A power grid load prediction method and system based on a large model

Abstract

The application discloses a power grid load prediction method and system based on a large model, and relates to the technical field of power grid load prediction.The method comprises the following steps: S1, collecting multi-source load comprehensive data, and performing edge processing and data preprocessing; S2, monitoring a mutation point in real time, constructing a behavior label sequence, and identifying a high-frequency disturbance point; S3, constructing a prediction input vector of the multi-source load comprehensive data, performing load data prediction, and calculating a physical constraint regularization value; S4, performing dynamic credible interval evaluation, judging the prediction result in real time according to a credible boundary value interval, and feeding back a control instruction; and S5, performing error analysis and retraining triggering, comprehensively evaluating prediction accuracy and output stability, and performing model optimization, version management and feedback learning.The method solves the problem that a prediction model is prone to oscillation due to fast load change in a high-density load scene.

Description

Technical Field

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

[0002] With the development of power system construction and smart city development, power grid operation and dispatch face increasingly complex load change patterns, especially in high-density urban areas, industrial parks, and commercial districts, where power load exhibits strong volatility, periodic overlap, and event-driven characteristics. With the development of big data and artificial intelligence technologies, large-scale time series modeling architectures have demonstrated powerful modeling capabilities in the fields of natural language processing and energy forecasting.

[0003] For example, invention patent CN112330009B discloses a power grid load forecasting method, relating to the field of power grid load forecasting technology. The method includes: acquiring historical load data, wherein the historical load data includes load data for multiple historical time periods; the duration of each historical time period is the same as the duration of the time period to be predicted, the historical time periods being historical days or historical weeks, and the multiple historical time periods being continuous with the time period to be predicted; forecasting standardized load data for the time period to be predicted using a preset method based on the historical load data; determining a first average load and a second average load for the time period to be predicted using a first forecasting method and a second forecasting method, respectively, based on the historical load data; determining a weighted average load for the time period to be predicted based on the first average load and the second average load; and determining the load data for the time period to be predicted based on the weighted average load and the standardized load data. This invention can improve the accuracy of power grid load forecasting.

[0004] For example, invention patent CN112101663B discloses a power grid load forecasting method, involving research on a power grid load verification and forecasting strategy based on big data and a partially observable Markov decision process. It uses historical load data collected by the power grid data center as a foundation, while considering load type classifications such as off-peak loads and transferable loads. Using a partially observable Markov decision process as the objective function, combined with an orthogonal matching pursuit algorithm, it establishes a mathematical model and performs current load analysis, verification, and short-term load forecasting, generating a day-ahead load analysis report. This research enables power grid load verification and short-term power grid load forecasting. The dispatch center can verify the collected load data and effectively schedule some controllable loads day-ahead based on the results, thereby achieving optimal economic operation for users and the power grid.

[0005] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0006] Existing power grid load forecasting methods generally suffer from insufficient modeling capabilities, mainly reflected in the lack of multi-source data fusion, making it difficult to respond to load disturbances in high-frequency scenarios caused by industrial equipment status, weather changes, and abnormal events; shallow model structure, failing to effectively capture long-term dependencies and changes during critical periods; at the same time, the forecast curve is prone to oscillation, affecting the stability of system scheduling; and the lack of health assessment and retraining mechanisms makes it difficult to adapt to long-term operation.

[0007] Therefore, in order to address the above problems, there is an urgent need for a power grid load forecasting method and system based on a large model. Summary of the Invention

[0008] Technical problems to be solved

[0009] To address the shortcomings of existing technologies, this invention provides a power grid load forecasting method and system based on a large model, which solves the problem of forecasting models being prone to oscillation due to the rapid load change rate in high-density load scenarios.

[0010] Technical solution

[0011] To achieve the above objectives, the present invention provides the following technical solution: a power grid load forecasting method based on a large model, comprising the following steps: S1, real-time acquisition of multi-source load integrated data, edge processing of the multi-source load integrated data, and data preprocessing; S2, real-time monitoring of abrupt change points, construction of behavioral label sequences to extract load change patterns, identification of high-frequency disturbance points, and implementation of optimization measures based on the identification results; S3, construction of a prediction input vector for multi-source load integrated data, load data prediction, introduction of physical constraint terms, and calculation of physical constraint regularization values ​​to determine whether the prediction results of the neural network time series prediction model violate physical boundary conditions and whether there are drastic jumps; S4, fusion of prediction results, dynamic confidence interval evaluation based on a sliding time window, real-time determination of prediction results based on the confidence boundary value interval, and feedback of control commands; S5, error analysis and retraining triggering, comprehensive evaluation of prediction accuracy and output stability, and optimization, version management, and feedback learning of the neural network time series prediction model.

[0012] Furthermore, the specific process of real-time acquisition of multi-source load integrated data, aggregation of multi-source load integrated data for edge processing, and data preprocessing is as follows: Real-time acquisition of multi-source load integrated data includes: deploying smart meters at load-dense points to collect load data; deploying regional micro-meteorological sensors to collect environmental data, including temperature and humidity data; accessing programmable logic controllers and manufacturing execution systems to collect industrial load behavior data, including equipment start-up and shutdown, and industrial equipment status; obtaining high-density activity information related to regional holidays and promotional information from the city event open platform interface; connecting with the data acquisition and monitoring control systems of factories and parks to collect equipment operating status, energy consumption data, and transformer capacity; implementing timestamp alignment to unify to minute-level granularity; performing event coding processing on unstructured data of high-density activities; using an edge computing gateway to process preliminary multi-source load integrated data compression, filtering, and caching; using linear interpolation to fill missing values, using the standard score method to remove outliers, and standardizing and normalizing the multi-source load integrated data before uploading it to the multi-source load integrated database.

[0013] Furthermore, the specific process of real-time monitoring of abrupt change points and constructing behavioral label sequences to extract load change patterns and identify high-frequency disturbance points is as follows: The volatility of load data is calculated using a sliding time window to monitor short-term load data change trends. A Bayesian mutation detection model is used to identify probabilistic structural mutations in the load data fluctuation curve in real time. The anomaly judgment threshold is adaptively fine-tuned based on the current time period and regional load density. Daily load data fluctuation curves are converted into symbolic pattern labels according to time period characteristics, and structurally expressed in units of behavioral segments. Unsupervised clustering algorithms are used to perform cluster analysis on these label sequences to identify potential periodic structural or abnormal pattern changes. Real-time load data is acquired, obtaining the load data at the current moment and the load data at the previous moment. Based on the sliding time window length... The average load data is obtained by calculating the mean value of the load data collected at various times within the specified range; real-time temperature data is acquired, and the temperature data at the current time and the temperature data at the previous time are obtained by subtracting the temperature data at the two times; the distribution of temperature data changes over a year in the multi-source load integrated database is statistically analyzed, and a temperature disturbance reference scale is determined based on the fact that more than 95% of the temperature data fluctuations do not exceed a certain value; the absolute change of load data is obtained by subtracting the load data at the previous time from the load data at the current time and taking the absolute value; the absolute change of load data is divided by the average load data to obtain the load fluctuation rate; the ratio of the temperature difference to the temperature disturbance reference scale is added to a constant to obtain a temperature influence term; the load fluctuation rate is multiplied by the temperature influence term to obtain the high-frequency disturbance point discrimination value.

