Power load prediction processing method and device and electronic equipment

Abstract

The invention discloses a power load prediction processing method and device and electronic equipment. The method comprises the following steps: determining a credible anchor point group based on an extreme dynamic fragment of a target area in a target time period; under the constraint of the credible anchor point group, disturbance processing is carried out on the feature data corresponding to the adjacent sub-regions of the target region, a distortion control boundary is determined, and the distortion control boundary is obtained based on a distortion propagation path and an amplification critical point of distortion features among the multiple sub-regions; generating region calibration priori data based on the distortion control boundary; and performing correction processing on the power load prediction results corresponding to the plurality of sub-regions based on the region calibration prior data to obtain corrected power load prediction results corresponding to the plurality of sub-regions. According to the invention, the technical problem of distortion characteristic propagation and amplification during multi-region distributed load prediction in a power system in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart grid, in particular, to a power load prediction processing method and device and electronic equipment. BACKGROUND

[0002] In modern power systems, with the continuous expansion of the power grid scale and the increasing complexity of power load prediction demand, multi-region distributed load prediction has become one of the key technologies to ensure the efficient and stable operation of the power system. However, when the related technology implements this function, it often faces a core problem: the propagation and amplification of distorted features. Since the collection of power load and meteorological information involves a wide range of geographical areas and complex network environments, distorted data, i.e. those feature data with abnormal change amplitude and deviating from the normal trend, may originate from equipment failure, measurement error or data transmission interruption and other factors. Once they appear in the input data of a certain sub-region, they can quickly spread to the entire network through the existing information sharing mechanism. Especially under the distributed prediction architecture, each sub-region shares low-dimensional features to reduce data transmission burden, but this also leads to a cascading amplification effect of distorted features, i.e. distorted feature data is continuously amplified in cross-regional information interaction, ultimately seriously affecting the global prediction accuracy. In addition, the global optimization strategy of the central coordination node is often based on the feature input provided by each sub-region, so the existence of distorted features will cause the deviation of the global optimization direction, and thus affect the decision accuracy of power dispatching.

[0003] At present, no effective solution has been proposed to solve the above problems. SUMMARY

[0004] The embodiments of the present application provide a power load prediction processing method, device and electronic equipment, to at least solve the technical problem of propagation and amplification of distorted features in the related art when multi-region distributed load prediction is performed in a power system.

[0005] According to an aspect of the embodiments of the present application, there is provided a power load prediction processing method, comprising: determining a group of trusted anchor points based on extreme dynamic segments of a target region in a target period, wherein the target region is a region with distortion characteristics in a plurality of sub-regions, the extreme dynamic segments are characteristic data with a variation amplitude greater than a preset amplitude, the characteristic data comprises power load data and meteorological data, and the group of trusted anchor points is used to indicate characteristic reference points under a normal load variation trend; performing perturbation processing on characteristic data corresponding to adjacent sub-regions of the target region under the constraint of the group of trusted anchor points to determine a distortion control boundary, wherein the distortion control boundary is obtained based on a distortion propagation path and an amplification critical point of the distortion characteristics among the plurality of sub-regions; generating region calibration prior data based on the distortion control boundary, wherein the region calibration prior data is used to indicate a power load prediction trend after elimination of the distortion characteristics and an adjustment standard of characteristic weights of each region, and a power load prediction offset compensation amount under a specific condition; and performing correction processing on power load prediction results corresponding to the plurality of sub-regions respectively based on the region calibration prior data to obtain corrected power load prediction results corresponding to the plurality of sub-regions respectively.

[0006] According to another aspect of the embodiments of the present application, there is also provided a power load prediction processing device, comprising: an anchor point group determination module configured to determine a group of trusted anchor points based on extreme dynamic segments of a target region in a target period, wherein the target region is a region with distortion characteristics in a plurality of sub-regions, the extreme dynamic segments are characteristic data with a variation amplitude greater than a preset amplitude, the characteristic data comprises power load data and meteorological data, and the group of trusted anchor points is used to indicate characteristic reference points under a normal load variation trend; a boundary determination module configured to perform perturbation processing on characteristic data corresponding to adjacent sub-regions of the target region under the constraint of the group of trusted anchor points to determine a distortion control boundary, wherein the distortion control boundary is obtained based on a distortion propagation path and an amplification critical point of the distortion characteristics among the plurality of sub-regions; a prior data generation module configured to generate region calibration prior data based on the distortion control boundary, wherein the region calibration prior data is used to indicate a power load prediction trend after elimination of the distortion characteristics and an adjustment standard of characteristic weights of each region, and a power load prediction offset compensation amount under a specific condition; and a load prediction correction module configured to perform correction processing on power load prediction results corresponding to the plurality of sub-regions respectively based on the region calibration prior data to obtain corrected power load prediction results corresponding to the plurality of sub-regions respectively.

