter is used to amplify the effect of the abnormal correction term when the overall system fluctuates violently.

[0057] Is a node The abnormal response sensitivity factor is used to characterize the load response intensity of the node under external disturbance conditions such as meteorology. It is obtained by the following regression statistics: ; in It is a node during the period of abnormal weather The mean value of load variation; It is the variation range of meteorological indicators (such as temperature, humidity, etc.) during the corresponding period.

[0058] For Node At the moment The behavior disturbance index is derived from the "volatility" feature dimension in the behavior label matrix, such as whether the user enters the demand response state at this moment, enables high-load equipment, participates in energy-saving strategies, etc. It can be extracted by the following formula: ; in Representation Node In the behavioral characteristics matrix perturbation dimension; is the disturbance importance coefficient of the corresponding dimension. Represents the total number of perturbation dimensions, usually taken as ; It can be automatically learned through training or assigned by experience (such as 0.5 for load switch behavior and 0.3 for response state, etc.). Common Control Index Function Section , realizing a nonlinear response function. When the disturbance intensity is small, the exponential term approaches zero and no correction is made; when the disturbance is strong and the sensitivity is high, the term approaches 1, resulting in a significant correction.

[0059] For Node The topological embedding vector is generated by the topological encoding submodule (such as the neural network in the figure) according to the electrical connection structure, physical properties and positional relationship of the node. It is usually a dense vector with a dimension between 16 and 64, reflecting the structural role of the node in the distribution network.

[0060] For the moment The dynamic weight vector of the behavior feature is calculated by the feature fusion submodule. This vector represents the influence of the current behavior feature in the global prediction through the attention mechanism, and its dimension is Matching is used to do the inner product with the topological vector to form the final disturbance response weight term.

[0061] Inner Product Structural regulation is achieved, that is, the correction term will be amplified only when the topological structure of the node is significantly correlated with the current behavioral characteristics, thereby achieving dynamic regulation that combines locality and contextual relevance.

[0062] Overall, the formula forms an explainable, learnable, and personalized load forecast correction mechanism, which integrates system-level fluctuations, node-level historical behaviors, topological locations, and disturbance events, thereby providing highly reliable forecast support for the intelligent distribution scheduling in the present invention, which is particularly suitable for distribution scenarios with unstable loads, significant behavior-driven, and tight topological coupling.

[0063] In the short-term load forecasting model of the present invention, the feature fusion submodule is used to integrate information from different sources into a set of structured dynamic input features to drive the subsequent time series forecasting process. Considering that different nodes have significant differences in behavioral drive and meteorological sensitivity, the system does not adopt a unified linear splicing method, but designs a feature fusion method based on the behavioral meteorological joint decoupling attention mechanism. The core is to weight the mapping results of the behavioral features and meteorological features in different semantic spaces in proportion to achieve personalized fusion expression of each node.

[0064] Furthermore, the feature fusion submodule includes a feature construction structure based on the behavioral meteorology joint decoupled attention mechanism, and generates the fusion input representation of each node through the following formula 2: ; in, Representation Node At the moment The fused input vector is used as the input of the time series prediction submodule (such as the Bi-GRU network) to predict the load evolution trend at several moments in the future. Its essence is a structured expression of multi-source heterogeneous features, taking into account the different impact mechanisms of behavioral characteristics and meteorological characteristics on load.

[0065] Represents the nodes extracted from the behavior label feature matrix At the moment The behavior driving vector reflects the characteristic state of the node based on the user's electricity consumption behavior at the current moment. It can include information such as the user's periodic load type (such as night-dominant type), whether it is in the demand response cycle, electricity price sensitivity index, controllable load proportion, user fluctuation frequency, etc. It is usually expressed in the form of a numerical vector with a dimension range of 5 to 20, depending on the feature selection strategy.

[0066] Indicates the meteorological characteristics after spatial interpolation at the node The local environmental factor vector at the node is the meteorological observation data (such as temperature, humidity, wind speed, sunshine intensity, weather category, etc.) mapped to the node through spatial interpolation method The local meteorological vector is obtained after the geographical location of the node. Meteorological interpolation can use Kriging, inverse distance weighted (IDW) or geographic mapping algorithm based on Voronoi polygons. This vector expresses the physical impact of the current environment of the node on the load change, and the dimension is generally between 5 and 10.

[0067] are trainable weight matrices used for decoupling transformation of behavior and meteorological features, with the goal of mapping the original behavior or meteorological vector into a unified target feature space. The matrix sizes are and ,in and are the dimensions of original behavioral characteristics and meteorological characteristics, is the dimension of the target fusion vector, the recommended value is or 64. The behavior vector is obtained by the hyperbolic tangent function Nonlinear compression is performed to retain detail changes and control their numerical range; the meteorological vector maintains positive sensitivity through the modified linear rectification function ReLU to reflect the incentive effect of weather changes on load.

[0068] For Node The individual behavior weight coefficient is calculated using the following formula 3: ; in, Representation Node The historical behavior dominance index is obtained by statistical learning based on the explanatory strength of user behavior on load changes in the past scheduling cycle. The specific calculation method is to use linear regression or feature importance scoring (such as SHAP value) to evaluate the contribution of behavioral input features to prediction accuracy in the historical prediction process.

[0069] Representation Node The meteorological sensitivity parameter is estimated based on the partial derivative of the node load change during historical meteorological drastic events. This value can be estimated by analyzing the load change curve during historical meteorological drastic events (such as high temperature, typhoon, rainstorm), and taking the derivative of the load change and the meteorological variable change or fitting the linear slope, so as to estimate the response sensitivity of the node.

[0070] In the formula It realizes personalized and continuous control of behavior and meteorological weights, rather than fusion through static weighting or fixed strategies. This enables the system to dynamically determine the dominant weight of behavior and environmental factors in feature fusion based on the physical properties of different nodes and user types, thereby improving the accuracy and generalization of predictions.

[0071] Finally, the fused input vector constructed by the above mechanism is It has the advantages of strong interpretability and flexible fusion strategy, providing a high-quality input basis for time series prediction models and has wide adaptability in multi-source heterogeneous input scenarios.

[0072] Step S103: Based on the predicted load time sequence, determine the voltage deviation, load imbalance or equipment overload risk that may occur at each node and feeder within the predicted time period.

[0073] The core of step S103 is: after obtaining the load forecast results of each feeder and key node in the future period, further evaluate the operation risk of the distribution system in the forecast period, including various potential abnormal conditions such as voltage deviation, load imbalance and equipment overload, so as to provide a quantitative basis and target constraints for subsequent scheduling optimization.