[0014] Furthermore, the specific process of implementing optimization measures based on the identification results is as follows: The high-frequency disturbance point discrimination value is compared with the first and second-level disturbance thresholds in real time. When the high-frequency disturbance point discrimination value is less than or equal to the disturbance threshold, the current load fluctuation is considered stable, and the current neural network time-series prediction model continues to be used for load prediction. Prediction caching and lightweight inference models are appropriately enabled, and edge computing nodes are run at reduced frequencies. The current stable segment is marked as a normal sample. When the high-frequency disturbance point discrimination value is greater than the first-level disturbance threshold but less than or equal to the second-level disturbance threshold, it indicates that the load change relative to the average value exceeds 20%, and the prediction cycle is shortened. To enhance instantaneous response, edge computing nodes record abnormal periods of high-frequency disturbance point discrimination values ​​in real time, marking them as disturbance segments. The platform triggers a prediction accuracy warning state and displays the area requiring attention. When the high-frequency disturbance point discrimination value exceeds the secondary disturbance threshold, it indicates a drastic change in load accompanied by a dramatic change in the environment. The energy storage system is then coordinated for pre-scheduling, and the coordinated equipment performs load reduction as needed according to priority. Power consumption limit signals are issued through the programmable logic controller. If abnormal fluctuations occur and the cloud platform is unreachable, a protection strategy is initiated on the edge side: static prediction and standby value output are entered, and the disturbance segment is compared with urban events and industrial production scheduling data.

[0015] Furthermore, the specific process of constructing a prediction input vector from multi-source load composite data for load data prediction is as follows: A load data feature vector based on multi-source load composite data is constructed as the training set for the neural network time-series prediction model. Key features of multiple key dimensions of data, including historical load data sequences, industrial equipment operating status, environmental data, and high-density activity information, are extracted, encoded, and aligned, transforming them into corresponding load data feature vectors. These vectors are then slidably stitched together according to a time window to form a load data feature vector. The load data feature vector is organized in a sliding window structure along the time dimension to form a unified three-dimensional tensor. A neural network time-series prediction model with a converter model as its core structure is constructed using the load data feature vector. Backpropagation training is performed by minimizing the prediction error loss function to output predicted load data for several future time points.

[0016] Furthermore, a physical constraint term is introduced, and a physical constraint regularization value is calculated to determine whether the prediction results of the neural network time series prediction model violate physical boundary conditions and whether there are drastic jumps. The specific process is as follows: Transformer capacity, load data change rate, and safety margin are used as hard boundary conditions. A physical constraint regularization value is added to the training loss function. The predicted load data at the current moment is obtained, and the transformer capacity is used as the maximum allowable load of the region. The difference between the predicted load data at the current moment and the maximum allowable load of the region is calculated. The difference is divided by the maximum allowable load of the region to obtain the relative proportion by which the predicted value exceeds the maximum allowable load of the region. A capacity penalty term is obtained by squaring the relative proportion. A maximum function operation is performed on the capacity penalty term. That is, if the predicted load data at the current moment is less than or equal to the maximum allowable load of the region, the capacity penalty term is zero; if the predicted load data at the current moment is greater than the maximum allowable load of the region, the capacity penalty term is the square of the relative proportion. Based on a sliding time window, the predicted load at the current moment is used... The load data is differentiated with respect to time, and the result is squared and multiplied by the rate of change weighting factor to obtain the jump penalty term. The capacity penalty term after the maximum function operation is added to the jump penalty term to obtain the physical constraint regularization value. The physical constraint regularization value is compared with the low-risk threshold and the high-risk threshold in real time. When the physical constraint regularization value is less than or equal to the low-risk threshold, it indicates that the neural network time series prediction model is stable and has not touched the capacity boundary, and is in a safe state. The current neural network time series prediction model is maintained. When the physical constraint regularization value is greater than the low-risk threshold and less than or equal to the high-risk threshold, it indicates that there is a moderate risk, the predicted value is close to the transformer capacity, and a slight jump occurs. Peak shaving warning is activated, and flexible load is linked to enter smooth adjustment. When the physical constraint regularization value is greater than the high-risk threshold, it is judged as a serious risk. The predicted value has seriously exceeded the limit and is fluctuating violently. Energy storage discharge and load reduction commands are triggered, the current prediction result output is frozen, the moving average is called for replacement, and risk logs and alarm information are uploaded to the cloud scheduling platform simultaneously.

[0017] Furthermore, the specific process of integrating prediction results, performing dynamic confidence interval assessment based on a sliding time window, and making real-time judgments on prediction results based on the confidence boundary value interval, and feeding back control instructions is as follows: Based on the predicted load data, combined with the sliding time window, the oscillating predicted load data is processed using a medium-range filter; the confidence boundary value interval of the prediction results is calculated; the predicted load data at the current moment is obtained; through a multi-source load comprehensive database and based on the sliding time window, the load data of the most recent n moments is obtained, and the average value of the load data of the most recent n moments is calculated, and the variance of the load data of the most recent n moments is obtained; the confidence factor constant term is obtained through the standard normal distribution table; the high-frequency disturbance point discrimination value at the current moment is calculated; the variance of the load data of the most recent n moments is multiplied by its own fluctuation weight factor to obtain the load fluctuation term, and the high-frequency disturbance point discrimination value at the current moment is multiplied by the control disturbance weight factor to obtain the disturbance discrimination term, and the load fluctuation term is compared with... The disturbance discrimination term is summed, and the summation result is multiplied by the confidence factor constant term to obtain the confidence interval width adjustment term. The upper limit of the predicted confidence boundary value is obtained by adding the confidence interval width adjustment term to the predicted load data at the current time, and the lower limit of the predicted confidence boundary value is obtained by subtracting the confidence interval width adjustment term from the predicted load data at the current time. The predicted confidence boundary value interval is obtained by combining the upper and lower limits of the confidence boundary value. Based on the predicted confidence boundary value interval, when the collected actual load data is not within the predicted confidence boundary value interval, the current neural network time series prediction model is determined to be inaccurate, an abnormal event reminder is pushed, and the backup historical mean prediction model is activated. When multiple consecutive actual load data fall outside the predicted confidence boundary value interval, the neural network time series prediction model is replaced, and disturbance segment samples are added for reinforcement learning. When the upper limit of the predicted confidence boundary value is greater than or equal to the transformer capacity, the energy storage equipment is dispatched, some non-critical loads are shut down, and the energy-saving operation mode is entered.