[0007] According to another aspect of the embodiments of the present application, there is also provided a non-volatile storage medium, which stores a plurality of instructions, and the instructions are adapted to be loaded and executed by a processor to perform any one of the power load prediction processing methods.

[0008] According to another aspect of the embodiments of the present application, an electronic device is provided, comprising one or more processors and a memory storing one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement any of the power load prediction processing methods.

[0009] According to another aspect of the embodiments of the present application, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of any of the power load prediction processing methods.

[0010] In the embodiments of the present application, the trusted anchor point group is determined based on the extreme dynamic segment of the target region in the target period, wherein the target region is a region with distortion features in a plurality of sub-regions, the extreme dynamic segment is feature data with a change amplitude greater than a preset amplitude, the feature data includes power load data and meteorological data, and the trusted anchor point group is used to indicate a feature reference point under a normal load change trend; under the constraint of the trusted anchor point group, the feature data corresponding to the adjacent sub-regions of the target region is subjected to disturbance processing to determine a distortion control boundary, wherein the distortion control boundary is obtained based on a distortion propagation path and an amplification critical point of the distortion features among the plurality of sub-regions; based on the distortion control boundary, region calibration prior data is generated, wherein the region calibration prior data is used to indicate the power load prediction trend of each region after the distortion features are eliminated, the adjustment standard of the feature weight, and the power load prediction offset compensation amount under a specific condition; based on the region calibration prior data, the power load prediction results corresponding to the plurality of sub-regions are corrected to obtain the corrected power load prediction results corresponding to the plurality of sub-regions, achieving the purpose of using the trusted anchor point group to guide abnormal detection and control, preventing the propagation of distortion signals by constructing a distortion control boundary, and accurately calibrating the power load prediction results, thereby realizing the technical effect of improving the accuracy of the power load prediction results and effectively suppressing the cascading effect of distortion signals, and further solving the technical problems of distortion feature propagation and amplification in the related art when predicting the distributed load in the power system. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0012] Figure 1 is a flowchart of a power load prediction processing method according to an embodiment of the present application;

[0013] Figure 2 is a flowchart of an optional power load prediction processing method according to an embodiment of the present application;

[0014] Figure 3 This is a schematic diagram of an optional power load forecasting and processing system according to an embodiment of the present invention;

[0015] Figure 4 This is a schematic diagram of an electrical load forecasting processing device according to an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] According to an embodiment of the present invention, a method embodiment for power load forecasting is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0019] Figure 1 This is a flowchart of a power load forecasting processing method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0020] Step S102: Based on the extreme dynamic segments of the target area within the target time period, determine a group of reliable anchor points. The target area is the area with distortion characteristics in multiple sub-regions. The extreme dynamic segments are feature data with a change amplitude greater than a preset amplitude. The feature data includes power load data and meteorological data. The group of reliable anchor points is used to indicate the feature benchmark points under the normal load change trend.

[0021] Optionally, the method in this embodiment is applied to a multi-regional distributed load forecasting scenario. The target region refers to any region among multiple geographically subdivided sub-regions of the power system, which contains anomalous or distorted characteristic data (i.e., distorted features) that may affect the accuracy of power load forecasting. The target time period refers to a specific time window, which may be a period that needs to be focused on in power load forecasting. Extreme dynamic segments refer to data segments within the target time period whose changes exceed a preset threshold. These segments may originate from abnormal fluctuations in power load or related meteorological data. The preset amplitude is a threshold used to distinguish between normal and abnormal fluctuations. The trusted anchor point group represents a set of data points selected from the extreme dynamic segments. They exhibit stable characteristics under normal load change trends and are used as benchmarks for identifying and correcting anomalous data. These anchor points have high data quality and can be used to verify and adjust the credibility of other data points.