[0074] This step first simulates and deduces the voltage and current change trends in the future time period based on the node-level and feeder-level predicted load time series data output in step S102, combined with the static structural parameters and dynamic state quantities of the power grid operation. The judgment of voltage deviation is based on the deviation between the node voltage value calculated after the node predicted load is injected and the rated voltage. At the implementation level, the electrical distribution of the future time period can be simulated by combining the current system topology, power supply side output, voltage regulation equipment settings and other factors through power flow calculation algorithms, such as Newton-Raphson method or fast decoupling method, and the voltage of each node is normalized and compared with the upper and lower allowable values. When the voltage prediction value of a node continues to deviate from the normal range, and the deviation exceeds the set threshold (for example, ±5% or ±7%), it can be determined that the node has a potential voltage deviation risk.

[0075] In the process of determining the unbalanced load of a feeder or node, the system needs to make phase-to-phase comparisons based on the predicted injection amount of the three-phase load. If the system adopts a three-phase asymmetric modeling method, each feeder section or node will have a three-phase power prediction value. The system can calculate the difference or imbalance rate between the power of each phase (such as the load deviation rate or the zero-sequence current estimate) and compare it with the safe operation threshold to identify areas where severe phase imbalance may occur during the prediction period, prompting the subsequent dispatching system to perform phase reconstruction or phase shift adjustment on the load.

[0076] The judgment of equipment overload risk is usually based on the relationship between the predicted load and the rated capacity of the equipment. For key equipment such as transformers, feeders, switches, capacitor banks, etc., the system records their rated capacity and overload tolerance, and summarizes them in combination with the predicted value of the connected load. When a device continuously carries a power exceeding its rated capacity or exceeds the thermal stability limit (which can be dynamically set according to the thermal characteristic curve of the equipment) during the predicted period, it can be considered that the equipment has an overload risk. This type of judgment can be assisted by tools such as equipment thermal models and load-temperature rise relationships to improve identification accuracy.

[0077] In actual applications, this judgment process is usually performed in a rolling manner, that is, after the prediction results are generated, the risk scan is performed for each future period at a frequency of minutes or hours to ensure that any potential anomalies can be identified in advance. In addition, the system can also introduce a risk level assessment mechanism to classify risks according to factors such as deviation amplitude, duration, and equipment level, thereby providing a priority ranking basis for subsequent scheduling optimization strategies.

[0078] This step is not limited to detecting the risk itself. What is more important is to output the risk identification results in a structured manner, including the potential risk type, risk location, occurrence period, severity, and impact range, so that the scheduling module can clearly identify the intervention object and target period and achieve targeted optimization.

[0079] In summary, step S103 realizes a closed-loop connection from data-driven prediction to physical constraint identification through quantitative analysis and risk identification of the predicted load results, which significantly enhances the foresight, accuracy and reliability of the dispatching control strategy in the present invention.

[0080] Furthermore, judging the voltage deviation, load imbalance or equipment overload risk that may occur at each node and feeder within the predicted period according to the predicted load time sequence includes: The predicted load time series is mapped to the distribution network topology, and the predicted voltage response trend is derived by combining the node voltage regulation capability, electrical connection relationship and load conduction path. Based on the prediction results of the three-phase load distribution, the feeder phase imbalance rate is calculated, and the continuous imbalance evolution trend is identified by combining the historical asymmetric operation records. According to the rated capacity of the equipment and the load evolution rate, the load rate of each node and feeder within the prediction interval is analyzed in time series to determine whether it crosses the equipment operation safety threshold; Cross-compare the judgment results with the load surge characteristics in the behavior labels, screen out highly sensitive nodes that are significantly affected by behavior and have overload risks, and build a node-level risk list.

[0081] This embodiment constructs a dynamic, data-driven grid operation risk identification mechanism based on the prediction results of node-level and feeder-level loads in future time periods. First, the system accurately maps the predicted load time series with the topological structure of the power grid. The load change trend of each node is embedded in its specific position in the topological network, and combined with the voltage regulation capability of the branch where the node is located, the electrical connection method with the upper node, and the active and reactive power conduction paths from the main substation to the node. And other key factors, a sensitive model of node voltage response to load is established. The system analyzes the change amplitude and rate in the load time series, combined with the power grid flow distribution rules, and derives the voltage response trend of the node in the predicted period, thereby identifying possible voltage deviation risk areas.

[0082] When judging the risk of load imbalance, the system will perform phase deconstruction on the predicted load according to the three-phase structure of each feeder to form time-series A, B, and C three-phase load distribution data. The system calculates the imbalance degree of phase current or power, and establishes a trend tracking model in a continuous time window based on the phase deviation rate and imbalance rate indicators. If a feeder is identified to have a systematic imbalance growth trend in multiple consecutive prediction time steps, and it cannot be corrected by normal load transfer methods, then the feeder is judged to have a potential load imbalance risk.

[0083] In order to analyze the risk of equipment overload, the system introduces the operating characteristics of the equipment itself based on load forecasting, especially parameters such as rated capacity, temperature rise limit and short-term overload tolerance. Combined with the time-series evolution trend of the predicted load, the system calculates the load rate curve of each device (including distribution transformers, feeders, tie switches, etc.) in real time, and determines whether it will exceed the safe operating range in the future. In particular, for equipment with a fast load growth rate or close to the capacity boundary, the system will mark it as a key monitoring target.

[0084] In order to further enhance the interpretability and accuracy of risk identification, the system also associates and analyzes the above-mentioned physical indicators with user-side behavior tags. The behavior tag matrix identifies the user types with high probability of load surge in the current forecast period (such as centralized charging at night, centralized startup of high-power equipment, participation in demand response to release load, etc.), and analyzes whether these behaviors are concentrated on certain nodes or branches, thereby forming behavior-driven risk indicators. If a node has both electrical operation risks (such as voltage overlimit or overload tendency) and behavioral driving signs of sudden load increase, it will be identified as a highly sensitive node and included in the node-level risk list for priority intervention in subsequent scheduling strategies.

[0085] In summary, this step builds a multi-category risk judgment mechanism that can be updated in real time by integrating the prediction data with the electrical structure, equipment capabilities, and user behavior in multiple dimensions. It does not rely on cumbersome static simulation processes and has high adaptability and prediction accuracy.

[0086] Step S104: Based on the judgment results of the grid topology, electrical parameter constraints and load forecast timing, a rolling time domain optimization method is used to generate a dispatching control strategy, which includes advance adjustment of transformer tap positions, advance switching plans of capacitor banks, reclosing sequence of feeder tie switches and load distribution logic, charging and discharging time and power settings of distributed energy storage devices, or demand response execution instructions for the user side.

[0087] This step is one of the core of the present invention. Its function is to generate a dynamically executable dispatching control strategy based on the physical structure and operating constraints of the current distribution network by using a rolling time domain optimization strategy after obtaining the load forecast results and potential risk assessment results, so that the system can actively adjust the operating state before the load mutation or operating pressure occurs, and realize the adaptive coordination of the load of the whole network.