[0018] Further, the specific process of error analysis and retraining triggering to comprehensively evaluate prediction accuracy and output stability is as follows: Monthly statistical analysis is performed on the residuals between predicted values ​​and actual load data to evaluate key error indicators and compare them with set performance thresholds. If errors exceed limits for two consecutive months, the data is automatically marked as performance degradation and included in the retraining candidate queue. The predicted load data at time i is obtained, and the actual load data at time i is obtained from the multi-source load database. Based on the predicted load data at time i and the average of the predicted load data at all times within the sliding time window, the variance of the predicted load data within the sliding time window is calculated to obtain the prediction volatility. The difference between the predicted load data at time i and the actual load data at time i is calculated. The difference is divided by the actual load data at time i, and the absolute value is taken to obtain the relative error. The sum of the relative errors at all times is divided by the total time length to obtain the average relative error term. The prediction volatility is multiplied by the jump penalty weighting factor to obtain the jump penalty term. The average relative error term and the jump penalty term are subtracted from a constant to obtain the health assessment value.

[0019] Furthermore, the specific process of optimizing the neural network time series prediction model and managing its version and feedback learning is as follows: The health assessment value is compared with the health assessment threshold in real time. When the health assessment value is less than the health assessment threshold, it indicates that the prediction has a large error and output oscillation problem, triggering an adjustment mechanism, including freezing abnormal prediction segments, reducing the control response frequency, and simultaneously marking the current neural network time series prediction model as needing fine-tuning, entering the training re-evaluation process; when the health assessment value is greater than or equal to the health assessment threshold, it indicates that the prediction is accurate and the fluctuation is stable, the current neural network time series prediction model continues to be used, its performance indicators are recorded as reference standards, and it is allowed to participate in core scheduling tasks; each update records the version number, indicator changes, and training data hash value, supporting rollback and comparative experiments.

[0020] The second aspect of this invention provides a large-scale model-based power grid load forecasting system, comprising a multi-source sensing and data preprocessing module, a disturbance identification and trend detection module, a physical fusion large-scale model prediction engine module, a steady-state fusion and feedback control module, and a model optimization and evolution module. The multi-source sensing and data preprocessing module is used to collect multi-source load comprehensive data in real time, aggregate the multi-source load comprehensive data for edge processing, and perform data preprocessing. The disturbance identification and trend detection module is used to monitor abrupt change points in real time, construct behavioral label sequences to extract load change patterns, identify high-frequency disturbance points, and implement optimization measures based on the identification results. The physical fusion large-scale model prediction engine module… The system is divided into four modules: a prediction input vector for constructing multi-source load composite data, a load data prediction module, a physical constraint term, and a physical constraint regularization value to determine whether the prediction results of the neural network time series prediction model violate physical boundary conditions and whether there are drastic jumps; a steady-state fusion and feedback control module for fusing prediction results, performing dynamic confidence interval evaluation based on a sliding time window, making real-time judgments on prediction results according to the confidence boundary value interval, and providing feedback control instructions; and a model optimization and evolution module for error analysis and retraining triggering, comprehensively evaluating prediction accuracy and output stability, and optimizing the neural network time series prediction model, as well as version management and feedback learning.

[0021] Beneficial effects

[0022] The present invention has the following beneficial effects:

[0023] (1) This invention integrates multi-source heterogeneous data such as load data, industrial operation status, environmental data, urban events and holiday information to construct a unified prediction input vector, breaking through the limitation of traditional methods that rely solely on single load data for modeling. This significantly enhances the model's ability to respond to abnormal events and non-periodic disturbances, thereby improving prediction accuracy and the system's adaptability to complex scenarios.

[0024] (2) This invention introduces physical constraints such as transformer capacity, safety margin and load change rate during the neural network training process, and makes real-time judgments through physical constraint regularization values ​​to prevent the predicted values ​​from exceeding the limits or drastic changes, thereby improving the physical consistency of the model output and the practical feasibility of power system dispatch, and significantly enhancing the safety and controllability of the prediction results in engineering.

[0025] (3) This invention calculates volatility through a sliding window, introduces Bayesian mutation detection and behavioral label sequence clustering, identifies high-frequency disturbance points in real time, and sets up a multi-level response mechanism to achieve dynamic adaptation of model structure and operation strategy, effectively improving the stability and robustness of the system under severe disturbance scenarios.

[0026] (4) This invention introduces a health assessment value to quantitatively monitor prediction error and output fluctuation, combines monthly residual analysis to trigger a retraining mechanism, and integrates meta-information such as version number, performance index, and data hash to realize the model version management and feedback learning mechanism, ensuring the continuous performance optimization and risk traceability of the system in long-term operation.

[0027] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0028] Figure 1 This is a flowchart of a power grid load forecasting method based on a large model.

[0029] Figure 2 This is a structural diagram of a power grid load forecasting system based on a large model;

[0030] Figure 3 A trend comparison chart of predicted load data and physical constraint regularization values. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-3 This invention provides a technical solution: a power grid load forecasting method and system based on a large model, comprising the following steps: S1, real-time acquisition of multi-source load integrated data, edge processing of the multi-source load integrated data, and data preprocessing; S2, real-time monitoring of abrupt change points, construction of behavioral label sequences to extract load change patterns, identification of high-frequency disturbance points, and implementation of optimization measures based on the identification results; S3, construction of a prediction input vector for multi-source load integrated data, load data prediction, introduction of physical constraint terms, and calculation of physical constraint regularization values ​​to determine whether the prediction results of the neural network time series prediction model violate physical boundary conditions and whether there are drastic jumps; S4, fusion of prediction results, dynamic confidence interval evaluation based on a sliding time window, real-time determination of prediction results based on the confidence boundary value interval, and feedback of control commands; S5, error analysis and retraining triggering, comprehensive evaluation of prediction accuracy and output stability, and optimization, version management, and feedback learning of the neural network time series prediction model.

[0033] Specifically, the real-time acquisition of multi-source load data, the aggregation of this multi-source load data for edge processing, and the data preprocessing process are as follows: Real-time acquisition of multi-source load data includes: deploying smart meters at load-dense points to collect load data; deploying regional micro-meteorological sensors to collect environmental data, including temperature and humidity data; connecting to programmable logic controllers and manufacturing execution systems to collect industrial load behavior data, including equipment start-up and shutdown, and industrial equipment status; obtaining high-density activity information related to regional holidays and promotional information from the city event open platform interface; and connecting to the data acquisition and monitoring control systems of factories and industrial parks to collect equipment operation data. The system collects data on load status, energy consumption, and transformer capacity; implements timestamp alignment to a minute-level granularity to ensure data synchronization across multiple sources and avoid time-series misalignment; performs event encoding on high-density activity unstructured data, transforming text events into structured features recognizable by the model; uses an edge computing gateway to process preliminary multi-source load data compression, filtering, and caching; uses linear interpolation to fill in missing values, uses standard score method to remove outliers, and standardizes and normalizes the multi-source load data before uploading it to a multi-source load database as a unified data foundation for subsequent prediction and analysis.