[0022] In an optional embodiment, before determining a group of reliable anchor points based on extreme dynamic segments of the target region within a target time period, the method further includes: constructing a feature quality spectrum based on feature data of the corresponding regions of multiple sub-regions, wherein the feature quality spectrum is used to characterize the anomalous fluctuation trend in the feature data and the degree of coordination between load data and meteorological data; determining noise fingerprint data and drift slope corresponding to each of the multiple sub-regions based on the feature quality spectrum, wherein the noise fingerprint data is used to indicate the statistical characteristics of predetermined anomalous patterns in the feature data; the drift slope is used to indicate the stability of the feature data over time and to identify regions where the feature data continuously shifts; and mapping the noise fingerprint data and drift slope corresponding to each of the multiple sub-regions to a cross-regional topological time series. The diagram shows the mapped topology time series, along with the cumulative anomaly scores and collaborative relationship identifiers for each of the multiple sub-regions. The cross-regional topology time series is constructed based on the power grid physical topology and geographical adjacency relationships between the multiple sub-regions, representing the dynamic coupling and information exchange paths between them. The collaborative relationship identifiers indicate the degree of synchronization and spatial correlation of the characteristic data of each region over time. Multiple candidate regions are identified in the mapped topology time series. Based on whether there are anomaly propagation paths between the candidate regions and their corresponding neighboring regions, a target region is determined from the candidate regions. Candidate regions are those sub-regions where the cumulative anomaly score is greater than a preset score threshold, and neighboring regions are those within a predetermined neighborhood of the corresponding candidate region.

[0023] Optionally, a comprehensive spectral matrix can be constructed by calculating various statistical characteristics, such as moving average, variance, and periodic recurrence rate, to comprehensively assess the quality and stability of the data. Characteristic quality spectra not only reflect the fluctuation trend of the data but also reveal the correlation between load data and meteorological data, providing a detailed and multi-dimensional perspective for subsequent analysis. Noise fingerprint data identifies anomalies or noise patterns in the data. Through cluster analysis and pattern matching, it identifies characteristic shapes that significantly deviate from the normal distribution, which helps identify occasional or persistent noise sources in the data. Drift slope quantifies the trend of data shifting gradually over time. It is used to identify areas where the data changes slowly or is continuously distorted over time by calculating the rate of change of the characteristic mean over time. The combined effect of noise fingerprint data and drift slope can more accurately locate anomalies and distortions in the data. The cross-regional topology time series diagram is a complex information network constructed based on the actual physical layout of the power grid and the geographical adjacency relationships between sub-regions. Each node in the diagram represents a sub-region, and the connections between nodes represent the dynamic coupling of loads and information propagation paths between regions. By mapping noise fingerprint data and drift slope onto this graph, the distribution and propagation patterns of anomalous data can be visually observed within the network structure. Each node also includes an anomaly cumulative score, which combines the anomaly severity of the noise fingerprint and drift slope, helping to identify areas more prone to data distortion. Candidate regions are selected based on their anomaly cumulative scores exceeding a preset threshold (e.g., regions with significantly higher anomaly cumulative scores than the average), indicating a higher degree of data anomaly. Subsequently, by analyzing the connections and information exchange paths between candidate regions and their surrounding neighboring regions in the graph, it is determined whether there are effective propagation channels for anomalous signals. Finally, regions that are not only severely anomalous themselves but also capable of influencing surrounding areas through anomalous propagation paths are selected as target regions. Through this meticulous anomaly detection and topology analysis, distorted portions of the data can be more accurately identified and isolated, laying a solid foundation for subsequent power load forecasting processing.

[0024] As an optional implementation, before constructing the characteristic quality spectrum, the characteristic data of each corresponding region in multiple sub-regions are aligned with the time baseline. That is, a cross-regional topological time series map is constructed under a unified event time baseline. By processing the load data and meteorological data, the characteristic quality spectrum is extracted, and noise fingerprint data and drift slope are calculated to locate potential distortion source regions at the system level as target regions, establishing an analytical benchmark for distortion tracing. The specific implementation process is as follows:

[0025] A unified event time baseline is constructed to time-align power load and meteorological data from different geographical regions (i.e., multiple sub-regions). First, edge nodes with high time acquisition accuracy and stable network synchronization performance are selected from all sub-regions as reference nodes. The actual time deviation between each data acquisition terminal and the reference node is measured to construct a time offset matrix between multiple regions. Next, a weighted time fusion method is used, assigning weights based on the time stability index of each node during historical sampling, forming a statistically stable logical time axis as a unified reference time line for the entire network. After the unified time baseline is constructed, load and meteorological data from all regions are re-interpolated onto this time baseline to resolve asynchrony issues caused by geographical differences, grid communication delays, or differences in sampling accuracy. This global time alignment mechanism ensures consistent temporal semantics for cross-regional data in subsequent analysis, guaranteeing the comparability and integrability of spatial characteristics and temporal dynamics.