[0088] In this step, three key information sources need to be input first, namely the topological structure of the distribution network, the electrical parameter constraints, and the load forecast results and risk determination information output by step S103. The topological structure of the distribution network includes the location, connection relationship and physical parameters of substations, transformers, feeders, nodes, tie switches and the terminal loads or distributed power sources connected thereto, such as impedance, conductivity, etc. This information is usually stored in the form of a topological diagram or matrix and loaded into the optimization engine before the scheduling calculation. Electrical parameter constraints include actual operating restrictions such as the rated capacity, voltage level, adjustment range, switching time delay, energy storage power boundary, charging and discharging rate, and operating cost of each device. These constraints are crucial to ensure the feasibility of the scheduling strategy at the physical level.

[0089] The rolling time domain optimization method is the main technical means used by the present invention to generate dispatch strategies. The core idea of ​​this method is to divide the load forecast cycle into multiple continuous forecast windows, perform an optimization calculation in each window, and recalculate the dispatch strategy in the next cycle based on the latest load forecast and actual feedback, so as to achieve closed-loop control with sequential advancement and real-time correction. The optimization model usually adopts a mixed integer linear programming (MILP), dynamic programming (DP) or an algorithm framework based on predictive control (MPC). The objective function can be set according to actual needs to minimize system energy consumption, reduce regulation costs, smooth load fluctuations or improve voltage qualification rate. Constraints include power flow balance equations, equipment state boundaries, action intervals, and user response capabilities.

[0090] The generated dispatch control strategy should be refined into executable specific control instructions and be able to achieve multi-level coordination in time and space dimensions. The strategy may include the following types of operations: First, the transformer tap position is adjusted in advance. The dispatching system calculates and sets the gear switching plan of the transformer voltage regulator in advance based on the load forecast and voltage deviation trend, so that the gear conversion can be completed before the load increases or decreases, thereby stabilizing the downstream voltage level. The gear adjustment needs to consider the minimum time interval and action cost of gear change to avoid frequent switching and impact on the equipment.

[0091] Secondly, it is the pre-switching of capacitor banks. Capacitor banks are used to adjust reactive power and improve voltage quality. Based on the reactive load forecast results, the system predetermines the combination of capacitor banks to be put into or out before the low voltage risk occurs. The action instructions include the specific group number, action time and control node information. In order to achieve regional coordination, group control and bus voltage closed-loop correction mechanisms can also be used.

[0092] The next step is the reconstruction strategy of the feeder tie switch. The system optimizes the load distribution relationship between feeders and dynamically generates the closing and opening sequence of the tie switch to realize the active transfer of load from the overloaded feeder to the spare capacity feeder, thereby alleviating the load concentration phenomenon. The tie switch operation needs to consider the flow direction, line protection coordination and fault isolation strategy at the same time to ensure that the system stability is not destroyed after reconstruction.

[0093] In addition, it also includes the charging and discharging scheduling of distributed energy storage devices. According to the load forecast curve of the entire network, the remaining capacity of energy storage and its health status, the system arranges to release energy before the load peak and absorbs excess energy when the load is low, thereby achieving peak shaving and valley filling. The scheduling strategy needs to specify the start and end time of charging and discharging, power size, participating unit number, and consider technical details such as battery efficiency, maximum number of cycles and converter limitations.

[0094] Finally, demand response instructions are issued on the user side. Based on the overload or voltage risk areas that may appear in the forecast, the system selects users with response capabilities and temporarily adjusts their load levels through price signals, incentive compensation or direct control. For example, some users can be instructed to reduce air conditioning loads, delay the start of high-energy-consuming equipment, or dispatch industrial loads to standby operation mode, thereby reducing system peak pressure.

[0095] It should be noted that all scheduling instructions must be compatible with the real-time execution system and can be distributed to field devices through the automation platform. After the strategy is generated, the system retains the execution time point, object identification, operation parameters and expected impact of each control operation as the basis for subsequent strategy execution confirmation and effect evaluation.

[0096] Through this step, the distribution system can intervene in potential operational risks in advance and actively optimize various control resources, thereby significantly improving the overall operational intelligence level and adaptability.

[0097] Furthermore, the above-mentioned determination result based on the grid topology, electrical parameter constraints and load forecast timing sequence adopts a rolling time domain optimization method to generate a dispatch control strategy, including: The grid topology of the target distribution area is mapped into a graph data structure using a graph structure modeling method, where nodes represent feeder connection points, transformer locations or load access points, and edges represent line connection relationships. The line impedance, voltage level, and switch state electrical properties are considered to construct a graph topology representation that integrates topology and parameters. The node-level load forecast time series obtained according to the load forecasting model is embedded into the above graph structure as a dynamic additional attribute of the time evolution feature. The load response feature distribution under the dependency of the whole network structure is extracted through the graph neural network to capture the coupling effect between system levels. Based on the extracted state characteristics of the nodes in the entire network, a rolling time domain optimization framework is introduced to construct a multi-objective nonlinear optimization problem in each rolling cycle. Minimization of voltage deviation, suppression of feeder imbalance, reduction of equipment overload risk and minimization of load adjustment cost are used as joint objective functions, and the operation boundaries, action frequency constraints and chain action logic of various electrical equipment are integrated as constraint conditions. Through evolutionary strategy search, a joint operation sequence covering different control resources is dynamically generated in each rolling cycle, including transformer tap adjustment, capacitor bank switching, tie switch opening and closing, energy storage unit power allocation and user load priority sorting. Each strategy is attached with its execution window, predicted risk area and priority level information to ensure that the dispatch strategy has regional awareness, adaptability and high timeliness.

[0098] In this embodiment, the core of the rolling time domain optimization method is to establish a dynamic scheduling strategy formulation mechanism that can be adjusted in real time as the forecast is updated. First, in order to make the power grid structure computable and learnable in the data space, the system maps the topological structure of the target distribution area into a graph data structure. In this graph, all nodes represent feeder connection points, distribution transformer locations, or user load access points in the actual power grid, while the edges in the graph represent physical line connections and attach corresponding electrical properties, such as line impedance, voltage level, maximum transmission capacity, and switch closing and opening states. This graph structure not only retains the spatial relationship of electrical connections, but also incorporates the operating boundaries and control properties of the equipment, thereby providing a highly structured power grid foundation for subsequent scheduling optimization.