[0034] In this implementation plan, multi-source data fusion and acquisition are achieved through smart meters, sensors, industrial systems and open platforms. Time alignment, event coding and edge computing are combined to ensure data synchronization and real-time performance. Interpolation completion and anomaly removal are used to improve data quality. Finally, a standardized unified data foundation is built to provide high-quality support for load forecasting and analysis.

[0035] Specifically, the process of real-time monitoring of abrupt change points and constructing behavioral label sequences to extract load change patterns and identify high-frequency disturbance points is as follows: The volatility of load data is calculated using a sliding time window to monitor short-term load data change trends for dynamic perception of load change intensity; a Bayesian mutation detection model is integrated to identify probabilistic structural mutations in the load data fluctuation curve in real time, improving the accuracy and response time of sudden jumps; the anomaly judgment threshold is adaptively fine-tuned based on the current time period and regional load density to enhance scenario adaptability and prevent false alarms; daily load data fluctuation curves are converted into symbolic pattern labels according to time period characteristics, and structurally expressed in units of behavioral segments. Unsupervised clustering algorithms are used to perform cluster analysis on these label sequences to identify potential periodic structural or abnormal pattern changes for behavioral modeling and strategy layering; real-time load data is acquired, obtaining the load data at the current moment and the load data at the previous moment; and the load data at each moment collected within the sliding time window is analyzed. The average load data is calculated from the mean value of the data and used as a baseline reference for fluctuations. Real-time temperature data is acquired, and the temperature data at the current moment and the temperature data at the previous moment are obtained. The difference between the two temperature data is used to detect the load response driven by environmental changes. By statistically analyzing the distribution of temperature data changes over a year in the multi-source load integrated database, and based on the fact that more than 95% of the temperature data fluctuations do not exceed a certain value, a temperature disturbance reference scale is determined to construct a temperature normalization scale, enhancing the comparability of indicators. The absolute change of load data is obtained by subtracting the load data at the previous moment from the current moment's load data and taking the absolute value. The absolute change of load data is divided by the average load data to obtain the load fluctuation rate, which reflects the relative intensity of the current fluctuation within the historical baseline. The ratio of the temperature difference to the temperature disturbance reference scale is added to a constant to obtain a temperature influence term. The load fluctuation rate is multiplied by the temperature influence term to obtain the high-frequency disturbance point discrimination value, which serves as an important indicator for triggering prediction model switching or early warning mechanisms.

[0036] The specific formula for the high-frequency disturbance point discrimination value is as follows:

[0037] ;

[0038] In the formula, D t P is the discriminant value for high-frequency disturbance points. t This represents the load data at the current moment; P t-1 This is the load data from the previous moment; This is based on the average load data of the most recent n times within the sliding time window; ΔT t T represents the temperature difference between the current moment and the previous moment. r It serves as a reference scale for temperature perturbations and is used for normalization.

[0039] In this implementation scheme, by integrating slip volatility calculation, Bayesian mutation detection and behavioral label clustering, high-frequency disturbances and periodic changes are accurately identified; by combining temperature disturbance factors to construct high-frequency disturbance point discrimination values, real-time perception and dynamic early warning of load mutations are achieved, providing key basis for model switching and strategy response, and significantly improving the system's sensitivity and adaptability.

[0040] Specifically, the process of implementing optimization measures based on the identification results is as follows: Real-time comparison of the high-frequency disturbance point discrimination value with the first and second-level disturbance thresholds. When the high-frequency disturbance point discrimination value is less than or equal to the disturbance threshold, the current load fluctuation is considered stable, and the current neural network time-series prediction model continues to be used for load prediction. Prediction caching and lightweight inference models are appropriately enabled, and edge computing nodes are run at reduced frequencies. The current stable segment is marked as a normal sample for subsequent model reinforcement learning and low-risk control baseline construction. When the high-frequency disturbance point discrimination value is greater than the first-level disturbance threshold but less than or equal to the second-level disturbance threshold, it indicates that the load change relative to the average value exceeds 20%. The prediction cycle is shortened, instantaneous response is strengthened, and edge computing nodes... The system records high-frequency disturbance point discrimination values ​​in real time during periods of abnormality, marking them as disturbance segments. This triggers a platform-level prediction accuracy warning and displays the area as a region requiring attention. This measure can improve the system's rapid adaptive capability under mild disturbances. When the high-frequency disturbance point discrimination value exceeds the secondary disturbance threshold, it indicates a drastic change in load accompanied by a dramatic change in the environment. The system then initiates pre-scheduling of the energy storage system, and the linked equipment performs load reduction as needed according to priority. Power consumption limit signals are issued through the programmable logic controller. If abnormal fluctuations occur simultaneously and the cloud platform becomes unreachable, a protection strategy is activated at the edge: static prediction and standby value output are entered, and the disturbance segment is compared with urban event and industrial production data to assist in tracing the root cause of the disturbance and optimizing subsequent model retraining sampling.

[0041] In this implementation plan, a dynamic optimization mechanism is achieved by using graded response to high-frequency disturbance point discrimination values, which includes switching prediction models, adjusting edge resources, and linking energy storage. This ensures low-consumption operation under stable conditions, strengthens rapid response and safety control under abnormal conditions, and provides data support for model retraining and disturbance source tracing, thereby improving the overall intelligence and robustness of the system.

[0042] Specifically, the process of constructing a predictive input vector from multi-source load data for load data prediction is as follows: A load data feature vector based on multi-source load data is constructed as the training set for a neural network time-series prediction model. Key features from multiple key dimensions of data, including historical load data sequences, industrial equipment operating status, environmental data, and high-density activity information, are extracted, encoded, and aligned to construct an input feature system with temporal memory and event response capabilities. These features are then transformed into corresponding load data feature vectors and assembled using a sliding window approach to form the load data feature vector. This method preserves the contextual relationships of the time series and the correlation before and after abrupt changes. The load data feature vectors are organized in a sliding window structure along the time dimension to form a unified three-dimensional tensor, providing structured input support for the deep learning model. A neural network time-series prediction model with a converter model as its core structure is constructed using the load data feature vectors. A self-attention mechanism is introduced to enhance the modeling ability for key periods and key variables. Backpropagation training is performed by minimizing the prediction error loss function to continuously optimize model parameters and improve fitting performance. Finally, predicted load data for several future moments are output, providing a forward-looking decision-making basis for system regulation and energy consumption management.

[0043] In this implementation scheme, a time-series input vector is constructed by fusing key features of multi-source load data. By combining a sliding window and a self-attention mechanism, the model's ability to perceive and predict load changes is improved. Ultimately, a structured and robust deep neural network time-series prediction framework is formed, which effectively supports accurate load forecasting and the implementation of forward-looking control strategies.