[0026] After data time alignment, the aligned power load and meteorological data are structured to extract characteristic indicators that can be used to assess data stability and variability. In this process, firstly, the load data for each sub-region is continuously checked on an hourly basis to identify sampling discontinuities, anomalous jumps, and numerical abrupt changes, and missing values ​​are reconstructed using a sliding window. Then, the hourly load data is combined one by one with the meteorological parameter data for that time period, including eight indicators: real-time temperature, average temperature, relative humidity, wind speed, wind direction, sunshine duration, rainfall, and atmospheric pressure. For this multi-source joint characteristic data sequence, seven statistical characteristics are calculated: moving average, variance, periodicity, maximum fluctuation range, frequency of local extrema, sample skewness, and rate of change of volatility, constructing a complete characteristic quality spectrum in matrix form. This characteristic quality spectrum reflects the stability and synergy of load and meteorological characteristics in each region across different time periods, thereby identifying potential abnormal fluctuation trends. Compared to related techniques that only analyze load data waveforms, this method enhances sensitivity to interference from complex factors through multivariate fusion extraction.

[0027] Based on the characteristic quality spectrum generated in the previous stage, noise fingerprint models and drift slope models are constructed for each region to identify regions and time periods with potential distortion characteristics. The noise fingerprint model is built on large-scale historical data. Low-frequency anomaly pattern fragments are extracted for each feature dimension, and typical distortion patterns are identified through clustering, including short-term spikes and drops, periodic unstable fluctuations, and random pulse insertions. These patterns are then standardized to form a reference fingerprint database. The drift slope model identifies regions where data continuously shifts over time by calculating the rate of change of the feature mean and the rate of increase in distance from the central trend over a continuous time window, paying particular attention to time periods where stability declines simultaneously across multiple feature dimensions. By matching the current data trajectory in the feature space of each region with the noise fingerprint database and simultaneously evaluating the drift slope index, two independent but cross-validated anomaly scores are generated. This approach effectively avoids misjudgments of single-dimensional features in related technologies, improving the accuracy and stability of multivariate distortion detection.

[0028] The anomaly scores of each sub-region are mapped onto a cross-regional topological time-series graph constructed based on physical structure and geographical relationships. Specifically, with each geographical region as a graph node, the historical load dynamic correlation between adjacent regions is used as edge weights to construct a weighted graph representing the dynamic coupling relationship of power regions. In this graph, each node is weighted with its corresponding feature quality score, fingerprint matching probability, and drift slope, and its cumulative anomaly score is calculated. By traversing the entire graph, cluster analysis is performed on regional nodes with scores significantly higher than the average, and further analysis is conducted to determine whether there are significant anomaly propagation channels between them and their neighboring regions. If a highly correlated path exists and is accompanied by a synchronous anomaly trend in adjacent nodes, the region can be preliminarily located as a potential distortion source area. This distortion source area location result is not only used to identify the geographical location of the initial anomaly, but also serves as a reference basis for subsequent anomaly dynamic playback, anomaly feature extraction, and prediction and correction operations, ensuring that the entire process is based on a unified temporal semantics and spatial structure, and has traceability and analytical consistency.

[0029] In an optional embodiment, before determining a group of credible anchor points based on extreme dynamic segments of the target region within a target time period, the method further includes: acquiring feature data of the target region within a predetermined time period, wherein the predetermined time period is a time period of predetermined duration before and after the occurrence of distortion in the target region; performing frame processing on the feature data of the predetermined time period according to a preset sliding window to obtain a multi-frame data sequence; and determining a target number of data sequences from the multi-frame data sequences as extreme dynamic segments based on the time-varying indices corresponding to each of the multi-frame data sequences, wherein the time-varying indices are used to quantify the dynamic characteristics of each frame of data as it develops over time, and the time-varying indices include the instantaneous rate of change of the load curve, the curvature of the temperature change, the wind speed vector transformation rate, the atmospheric pressure gradient, the amplitude of rainfall fluctuations, and the rate of increase in solar radiation.