[0099] After constructing the topology graph, the system embeds the node-level load forecast time series output by the load forecasting model into the graph structure. The predicted load curve of each node over time will be used as a time evolution feature, and together with the static structural properties of the nodes in the graph, it constitutes the dynamic graph node feature. The system uses a graph neural network (GNN) to extract the state representation of each node in the graph under the topological constraints of the power grid. This representation not only reflects the load changes of the node itself, but also captures the mutual influence between it and the neighboring nodes due to topological coupling, such as the pressure caused by the load change of a line on the upper transformer or adjacent branches.

[0100] Based on the graph structure, the system introduces a rolling time domain optimization framework. This framework takes each prediction period as a rolling window and constructs an optimization model for the joint control problem within the current window. The optimization model adopts a nonlinear multi-objective function form, which comprehensively considers multiple goals such as voltage deviation minimization, three-phase load imbalance suppression, equipment overload risk reduction, and minimization of control operation cost. Among them, voltage deviation minimization refers to ensuring that the voltage of all nodes remains within the allowable fluctuation range, the three-phase imbalance suppression goal aims to reduce the deviation of the load distribution between phases, overload risk minimization is used to avoid key equipment from being in an overloaded state for a long time, and operation cost minimization ensures that the control measures will not be frequently activated to reduce the equipment life or cause secondary disturbances.

[0101] This optimization problem also incorporates a series of constraints related to the operation of distribution equipment. For example, for transformer taps, it is necessary to consider the allowable switching frequency and upper and lower ranges; the switching of capacitor banks is limited by the number of switching times and the voltage response rate; the operation of feeder tie switches must comply with network connectivity constraints and cannot cause island effects or form ring networks; the regulation of energy storage equipment is subject to physical limitations such as power, current rate, and charge and discharge cycles; and the response on the user load side needs to consider its adjustability label, response delay, and user priority.

[0102] In the process of generating the dispatching control strategy of the present invention, in order to achieve high-precision and high-reliability optimization decisions, the system designs and solves a nonlinear multi-objective optimization problem. This optimization model not only focuses on the safety of the operating state and the quality of power, but also emphasizes the economy and feasibility of the control behavior. Its objective function covers multiple performance indicators with practical significance and is coordinated and solved in the dynamic environment of the distribution network.

[0103] First, the voltage deviation minimization objective is used to ensure that the voltage level of each node in the distribution system is maintained within the allowable fluctuation range of the rated voltage as much as possible, usually ±5% or ±10%, depending on the national power standards. In the model, by introducing the deviation term between the predicted value and the target value of the node voltage, the voltage deviation of all key nodes is accumulated to form a penalty amount. This indicator has a direct impact on the power quality and equipment stability on the user side, so it is given a higher weight in the optimization.

[0104] Secondly, the three-phase load imbalance suppression target is used to reduce the problem of uneven power or current between phases at the feeder level. The system calculates the phase difference of the three-phase load forecast value in each feeder and introduces a standardized imbalance index to reflect the degree of load deviation between phase A, phase B, and phase C. The introduction of this target helps improve the system operation efficiency and avoid problems such as increased zero-sequence current, transformer overload, and increased energy loss caused by phase imbalance.

[0105] The third optimization goal is to reduce the risk of equipment overload, which aims to avoid the continuous overload state that may occur in the future period of important equipment in the distribution system in advance. By comparing the predicted load curve with the equipment rated capacity or allowable load curve (such as transformer overload curve, line temperature rise limit, etc.), the model evaluates the overload risk level of each key node or line, and introduces the load excess degree of high-risk nodes into the objective function in a weighted form. This design can effectively prevent the equipment from gradually entering the dangerous working area without control intervention, and improve the safety margin of the overall power supply system.

[0106] In addition, in order to avoid the dispatch strategy being effective in theory but executed too frequently in practice, resulting in mechanical fatigue of equipment or accumulation of system disturbances, the model introduces the minimization of the control operation cost as the fourth goal. In this part, each control action, whether it is transformer tap adjustment, capacitor bank switching, tie switch opening and closing, or energy storage unit scheduling, user response call, is assigned an operation cost weight based on parameters such as action frequency, response delay, and recovery cost. The optimization process tends to select a strategy with a smaller control cost and a better action path while meeting the requirements of power quality and safety, thereby ensuring the economy and stability of the control behavior.

[0107] In order to ensure the feasibility of the above multi-objective optimization under actual power grid conditions, multiple types of constraints related to the operating boundaries of electrical equipment are also introduced into the model. For example, the adjustment of the transformer tap needs to take into account its mechanical structure limitations, frequent switching is not allowed, and each adjustment must be performed within a limited gear range; the capacitor bank is affected by the action response time and the number of switching times, and its minimum action interval and maximum continuous switching times need to be controlled; the operation of the interconnecting switch must ensure that the system topology remains connected after the operation to prevent the emergence of power supply islands or the formation of unacceptable closed-loop structures; the energy storage equipment must simultaneously meet the physical boundary conditions of its power state, current change rate, and periodic charge and discharge times during scheduling; the participation ability of the user load side is also limited by behavioral labels. For example, a certain type of user only supports low-level demand response, has obvious startup delays, or has constraints that cannot be adjusted for certain loads. All this information needs to be dynamically incorporated into the optimization model.

[0108] Finally, the optimization model constructs a nonlinear joint solution problem based on the above objectives and constraints in each rolling time domain, uses a multi-objective optimization algorithm or evolutionary search mechanism for iterative solution, and outputs a set of joint dispatch strategies that meet the system operation objectives and physical boundary conditions. This strategy not only covers various control resources, but also has the attributes of execution priority and effective period, which facilitates the downstream control system to execute efficiently by region and level, thereby realizing truly intelligent and adaptive distribution network dispatch control.

[0109] In the optimization process, the system uses evolutionary strategy search algorithms, such as reinforcement learning, genetic algorithms or simulated annealing, to improve the efficiency of solving nonlinear multi-objective optimization and the ability to explore the solution space. Each rolling cycle optimization will output a set of joint operation sequences, which will cover multiple categories of control objects, such as which tapping level the transformer should be adjusted to within a certain period of time, whether a certain capacitor bank should be switched in advance, which interconnecting switches should be closed to reconstruct the load path, when and how much power a certain energy storage system should be charged or discharged, or what level of demand response instructions should be issued to which users.

[0110] In addition, in order to ensure the effectiveness and local adaptability of the strategy during implementation, the system adds additional information such as execution window, risk-related area and control priority to each operation strategy. The execution window defines the start and end time period when the strategy takes effect, the risk area indicates the load risk point or line segment that the operation is mainly used to alleviate, and the control priority is used to make sorting decisions when there is a resource conflict or multiple strategies are concurrent. Through these auxiliary attributes, the system can quickly respond to sudden changes in actual execution, while avoiding the diffusion of uncertainty caused by the rigid execution of the global strategy.