[0044] Specifically, the process of introducing physical constraint terms and calculating physical constraint regularization values ​​to determine whether the prediction results of the neural network time series prediction model violate physical boundary conditions and whether there are drastic jumps is as follows: Transformer capacity, load data change rate, and safety margin are used as hard boundary conditions to constrain the model output from exceeding the physical safety limit. A physical constraint regularization value is added to the training loss function to guide the model to learn the load change trend within a controllable range. The predicted load data at the current moment is obtained, and the transformer capacity is used as the maximum allowable load for the region. The predicted load data at the current moment is then compared with the maximum allowable load for the region. The difference is divided by the maximum allowable load of the region to obtain the relative proportion by which the predicted value exceeds the maximum allowable load of the region. Squaring this relative proportion yields a capacity penalty term, which quantifies the degree to which the predicted result exceeds the transformer capacity limit. A maximum function operation is performed on the capacity penalty term: if the predicted load data at the current moment is less than or equal to the maximum allowable load of the region, the capacity penalty term is zero; if the predicted load data at the current moment is greater than the maximum allowable load of the region, the capacity penalty term is the square of the relative proportion, reflecting the non-linear penalty for over-limit behavior. Based on a sliding time window, the predicted load data at the current moment is used... The jump penalty term is obtained by taking the derivative with respect to time, squaring the result, and multiplying it by the rate of change weight factor. This term penalizes short-term drastic changes in the prediction result, improving the smoothness of the output. The capacity penalty term after the maximum function operation is added to the jump penalty term to obtain the physical constraint regularization value, which is used to measure the acceptability of the current model output. The physical constraint regularization value is compared with the low-risk threshold and the high-risk threshold in real time. When the physical constraint regularization value is less than or equal to the low-risk threshold, it indicates that the neural network time series prediction model is stable and has not reached the capacity boundary, and is in a safe state. The current neural network time series prediction model is maintained. Okay; when the physical constraint regularization value is greater than the low-risk threshold and less than or equal to the high-risk threshold, it indicates that there is a moderate risk, the predicted value is close to the transformer capacity and there is a slight jump, the peak shaving warning is activated, and the flexible load is linked to enter smooth adjustment to avoid the potential risk from expanding; when the physical constraint regularization value is greater than the high-risk threshold, it is judged as a serious risk, the predicted value has seriously exceeded the limit and is fluctuating violently, the energy storage discharge and load reduction command is triggered, the current prediction result output is frozen, the moving average is called to replace it, and the risk log and alarm information are uploaded to the cloud scheduling platform at the same time to ensure the safe operation of the system and retain risk records for subsequent analysis and learning.

[0045] The specific formula for the physical constraint regularization value is as follows:

[0046] ;

[0047] In the formula, L p The physical constraint regularization value; t represents a time variable used to identify a specific prediction time. Forecast load data at the current moment; C max γ is the transformer capacity, i.e., the maximum allowable load of the area; γ is the rate of change weighting factor, which is calculated based on the predicted load data of the most recent n times and the sliding time window, and the difference value sequence and derivative square sequence are obtained by Bayesian optimization algorithm, with a range between 0.1 and 0.2. This indicates that the maximum function operation is performed on the capacity penalty term, if Less than or equal to C max If no penalty is incurred, 0 is returned. Greater than C max The penalty is calculated by squared the relative proportion of the excess.

[0048] The rate of change weighting factor was set to 0.15, and the maximum allowable load for the region was 1000. Different capacity penalty terms and jump penalty terms were obtained based on the predicted load data at different times, and the physical constraint regularization value was calculated. The physical constraint regularization value data is shown in Table 1.

[0049] Table 1. Physical Constraint Regularization Values

[0050]

[0051] like Figure 3 As shown in Table 1, this is a comparison chart of the trend between the predicted load data and the physical constraint regularization value provided in the embodiments of this application. Figure 3 It can be seen that, with the rate of change weight factor and the maximum allowable load of the region remaining unchanged, and the predicted load data at different times being different, the trends of the predicted load data and the physical constraint regularization value, as well as the trend comparison between the predicted load data and the physical constraint regularization value, are observed.

[0052] In this implementation plan, physical boundary conditions such as transformer capacity and load change rate are introduced to construct physical constraint regularization values, accurately assessing whether the prediction results exceed limits or experience drastic jumps; combined with graded risk assessment and dynamic response strategies, steady-state control and risk protection of the prediction model are achieved, significantly improving the system's safety, controllability, and engineering practicality.

[0053] Specifically, the process of integrating forecast results, performing dynamic confidence interval assessment based on a sliding time window, and making real-time judgments on the forecast results based on the confidence boundary value interval, and feeding back control instructions, is as follows: Based on the forecast load data, combined with the sliding time window, apply medium-mean-law filtering to process the oscillating forecast load data to suppress short-term forecast anomalies and improve time series smoothness; calculate the confidence boundary value interval of the forecast results; obtain the forecast load data at the current moment; and obtain the load data for the most recent n moments through a multi-source load integrated database and based on the sliding time window, and calculate the most recent n moments... The average load data is used to calculate the variance of the load data over the most recent n time periods. This variance reflects the historical fluctuation intensity and is used for dynamic adjustment of the confidence boundary. A confidence factor constant term is obtained from the standard normal distribution table to control the confidence level range of the confidence boundary. The high-frequency disturbance point discrimination value at the current time is calculated, reflecting the intensity of the disturbance environment at the current prediction time. The load fluctuation term is obtained by multiplying the variance of the load data over the most recent n time periods by its own fluctuation weighting factor. The disturbance discrimination term is obtained by multiplying the high-frequency disturbance point discrimination value at the current time by the control disturbance weighting factor. The load fluctuation term and the disturbance discrimination term are then compared. The confidence interval width adjustment term is obtained by summing the other terms and multiplying the summation result by the confidence factor constant term, forming the confidence interval range value. The upper limit of the prediction confidence boundary value is obtained by adding the confidence interval width adjustment term to the current predicted load data, and the lower limit of the prediction confidence boundary value is obtained by subtracting the confidence interval width adjustment term from the current predicted load data. The upper and lower limits of the confidence boundary value are combined to obtain the prediction confidence boundary value interval, which is used to dynamically judge the current prediction confidence level. Based on the prediction confidence boundary value interval, when the collected actual load data is not within the prediction confidence boundary value interval, the current neural network time series prediction model is determined to be inaccurate, an abnormal event reminder is pushed, and the backup historical average prediction model is activated to ensure prediction continuity and control safety. When multiple consecutive actual load data fall outside the prediction confidence boundary value interval, the neural network time series prediction model is replaced, and disturbance segment samples are added for reinforcement learning to improve the model's generalization ability in high disturbance environments. When the upper limit of the prediction confidence boundary value is greater than or equal to the transformer capacity, energy storage equipment is dispatched, some non-critical loads are shut down, and energy-saving operation mode is entered to realize risk pre-control and intelligent linkage response based on confidence boundary.

[0054] The specific formula for the reliable boundary value interval is as follows:

[0055] ;

[0056] In the formula, U t To predict the upper bound of the reliable boundary value; L t This serves as the lower bound for predicting the reliable boundary values. Z represents the predicted load data for the current moment. αThe confidence factor is used to control the confidence level of the confidence interval. It is a constant term obtained from the standard normal distribution table. For example, the confidence factor corresponding to the 95% confidence level is 1.96. D is the variance of the load data over the most recent n time periods; t ω1 is the high-frequency disturbance point discrimination value; ω2 is the self-fluctuation weighting factor, which is obtained by statistically analyzing the proportion of actual load data falling into the confidence boundary value interval within a certain period based on the predicted load data of the most recent n times, comparing it with the 95% and 99% confidence interval coverage targets, selecting the closest one, and fitting it using the Bayesian optimization algorithm, with a range between 0.2 and 0.6; ω3 is the control disturbance weighting factor, which is obtained by analyzing the accuracy of high-frequency disturbance point discrimination value identification within the confidence boundary value interval using the hierarchical error analysis method based on the high-frequency disturbance point discrimination value, through cross-validation and minimum residual strategy, with a range between 0.5 and 2.0.

[0057] In this implementation scheme, a dynamic reliability interval is constructed by combining a sliding window and a disturbance factor to evaluate the stability and reliability of the prediction results in real time. Combined with anomaly detection and boundary early warning mechanisms, the system achieves linkage between model inaccuracy response, backup model switching and energy storage regulation, thereby improving the prediction robustness and operational safety of the system under uncertain environments.

[0058] Specifically, the process of error analysis and retraining triggering to comprehensively evaluate prediction accuracy and output stability is as follows: Monthly statistical analysis is performed on the residuals between predicted values ​​and actual load data to evaluate key error indicators and compare them with set performance thresholds to identify model performance trends. If errors exceed limits for two consecutive months, the model is automatically marked as performance degradation and included in the retraining candidate queue to ensure its self-repair and evolution capabilities during long-term operation. The predicted load data at time i is obtained, and the actual load data at time i is obtained from the multi-source load database. Based on the predicted load data at time i and the average of the predicted load data at all times within the sliding time window, the statistical sliding time window is calculated. The variance of the predicted load data within the specified length is used to obtain the prediction volatility, which characterizes the stability level of the model output. The difference between the predicted load data at time i and the actual load data at time i is calculated. The difference is divided by the actual load data at time i, and the absolute value is taken to obtain the relative error. The sum of the relative errors over all times is then divided by the total time length to obtain the average relative error term, which reflects the overall prediction accuracy. The prediction volatility is multiplied by the jump penalty weight factor to obtain the jump penalty term, which reflects the adverse evaluation of models with frequent jumps. The health assessment value is obtained by subtracting the average relative error term and the jump penalty term from a constant, which serves as an evaluation index for the overall prediction capability of the current model. The result will be used as the triggering basis for model retention, fine-tuning, or retraining strategies.

[0059] The specific formula for the health assessment value is as follows:

[0060] ;

[0061] In the formula, S h The value represents the health assessment score; N represents the length of the statistical time window. P represents the predicted load data at time i; i This represents the actual load data at time i. To predict volatility, i.e. the variance of the predicted load data within the sliding time window; δ is the jump penalty weighting factor. Based on the predicted load data and the actual load data, two objective functions are defined: absolute error and oscillation degree. Multiple datasets are constructed, and the optimal solution set composed of several candidate points is output using a multi-objective optimization method. The data with the smallest error and the smallest oscillation is selected, with a range between 0.1 and 0.2.

[0062] In this implementation plan, a health assessment value is constructed through periodic error analysis and volatility assessment to comprehensively reflect the model's prediction accuracy and stability. Combined with threshold judgment, performance degradation is identified and retraining is triggered, ensuring the model's continuous optimization and adaptive capabilities during long-term operation.

[0063] Specifically, the process of optimizing the neural network time series prediction model and managing its version and feedback learning is as follows: Real-time comparison of the health assessment value with the health assessment threshold. When the health assessment value is less than the health assessment threshold, it indicates that the prediction has a large error and output oscillation problem, triggering an adjustment mechanism, including freezing abnormal prediction segments and reducing the control response frequency, to prevent further impact of misprediction on actual regulation. Simultaneously, the current neural network time series prediction model is marked as needing fine-tuning and enters the training re-evaluation process to ensure rapid correction of the model in an unstable state. When the health assessment value is greater than or equal to the health assessment threshold, it indicates that the prediction is accurate and the fluctuation is stable. The current neural network time series prediction model continues to be used, its performance indicators are recorded as reference standards, and it is allowed to participate in core scheduling tasks to ensure the reliability and performance inheritance of the model's continuous operation. Each update records the version number, indicator changes, and training data hash value, supporting rollback and comparative experiments to achieve traceability, comparability, and anomaly rollback capabilities during model iteration, constructing a comprehensive version lifecycle management system and online feedback optimization mechanism.

[0064] In this implementation plan, a health assessment is used to monitor the model status in real time, triggering fine-tuning and retraining to ensure prediction reliability. A supporting version management and feedback learning mechanism is also provided to support performance tracking, anomaly rollback and comparative optimization, thus building a sustainable model management system.

[0065] Reference Figure 2As shown, the second aspect of this invention provides a power grid load forecasting system based on a large model, applied to the aforementioned power grid load forecasting method based on a large model. The system includes a multi-source sensing and data preprocessing module, a disturbance identification and trend detection module, a physical fusion large model prediction engine module, a steady-state fusion and feedback control module, and a model optimization and evolution module. The multi-source sensing and data preprocessing module is used to collect multi-source load comprehensive data in real time, aggregate the multi-source load comprehensive data for edge processing, and perform data preprocessing. The disturbance identification and trend detection module is used to monitor abrupt change points in real time, construct behavioral label sequences to extract load change patterns, identify high-frequency disturbance points, and implement optimization measures based on the identification results. The large-scale model prediction engine module, which integrates physical constraints, constructs the prediction input vector of multi-source load data, performs load data prediction, introduces physical constraint terms, and calculates physical constraint regularization values ​​to determine whether the prediction results of the neural network time series prediction model violate physical boundary conditions and whether there are drastic jumps. The steady-state fusion and feedback control module is used to fuse prediction results, perform dynamic confidence interval evaluation based on a sliding time window, make real-time judgments on prediction results according to the confidence boundary value range, and provide feedback control instructions. The model optimization and evolution module is used to perform error analysis and retraining triggering, comprehensively evaluate prediction accuracy and output stability, and optimize the neural network time series prediction model, manage versions, and perform feedback learning.