[0111] Furthermore, the graph structure modeling method includes: The initial topology diagram is generated based on the distribution network GIS data and the equipment master data in the SCADA system. The node type is determined by the equipment attribute field analysis, and the electrical parameters such as line impedance and voltage level are supplemented by real-time monitoring data and standard parameter library. The topology connectivity verification algorithm is used to identify the breakpoints, island nodes or loop structures in the graph, and the graph structure is closed in combination with the electrical operation rules to ensure the integrity and decoupling of the topology graph. Structured parameter vectors are embedded in the nodes and edges of the topology graph respectively. The node parameters include access capacity, node type and control capability identification, and the edge parameters include impedance value, maximum current capacity and current switch state, so as to realize the fusion expression of topological structure and operating characteristics.

[0112] The present invention adopts a modeling method based on graph neural network to represent the power grid structure in the process of scheduling optimization. Therefore, the accurate construction and parameter characterization of the graph structure are the basis of the entire intelligent distribution load prediction and scheduling method. In order to realize graph structure modeling, the system first needs to structure and abstract the information of the real distribution network. This step is completed by obtaining the GIS (geographic information system) data of the distribution network and the equipment master data recorded in the SCADA system. Among them, the GIS data provides the spatial distribution and connection relationship of facilities such as distribution lines, transformers, feeder switches, and user access points, and the SCADA system provides the operating attributes and control fields of each electrical equipment, such as equipment type, voltage level, installation location and operating status. By parsing the fields of these information, the system can automatically identify the physical nodes in the distribution network and determine the type of each node, such as load nodes, switch nodes, transformer nodes, or feeder contact points, so as to construct the initial network node set.

[0113] After the node is built, the system generates the corresponding edge structure based on the physical connection relationship between the devices. The edge represents the line connection between the nodes. These edges not only indicate the existence of the connection, but also carry key electrical parameter information, including the impedance value, voltage level, maximum transmission current, etc. of the line. Some parameters can be obtained from SCADA real-time monitoring data, such as current, voltage and power factor; other static parameters come from power design standards or power grid archives to complete the missing physical properties in the graph structure.

[0114] The constructed initial topology may have structural errors or be incomplete, so the system needs to perform topological connectivity verification on the graph structure. The connectivity check algorithm identifies breakpoints, island nodes (i.e., isolated devices not connected to the main network), or loop structures in the network, and closes these problems. The closing process refers to the grid operation procedures and safety constraints, appropriately adds virtual edges, completes disconnection information, or disconnects illegal loops according to procedures, thereby ensuring that the entire topology is closed and connected, and has the structural legitimacy required for network flow calculation and optimized scheduling.

[0115] After the topology is constructed, the operating parameters related to each node and edge need to be further embedded in the graph to form the attribute representation of the graph. For nodes, the embedded structured parameter vector includes the maximum access capacity of the node (such as the allowable load limit or energy storage capacity), the functional type label of the node (such as load node, transformer node), and whether control is supported (such as tap, circuit breaker, adjustable capacitor, etc.). For edges, it is necessary to embed the impedance value, maximum allowable current capacity, and the current switch state of the line (closed, open or tripped). These parameters are converted into standardized vector form for subsequent graph neural network input, so that it can not only learn the network topology, but also consider the actual operating physical characteristics and regulation capacity boundaries at the same time.

[0116] Through the above modeling process, the graph structure constructed by the system is no longer a single static connection graph, but a comprehensive expression that integrates structure, parameters and control capabilities. The graph structure constructed in this way can not only accurately reflect the physical topology of the power grid, but also support the deep learning algorithm to extract local or global load response relationships, providing a solid foundation for the generation of rolling optimization scheduling strategies and ensuring that the control measures are physically and logically executable. The graph structure can be continuously updated and supplemented in different rolling forecast cycles, thereby achieving the timeliness, accuracy and adaptability of the distribution network operation model.

[0117] Furthermore, the process of embedding the node-level load forecast time series into the graph structure includes: The load values ​​of each node at multiple prediction moments are constructed as dynamic feature vectors, and load confidence, prediction upper and lower bounds, and abnormal fluctuation signs are added to form a time series feature set with uncertainty. The time series feature set is mapped one-to-one with the topology graph nodes through the node and time dual index mechanism, and the prediction sequence is segmented in a rolling manner using the time window mechanism to ensure that the graph neural network input has time consistency within each rolling cycle. A multi-layer graph convolution structure is used to fuse static topological attributes and dynamic load characteristics, so that the node state representation not only reflects the evolution trend of its local load, but also reflects the coupling influence with neighboring nodes in structure and timing, providing context-related input for rolling optimization.

[0118] In the dispatch optimization scheme of the present invention, the graph structure not only carries the topological information of the distribution network, but also needs to dynamically integrate the time series prediction results from the short-term load forecasting model. To achieve this goal, the system first structures the predicted load of each node in the future. Specifically, the load forecasting model outputs the load forecast value of each node at multiple future moments, such as 96 15-minute values ​​or hourly forecasts for several hours in the next hour. These predicted values ​​are organized into a feature vector in the form of a time series to express the evolution trend of the future load of the node.

[0119] In order to enhance the identification and modeling of uncertain factors, the system will also attach prediction confidence, upper and lower boundaries, and abnormal markers to each prediction point. Prediction confidence reflects the model's credibility of the current prediction value of the node, which is usually calculated from the standard deviation or confidence interval width output by the prediction model; upper and lower boundaries are used to characterize the possible maximum fluctuation range, such as the upper and lower limits of the 95% confidence interval; abnormal fluctuation markers are used to mark nodes that experience mutations, drastic jumps, or high fluctuations during the prediction period. These additional information together constitute a "time series feature set with uncertainty" that has higher expressiveness and stronger model robustness than traditional point prediction sequences.

[0120] In order to integrate the above-mentioned time series feature set into the graph structure, the mapping problem between nodes and time must be solved. The present invention adopts a dual index mechanism of node index and time index to explicitly bind the features of each node at each prediction time point to the corresponding node in the graph. This means that in the input of the graph neural network, it is no longer just a static attribute vector provided for each node, but the prediction sequence and its confidence index in the current time window of the node are dynamically loaded for each rolling cycle. In order to ensure the consistency of the time series structure, the system introduces a sliding time window mechanism of fixed length to segment the load forecast sequence according to the rolling cycle. For example, if the rolling cycle is 15 minutes and the prediction window is 2 hours, only the 96 prediction points of the current 2 hours are loaded as part of the graph input each time the sliding occurs, thereby ensuring that the time window used in each round of network training or reasoning remains consistent, and improving the stability of the model in the time series dimension.

[0121] In the graph neural network part, the system uses a multi-layer graph convolution structure to update and fuse the node states in the graph. These graph convolution operations not only act on static electrical topology structures, such as the connection relationship between nodes, the impedance of edges, voltage levels, etc., but also process the timing load characteristics of dynamic loading. Through convolution operations, the model can aggregate the timing evolution information from adjacent nodes, and learn the coupling effect between nodes in the time dimension while maintaining structural information. For example, an abnormal increase in future load at a node may cause a power shock to its neighboring transformers or branches. This dependency can be automatically modeled through graph convolution and identified and responded to in subsequent optimization.

[0122] Finally, after multi-layer graph convolution calculations, each node will obtain a state representation that integrates its local load trend, the dynamic impact of neighboring nodes, and structural context constraints. This state representation not only reflects the current operating environment of the node, but also includes the coupling risks that may be caused by future load fluctuations, which is crucial for the rolling optimization module to build a control strategy.

[0123] In summary, this embodiment integrates the dynamic evolution of predicted load, the uncertainty expression of the model, and the structural dependence of the power grid topology into a unified graph structure, which not only enhances the generalization ability and predictive interpretability of the model, but also provides a high-dimensional and high-quality input foundation for subsequent graph-based scheduling optimization.

[0124] Furthermore, the process of generating a scheduling control strategy through evolutionary strategy search includes: Based on the extracted node state features, a scheduling encoding vector including the state of the control object, constraints, scheduling action set and initial strategy score is constructed to initialize the population or strategy library; In each rolling cycle, the scheduling action sequence is continuously optimized through cross-mutation and strategy screening mechanisms. Risk buffer distance and equipment protection margin are introduced as implicit penalty items in the strategy generation process to guide the search towards a robust operation path. For each output joint operation sequence, a scheduling area mask, time period priority label and policy coverage range identifier are attached to support the subsequent scheduling execution system to perform regional distribution, priority sorting and policy switching judgment, and realize distributed and hierarchical execution control.

[0125] In this embodiment, in order to achieve joint optimization scheduling of multiple distribution equipment and control resources, the system introduces an optimization method based on evolutionary strategy search, which is used to dynamically generate feasible and robust control strategy sequences within the rolling prediction cycle. The basic idea is to encode the state, physical constraints, control action set and historical execution feedback of the controllable objects in the power grid into a unified strategy representation, and use the evolutionary search algorithm to iteratively optimize in a solution space containing multiple candidate strategies.

[0126] Specifically, the system first constructs a scheduling encoding vector for each rolling cycle based on the node state features extracted by the graph neural network. The encoding vector not only contains the load state of the current node or line, the remaining adjustment capacity of the equipment, and the known physical boundary constraints, but also includes control constraint information such as equipment action frequency limit, linkage rules or regulation order. The scheduling action set covers transformer tap position adjustment, capacitor bank switching control, feeder tie switch opening and closing selection, energy storage unit charging and discharging strategy, and adjustable load response level. For each type of control resource, the system constructs a set of legal actions based on its current state and predicted reachable range. All these elements are packaged together to form a structured encoding vector, which is used to initialize the population or strategy library in the evolutionary algorithm as the starting point for strategy search.

[0127] During the evolutionary search process, the system uses genetic algorithms, reinforcement learning or other heuristic methods to perform crossover mutation and fitness evaluation on the strategy set. In the crossover stage, fragments of different strategy codes are combined to form a new operation sequence; in the mutation stage, slight disturbances are introduced to discover potential optimal solutions. In order to make the generated strategy more robust and engineering feasible, the system introduces "implicit penalty terms" in the fitness evaluation process, including two indicators: risk buffer distance and equipment protection margin. The risk buffer distance is used to measure whether the system is close to the voltage or load operating limit after the strategy is executed; the equipment protection margin reflects whether the strategy operates within the acceptable action frequency, overload duration or temperature rise range of the equipment. If a strategy meets the objective function (such as minimum voltage deviation, balanced load distribution, etc.) while also having strong risk isolation capabilities and protection tolerance, then its fitness will be given a higher weight, so that it will be retained first and used for subsequent iterations.

[0128] Finally, the system will output a set of optimized joint operation sequences before the end of the rolling cycle. Each operation sequence not only specifies the operation type and execution parameters, but also comes with multiple dispatch-related attributes to support actual deployment and instruction distribution. These attributes include dispatch area masks, which indicate the feeders or substations within which the operation should take effect; time period priority labels, which are used to sort execution priorities according to urgency, scope of impact, or intensity of behavior drivers when multiple operations coexist; and policy coverage identifiers, which are used to mark high-risk nodes or potential fault areas targeted by the policy, so that the dispatch center can perform local intervention, distributed execution, or centralized coordination according to the risk level.

[0129] Through the above process, this embodiment not only realizes the multi-objective global optimization of the dispatching strategy, but also introduces the adaptive adjustment capability for the complex dynamic characteristics of the power grid through the evolutionary mechanism, which can effectively handle large-scale, diversified, and continuously changing control scenarios.

[0130] Step S105: Send the dispatch control strategy to the corresponding device control unit, and perform voltage control, load reconstruction, energy regulation or user load response before the prediction period according to the dispatch control strategy.

[0131] The core of step S105 is to convert the dispatching control strategy generated in the previous step into control instructions that can be accurately executed by on-site equipment, and send these instructions to the corresponding equipment control units, so as to complete the pre-regulation of the power grid before the load fluctuation actually occurs, and realize dynamic optimization control of voltage, power distribution and energy utilization.

[0132] In this step, we first need to clarify the target execution object of the dispatching control strategy. Usually, it includes various intelligent devices installed in the distribution network, such as transformer tap adjustment mechanism, capacitor switching controller, remote switching control unit of feeder tie switch, energy storage converter controller, and user-side energy consumption management system or intelligent terminal with two-way communication capability. These devices must have remote controllability and be connected to the main dispatching center or edge control node through a communication link. The communication mode can be wired (such as optical fiber, PLC) or wireless (such as 4G / 5G, LoRa, Wi-SUN, etc.), and are required to have appropriate real-time and anti-interference capabilities.

[0133] The dispatch control strategy is issued by the central dispatch system or the distribution automation master station. The system will call the corresponding control logic module according to the control interface protocol of each type of equipment, and parse the abstract strategy content into standardized control instructions that comply with the field equipment control protocol. For example, the strategy of "transformer T1 adjusts the tap from +2 to +1 at 15:30" is converted into the Modbus TCP message format supported by the voltage regulator controller, with a checksum and timestamp to ensure accurate execution and traceability of the instruction.

[0134] In terms of time control, all control instructions must specify their execution time points and be strictly aligned with the predicted period. To this end, the system introduces a time synchronization mechanism to ensure that all terminal devices operate with a unified clock reference. The system can use IEEE1588 Precision Time Protocol (PTP) or GPS timing to ensure that scheduling behaviors are executed synchronously at the millisecond level. Especially for multi-point coordinated actions such as feeder reconstruction or demand response, time consistency is particularly important for operational safety.