[0066] In this implementation scheme, high-quality input data is constructed through multi-source sensing and edge preprocessing. The model's adaptability to sudden changes and boundaries is enhanced by combining disturbance identification and physical constraints. Real-time judgment and control response of prediction results are realized based on the confidence interval. The model is continuously optimized and evolved through error evaluation and version management, thereby comprehensively improving the accuracy, stability and adaptability of power grid load forecasting.

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0068] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A large model-based power grid load forecasting method, characterized in that, The method comprises the following steps: S1, collecting multi-source load comprehensive data in real time, collecting multi-source load comprehensive data for edge processing, and performing data preprocessing; S2, real-time monitoring of mutation points, construction of behavior label sequence, extraction of load change mode, identification of high-frequency disturbance points, and implementation of optimization measures according to the identification results; The specific process of real-time monitoring of mutation points and construction of behavior label sequence to extract the load change mode and identify high-frequency disturbance points is as follows: The volatility of the load data is calculated through a sliding time window, the trend of the short-time load data is monitored, the Bayesian mutation detection model is accessed to identify the probabilistic structural mutation in the load data fluctuation curve in real time, and the abnormal judgment threshold is adaptively fine-tuned in combination with the current time period and the regional load density; The load data fluctuation curve of each day is converted into a symbolic mode label according to the time period characteristics, and the labels are expressed in behavior segments. The label sequences are clustered and analyzed by an unsupervised clustering algorithm to identify potential periodic structures or abnormal mode changes. Real-time load data is obtained to obtain the load data at the current time and the load data at the previous time. The average load data is calculated based on the load data at each time collected within the sliding time window length. Real-time temperature data is obtained to obtain the temperature data at the current time and the temperature data at the previous time, and the difference between the temperature data at the two times is obtained. The distribution of the temperature data change amount in one year in the multi-source load comprehensive database is counted, and a temperature disturbance reference scale is determined according to the fact that more than 95% of the temperature data fluctuations do not exceed a certain value. The absolute change amount of the load data is obtained by subtracting the load data at the previous time from the load data at the current time and taking the absolute value. The load fluctuation rate is obtained by dividing the absolute change amount of the load data by the average load data. A temperature influence term is obtained by adding a constant one to the ratio of the temperature difference to the temperature disturbance reference scale. The high-frequency disturbance point discrimination value is obtained by multiplying the load fluctuation rate and the temperature influence term. The specific process of implementing optimization measures according to the identification results is as follows: The high-frequency disturbance point discrimination value is compared with the first-level and second-level disturbance threshold values in real time. When the high-frequency disturbance point discrimination value is less than or equal to the disturbance threshold value, it is considered that the current load fluctuation is stable, and the current neural network time series prediction model is continued to be used for load prediction. The prediction cache and the lightweight inference model are appropriately enabled, the edge computing node is operated at a reduced frequency, and the current stable section is marked as a normal sample. When the high-frequency disturbance point discrimination value is greater than the first-level disturbance threshold value and less than or equal to the second-level disturbance threshold value, it indicates that the load change exceeds 20% of the average value, the prediction period is shortened, the instantaneous response is strengthened, the edge computing node records the abnormal period of the high-frequency disturbance point discrimination value in real time, and marks it as a disturbance section. The platform triggers a prediction accuracy warning state, and displays the region that needs attention. When the high-frequency disturbance point discriminant value is greater than the secondary disturbance threshold value, it indicates that the load has a sharp jump and is accompanied by a dramatic change in the environment, and the energy storage system is pre-scheduled. The priority is executed according to the demand of the equipment, the power consumption power limit signal is issued through the programmable logic controller, and if the abnormal fluctuation is not accessible to the cloud platform, the protection strategy is started on the edge side: entering the static prediction and standby value output, and comparing the disturbance segment with the city event and industrial production data; S3, constructing a prediction input vector of multi-source load comprehensive data, performing load data prediction, introducing a physical constraint term, and calculating a physical constraint regularization value to determine whether the prediction result of the neural network time series prediction model violates the physical boundary condition and whether there is a sharp jump; S4, fusion prediction result, dynamic credible interval evaluation based on sliding time window, real-time judgment of prediction result according to credible boundary value interval, and feedback control instruction; S5, error analysis and retraining triggering, comprehensive evaluation of prediction accuracy and output stability, neural network time series prediction model optimization, version management and feedback learning.

2. The method of claim 1, wherein, The specific process of real-time collection of multi-source load comprehensive data, aggregation of multi-source load comprehensive data for edge processing, and data preprocessing is as follows: Real-time collection of multi-source load comprehensive data, including: deploying intelligent electric meters at load-intensive points to collect load data; deploying regional micro-meteorological sensors to collect environmental data including temperature and humidity data; accessing programmable logic controllers and manufacturing execution systems to collect industrial load behavior data including device start-stop and industrial device status; accessing city event open platform interfaces to obtain high-density activity information related to regional holidays and promotion information; accessing factory and park data collection and monitoring systems to collect device operating status, energy consumption data, and transformer capacity; Implementing timestamp alignment, unified to minute-level granularity; processing event coding of high-density activity unstructured data; using edge computing gateway to process preliminary multi-source load comprehensive data compression, filtering and caching; using linear interpolation to fill in missing values, using standard score method to remove outliers, and standardizing and normalizing multi-source load comprehensive data; uploading multi-source load comprehensive data to multi-source load comprehensive database.

3. The method of claim 1, wherein, The specific process of constructing a prediction input vector of multi-source load comprehensive data and performing load data prediction is as follows: Constructing a load data feature vector based on multi-source load comprehensive data as a training set of the neural network time series prediction model, extracting, encoding and aligning the key features of the historical load data sequence, industrial device operating status, environmental data and high-density activity information coding multiple key dimensions of data, respectively converting into corresponding load data feature vectors, and splicing according to time window sliding, integrating to form a load data feature vector; the load data feature vector is organized in time dimension according to sliding window structure to form a unified three-dimensional tensor, a neural network time series prediction model with a core structure of converter model is constructed through the load data feature vector, and the prediction error loss function is minimized for back propagation training, and the predicted load data at future time is output.