[0135] Voltage control is one of the most common strategies. The dispatch command can make the transformer automatically perform voltage step-up or step-down operations, and monitor in real time through the feedback loop whether the adjusted voltage value meets the expected target. If the adjustment fails, the system has an automatic fallback mechanism to restore to a safe position to avoid abnormal voltage expansion.

[0136] In terms of load reconfiguration, the instructions issued can control the closing or opening order of specific contact switches, thereby changing the distribution relationship of loads between feeders. For example, when a feeder is predicted to have an overload risk, the system can divert part of the load to the adjacent node through the backup feeder in advance. The instruction content must clearly specify the switch number, action delay time and reclosing protection setting.

[0137] For energy regulation operations, such as controlling the charging and discharging behavior of distributed energy storage systems, the dispatching system will send specific charging and discharging power values, start and end times, power curve control methods (constant power, tracking or predictive) and other parameters to the energy storage controller based on the optimization results. The system will also collect and dynamically adjust key state quantities such as SOC (state of charge), temperature, and voltage during the charging and discharging process in real time to avoid overcharging or thermal runaway.

[0138] In terms of user load response, the dispatch system can send load adjustment signals to users or aggregators participating in demand response through a two-way communication platform. The method can be automatic control (such as disconnecting non-critical loads, adjusting refrigeration set temperature) or guidance based on incentive mechanisms (such as sending temporary electricity price adjustment notifications, electricity consumption reward reminders, etc.). The user terminal should have the ability to receive dispatch signals and automatically execute response strategies, and the response results must be fed back to the main system for evaluating the response effect and dispatch achievement rate.

[0139] The entire scheduling strategy issuance process has a confirmation and receipt mechanism. After each instruction is issued, it is necessary to receive the execution status feedback sent back by the device, including success, failure, partial execution or manual confirmation. The system can adjust the strategy or initiate backup instructions accordingly. At the same time, the system supports archiving of issuance records, including timestamps, device numbers, execution parameters and response results, to facilitate post-audit, operation and maintenance analysis and strategy improvement.

[0140] In summary, step S105 is not only a command transmission process, but also a key link to deeply integrate the prediction-driven optimization strategy with the actual operation of the distribution network.

[0141] The second embodiment of the application provides an electronic device, the electronic device comprising: processor; The memory is used to store a program, and when the program is read and executed by the processor, it executes an intelligent distribution load prediction and adaptive scheduling method provided in the first embodiment of the present application.

[0142] The third embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, an intelligent distribution load forecasting and adaptive scheduling method provided in the first embodiment of the present application is executed.

[0143] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. An intelligent power distribution load forecasting and adaptive scheduling method, characterized in that: include: Obtain historical load data, real-time meteorological information, and user electricity consumption behavior data in the target distribution area, and establish a comprehensive load forecast input data set; Based on the comprehensive load forecast input data set, a short-term load forecast model is constructed to forecast the load levels of multiple feeders and key nodes in a preset future period, and obtain the forecast load time series at the node level and feeder level; According to the predicted load time sequence, determining the voltage deviation, load imbalance or equipment overload risk that may occur at each node and feeder within the predicted period; Based on the grid topology, electrical parameter constraints and load forecast timing, a rolling time domain optimization method is used to generate a dispatch control strategy, which includes the early adjustment of transformer tap positions, the advance switching plan of capacitor banks, the reclosing sequence and load distribution logic of feeder tie switches, the charging and discharging time and power setting of distributed energy storage devices, or the demand response execution instructions for the user side; The dispatching control strategy is sent to the corresponding device control unit, and voltage control, load reconstruction, energy regulation or user load response is performed before the predicted time period according to the dispatching control strategy.

2. The intelligent power distribution load prediction and adaptive scheduling method according to claim 1 is characterized in that: The acquisition of historical load data, real-time meteorological information and user electricity consumption behavior data in the target distribution area and the establishment of a comprehensive load forecast input data set include: Obtain historical data of active and reactive loads from multiple monitoring points within a time window, classify them by feeder and node structure partitions, and perform missing value completion, anomaly elimination, and trend stability analysis to form a basic load sample set with complete time series; Receive and integrate real-time meteorological information from multiple sources, including temperature, humidity, wind speed, sunshine intensity and weather type, map meteorological data to the coverage area of ​​each feeder through spatial interpolation algorithm, and construct a meteorological feature set with geographical annotations; Collect user-side electricity consumption behavior information, including periodic electricity consumption patterns, peak and valley electricity consumption response habits, electricity price sensitivity, and equipment operation logs, perform cluster modeling based on user categories, and generate a behavior label feature matrix; Based on the load sample set, meteorological feature set and behavior label feature matrix, a unified multi-dimensional input data structure is constructed, and the data set is dynamically updated according to the rolling forecast cycle to drive the subsequent short-term load forecasting model.

3. The intelligent power distribution load prediction and adaptive scheduling method according to claim 2 is characterized in that: The short-term load forecasting model includes a topology encoding submodule, a feature fusion submodule, a time series prediction submodule and an output reconstruction submodule; The topology encoding submodule takes the grid topology structure of the target distribution area as input, uses graph neural network to construct node vector embedding representation with electrical attributes, and encodes the connection relationship, impedance parameters, equipment constraints and physical position relationship of each feeder and node in the topology into a low-dimensional continuous vector as an explicit expression of the impact of the topology structure on load forecasting. The feature fusion submodule takes the topological coding results, historical load time series, meteorological features after spatial interpolation and behavior label matrix as joint inputs, adopts a multi-layer perception fusion mechanism, extracts key features based on the attention mechanism and feature gating mechanism, eliminates scale differences and semantic conflicts between data, and outputs a multi-dimensional dynamic feature vector sequence of uniform length as a comprehensive driving factor for load changes; The time series prediction submodule takes the fused multi-dimensional dynamic feature vector sequence as input, and uses a bidirectional gated recurrent unit network enhanced by time series attention to predict the node-level and feeder-level loads step by time. By introducing an adaptive time window mechanism and an abnormal state dynamic correction structure, the model has the ability to identify short-term load mutation events and actively suppress prediction errors. The output reconstruction submodule maps the predicted value back to the physical number of each node or feeder according to the relationship between the timing prediction results and the actual configuration of the power grid nodes, and adds confidence indicators, upper and lower bounds of the prediction error, and typical power consumption scenario labels to provide multi-level prediction information support for subsequent scheduling strategy optimization and execution control.

4. The intelligent power distribution load prediction and adaptive scheduling method according to claim 3 is characterized in that: The abnormal state dynamic correction structure in the time series prediction submodule performs node-level dynamic adjustment on the initial prediction value through the following formula 1: ； in, Indicates the corrected Nodes at time The predicted load value; For the Nodes at time The initial predicted value of For Node The individual correction coefficient is obtained by training based on the historical load variance and prediction confidence of the node; It represents the overall load variability index at the current moment, which is dynamically calculated based on the standard deviation of the forecast load series of the entire network; is the node sensitivity factor to abnormal load, which is obtained by statistical regression of the load offset frequency and amplitude of the node in the past meteorological disturbance period; It is the behavior disturbance index of the node at the current moment, which is constructed based on the fluctuation feature dimension in the behavior label matrix and reflects the degree to which the user behavior drives the prediction deviation. For Node The topological embedding vector of is generated by the topological encoding submodule; It is the dynamic weight vector of the behavior feature, calculated by the feature fusion module at the current time step, and is used to measure the local contribution weight of the behavior label to the load forecast.

5. The intelligent power distribution load prediction and adaptive scheduling method according to claim 3 is characterized in that: The feature fusion submodule includes a feature construction structure based on the behavioral meteorology joint decoupled attention mechanism, and generates the fusion input representation of each node through the following formula 2: ； in, Representation Node At the moment The fused input vector is used as the input of the time series prediction submodule; Represents the nodes extracted from the behavior label feature matrix At the moment The behavior driving vector; Indicates the meteorological characteristics after spatial interpolation at the node The local environmental factor vector at ; They are respectively the trainable weight matrices used for decoupling transformation of behavior and meteorological characteristics; For Node The individual behavior weight coefficient is calculated using the following formula 3: ； in, Representation Node The historical behavior dominance index is obtained by statistically learning the explanation intensity of user behavior on load changes in the past scheduling cycle; Representation Node The meteorological sensitivity parameters are estimated based on the partial derivatives of node load changes during historical meteorological upheavals.

6. The intelligent power distribution load prediction and adaptive scheduling method according to claim 1 is characterized in that: The step of determining, based on the predicted load time sequence, the voltage deviation, load imbalance or equipment overload risk that may occur at each node and feeder within the predicted period includes: The predicted load time series is mapped to the distribution network topology, and the predicted voltage response trend is derived by combining the node voltage regulation capability, electrical connection relationship and load conduction path; Based on the prediction results of three-phase load distribution, the feeder phase imbalance rate is calculated, and the continuous imbalance evolution trend is identified in combination with the historical asymmetric operation records; According to the rated capacity of the equipment and the load evolution rate, the load rate of each node and feeder within the prediction interval is analyzed in time series to determine whether it crosses the equipment operation safety threshold; Cross-compare the judgment results with the load surge characteristics in the behavior labels, screen out highly sensitive nodes that are significantly affected by behavior and have overload risks, and build a node-level risk list.

7. The intelligent power distribution load forecasting and adaptive scheduling method according to claim 1 is characterized in that: The method of generating a dispatch control strategy based on the judgment results of the power grid topology, electrical parameter constraints and load forecast timing by using a rolling time domain optimization method includes: The grid topology of the target distribution area is mapped into a graph data structure using a graph structure modeling method, where nodes represent feeder connection points, transformer locations or load access points, and edges represent line connection relationships. The line impedance, voltage level, and switch state electrical properties are considered to construct a graph topology representation that integrates topology and parameters. The node-level load forecast time series obtained according to the load forecasting model is embedded into the above graph structure as a dynamic additional attribute of the time evolution feature. The load response feature distribution under the dependency of the whole network structure is extracted through the graph neural network to capture the coupling effect between system levels. Based on the extracted state characteristics of the nodes in the entire network, a rolling time domain optimization framework is introduced to construct a multi-objective nonlinear optimization problem in each rolling cycle. Minimization of voltage deviation, suppression of feeder imbalance, reduction of equipment overload risk and minimization of load adjustment cost are used as joint objective functions, and the operation boundaries, action frequency constraints and chain action logic of various electrical equipment are integrated as constraint conditions. Through evolutionary strategy search, a joint operation sequence covering different control resources is dynamically generated in each rolling cycle, including transformer tap adjustment, capacitor bank switching, tie switch opening and closing, energy storage unit power allocation and user load priority sorting. Each strategy is attached with its execution window, predicted risk area and priority level information to ensure that the dispatch strategy has regional awareness, adaptability and high timeliness.

8. The intelligent power distribution load forecasting and adaptive scheduling method according to claim 7 is characterized in that: The graph structure modeling method includes: The initial topology diagram is generated based on the distribution network GIS data and the equipment master data in the SCADA system. The node type is determined by the equipment attribute field analysis, and the electrical parameters such as line impedance and voltage level are supplemented by real-time monitoring data and standard parameter library. The topology connectivity verification algorithm is used to identify the breakpoints, island nodes or loop structures in the graph, and the graph structure is closed in combination with the electrical operation rules to ensure the integrity and decoupling of the topology graph. Structured parameter vectors are embedded in the nodes and edges of the topology graph respectively. The node parameters include access capacity, node type and control capability identification, and the edge parameters include impedance value, maximum current capacity and current switch state, so as to realize the fusion expression of topological structure and operating characteristics.

9. The intelligent power distribution load forecasting and adaptive scheduling method according to claim 7 is characterized in that: The process of embedding the node-level load forecast time series into the graph structure includes: The load values ​​of each node at multiple prediction moments are constructed as dynamic feature vectors, and load confidence, prediction upper and lower bounds, and abnormal fluctuation signs are added to form a time series feature set with uncertainty. The time series feature set is mapped one-to-one with the topology graph nodes through the node and time dual index mechanism, and the prediction sequence is segmented in a rolling manner using the time window mechanism to ensure that the graph neural network input has time consistency within each rolling cycle. A multi-layer graph convolution structure is used to fuse static topological attributes and dynamic load characteristics, so that the node state representation not only reflects the evolution trend of its local load, but also reflects the coupling influence with neighboring nodes in structure and timing, providing context-related input for rolling optimization.

10. The intelligent power distribution load forecasting and adaptive scheduling method according to claim 7, characterized in that: The process of generating a scheduling control strategy through evolutionary strategy search includes: Based on the extracted node state features, a scheduling encoding vector including the state of the control object, constraints, scheduling action set and initial strategy score is constructed to initialize the population or strategy library; In each rolling cycle, the scheduling action sequence is continuously optimized through cross-mutation and strategy screening mechanisms. Risk buffer distance and equipment protection margin are introduced as implicit penalty items in the strategy generation process to guide the search towards a robust operation path. For each output joint operation sequence, a scheduling area mask, time period priority label and policy coverage range identifier are attached to support the subsequent scheduling execution system to perform regional distribution, priority sorting and policy switching judgment, and realize distributed and hierarchical execution control.

Images