4. The method of claim 1, wherein, The specific process of introducing the physical constraint term and calculating the physical constraint regularization value to determine whether the prediction result of the neural network time series prediction model violates the physical boundary condition and whether there is a sharp jump is as follows: The transformer capacity, load data change rate and safety margin are taken as hard boundary conditions, the physical constraint regularization value is added in the training loss function, the predicted load data at the current time is obtained, the transformer capacity is taken as the maximum allowed load of the region, the difference between the predicted load data at the current time and the maximum allowed load of the region is calculated, the relative proportion of the predicted value exceeding the maximum allowed load of the region is obtained by dividing the difference by the maximum allowed load of the region, and the relative proportion is squared to obtain a capacity penalty term. The capacity penalty term is subjected to maximum function operation, that is, if the predicted load data at the current time is less than or equal to the maximum allowed load of the region, the capacity penalty term is zero, and if the predicted load data at the current time is greater than the maximum allowed load of the region, the capacity penalty term is the square value of the relative proportion; Based on the sliding time window, the predicted load data at the current time is differentiated with respect to time, and the derivative result is squared and multiplied by a change rate weight factor to obtain a jump penalty term. The capacity penalty term after maximum function operation is added to the jump penalty term to obtain the physical constraint regularization value. The physical constraint regularization value is compared with the low risk threshold and the high risk threshold in real time. When the physical constraint regularization value is less than or equal to the low risk threshold, it indicates that the neural network time series prediction model is stable and has not touched the capacity boundary, and is in a safe state, so the current neural network time series prediction model is maintained. When the physical constraint regularization value is greater than the low risk threshold and less than or equal to the high risk threshold, it indicates that there is a moderate risk, the predicted value is close to the transformer capacity and a slight jump occurs, the peak shaving warning is started, and the flexible load is linked to enter smooth adjustment. When the physical constraint regularization value is greater than the high risk threshold, it is determined that there is a serious risk, the predicted value has been seriously out of limit and has a sharp shock, the energy storage discharge and load reduction instructions are triggered, the current prediction result output is frozen, the sliding average is called to replace, and the risk log and alarm information are uploaded to the cloud scheduling platform at the same time.

5. The method of claim 1, wherein, The specific process of fusing the prediction result, dynamically evaluating the credible interval based on the sliding time window, and determining the prediction result in real time according to the credible boundary value interval and feeding back the control instruction is as follows: Based on the predicted load data, the sliding time window is combined, the median filter is applied to process the shock predicted load data, the credible boundary value interval of the prediction result is calculated: the predicted load data at the current time is obtained; Through the multi-source load comprehensive database and based on the sliding time window, the load data at the last n times is obtained, and the average value of the load data at the last n times is calculated. The load data variance at the last n times is obtained by calculation; The confidence factor constant term is obtained through the standard normal distribution table; The high frequency disturbance point discrimination value at the current time is calculated; The load fluctuation term is obtained by multiplying the variance of the load data of the last n time points with the self fluctuation weight factor, the disturbance discrimination term is obtained by multiplying the high frequency disturbance point discrimination value of the current time point with the control disturbance weight factor, the confidence interval width adjustment term is obtained by summing the load fluctuation term and the disturbance discrimination term and multiplying the sum with the confidence factor constant term, the upper limit of the predicted confidence boundary value is obtained by adding the confidence interval width adjustment term to the predicted load data of the current time point, the lower limit of the predicted confidence boundary value is obtained by subtracting the confidence interval width adjustment term from the predicted load data of the current time point, and the predicted confidence boundary value interval is obtained by combining the upper limit and the lower limit of the predicted confidence boundary value. Based on the predicted confidence boundary value interval, when the collected actual load data is not in the predicted confidence boundary value interval, it is determined that the current neural network time series prediction model is inaccurate, an abnormal event reminder is pushed, and a backup historical mean prediction model is started; and when a plurality of consecutive actual load data falls out of the predicted confidence boundary value interval, the neural network time series prediction model is replaced, and a disturbance segment sample is added for reinforcement learning; when the upper limit of the predicted confidence boundary value is greater than or equal to the transformer capacity, the energy storage device is dispatched, part of the non-key load is turned off, and the energy-saving operation mode is entered.

6. The method of claim 1, wherein, The specific process of the error analysis and retraining triggering, comprehensive evaluation of prediction accuracy and output stability is as follows: Statistical analysis is performed on the residual between the predicted value and the actual load data every month, key error indicators are evaluated, and the performance threshold is compared; If the error exceeds the limit for two consecutive months, it is automatically marked as performance degradation and is included in the retraining candidate queue; The predicted load data of the i-th time point is obtained, the actual load data of the i-th time point is obtained from the multi-source load comprehensive database, the prediction fluctuation degree is calculated based on the predicted load data of the i-th time point and the average value of the predicted load data of all time points within the sliding time window length, the difference between the predicted load data of the i-th time point and the actual load data of the i-th time point is calculated, the relative error is obtained by dividing the difference by the actual load data of the i-th time point and taking the absolute value, and the average relative error term is obtained by summing the relative errors of all time points and dividing by the total time length; the jump penalty term is obtained by multiplying the prediction fluctuation degree by the jump penalty weight factor; the health degree evaluation value is obtained by subtracting the average relative error term and the jump penalty term from the constant.

7. The method of claim 1, wherein, The specific process of the neural network time series prediction model optimization and version management and feedback learning is as follows: The health degree evaluation value is compared with the health evaluation threshold in real time, when the health degree evaluation value is less than the health evaluation threshold, it indicates that the prediction has large error and output oscillation problem, the adjustment mechanism is triggered, including freezing the abnormal prediction segment, reducing the control response frequency, and at the same time marking the current neural network time series prediction model as needing fine-tuning state and entering the training reevaluation process; When the health degree evaluation value is greater than or equal to the health evaluation threshold, it indicates that the prediction is accurate and the fluctuation is stable, the current neural network time series prediction model is continued to be used, the performance indicators are recorded as reference standards, and the current neural network time series prediction model is allowed to participate in the core scheduling task. Each update record version number, index changes and training data hash value, support rollback and contrast experiment.

8. A large model-based power grid load forecasting system, applying the large model-based power grid load forecasting method according to any one of claims 1-7, characterized in that, Include: Multi-source perception and data preprocessing module, disturbance identification and trend detection module, physical fusion large model prediction engine module, steady-state fusion and feedback control module and model optimization and evolution module: Among them, the multi-source perception and data preprocessing module is used for real-time collection of multi-source load comprehensive data, aggregation of multi-source load comprehensive data for edge processing, and data preprocessing; The disturbance identification and trend detection module is used for real-time monitoring of mutation points, construction of behavior label sequence, extraction of load change mode, identification of high-frequency disturbance points, and implementation of optimization measures according to the identification results; The physical fusion large model prediction engine module is used for constructing a prediction input vector of multi-source load comprehensive data, predicting load data, introducing a physical constraint term, and calculating a physical constraint regularization value to determine whether the prediction result of the neural network time series prediction model violates the physical boundary condition and whether there is a sharp jump; The steady-state fusion and feedback control module is used for fusing the prediction results, performing dynamic credible interval evaluation based on a sliding time window, judging the prediction results in real time according to the credible boundary value interval, and feeding back control instructions; The model optimization and evolution module is used for error analysis and retraining triggering, comprehensive evaluation of prediction accuracy and output stability, and neural network time series prediction model optimization and version management and feedback learning.

Images