[0030] Optionally, feature data can be acquired within a certain time frame before and after the distortion occurs. This feature data includes, but is not limited to, power load data and related meteorological data such as temperature, humidity, wind speed, and rainfall. By presetting the sliding window size (e.g., every 15 minutes or hour), multiple data frames can be generated, each frame representing a segment of feature data within a predetermined time period. This processing method helps refine data features in the time domain, facilitating subsequent dynamic analysis. Time-varying indices are key parameters used to quantify the dynamic characteristics of data. For example, the instantaneous rate of change of the load curve can measure sudden changes in load; temperature curvature and wind speed vector transformation rate can help understand how weather factors affect the load. By calculating the time-varying indices of each data frame, data frames with significantly larger changes than usual can be identified—i.e., extreme dynamic segments. These segments are often direct evidence of data distortion, and their quantity and selection (target number) directly affect the determination of subsequent reliable anchor groups and the accurate location of distortion features.

[0031] Optionally, for the target area identified during the distortion source localization stage, power load data and multiple meteorological variables for the area within a continuous 48-hour period before and after the distortion occurred are extracted to construct a multi-dimensional data sequence with a time index. The selected meteorological variables include hourly collected real-time temperature, relative humidity, wind speed and direction, atmospheric pressure, hourly rainfall, sunshine duration, and upper-air temperature gradient. All the above raw data are uniformly mapped to the unified time baseline constructed in the previous stage, and missing values ​​or data gaps are filled using linear interpolation to ensure that the data at each time point is complete and usable. Based on this, a sliding time window of one hour is set with a time step of five minutes to segment and extract the entire data sequence. Each time period is organized into a frame, forming a high-density frame sequence data structure covering the target time period. For each frame, time-varying indicators such as the instantaneous rate of change of the load curve, the curvature of temperature change, the wind speed vector transformation rate, the atmospheric pressure gradient, the amplitude of rainfall fluctuations, and the rate of sunshine growth are calculated to dynamically calibrate the state of each frame, and the intensity of change is identified using frame-level data tags. Based on the dynamic intensity reflected in the frame-level data tags, the top 20 frame sequences with the highest rate of change were selected as extreme dynamic segments.

[0032] In one optional embodiment, determining a group of trustworthy anchor points based on extreme dynamic segments of the target region within a target time period includes: performing phase verification on the extreme dynamic segments to obtain phase verification results, wherein phase verification is used to verify the degree of synchronous change between load variables and meteorological variables in the time-varying indicators of the data sequence; performing structural verification on the extreme dynamic segments to obtain structural verification results, wherein structural verification is used to verify the degree of clustering and isolation of the data sequence in the low-dimensional feature space mapping; determining multiple candidate trustworthy anchor points from the extreme dynamic segments based on the phase verification results and structural verification results; evaluating the trustworthiness of the multiple candidate trustworthy anchor points to obtain trustworthiness evaluation results; and selecting trustworthy anchor points that meet preset constraints from the multiple candidate trustworthy anchor points based on the trustworthiness evaluation results to obtain a group of trustworthy anchor points.

[0033] Optionally, phase verification assesses the synchronicity and consistency between power load data and meteorological data by analyzing the relative changes in these data points during extreme dynamic segments. In this process, time-varying indices of load and meteorological variables are used to measure the degree of synchronicity in their responses. For example, if at a certain moment, an increase in temperature is accompanied by an increase in electricity demand, this synchronicity will be reflected in the phase verification results, helping to identify characteristic points under normal load change trends. Structure verification maps extreme dynamic segments to a low-dimensional feature space and observes their distribution within that space, paying particular attention to clustering and isolation. In the low-dimensional space, if data points are closely clustered, it indicates similarity in the feature space, while extremely isolated data points may represent anomalous or distorted characteristics. Structure verification results provide an additional perspective, helping to determine whether data points truly reflect typical characteristics of load changes. By combining the results of phase and structure verification, a preliminary set of data points can be selected that exhibit both good synchronicity under load change trends and reasonable clustering in the low-dimensional feature space, making them candidates for reliable anchor points. The credibility assessment further analyzes the stability and reliability of candidate credible anchors. This may include, but is not limited to, checking the frequency of data point recurrence in historical records, their similarity to other stable data points, and whether they co-occur in multiple data sources, to ensure that these anchors truly represent stable and normal load change trends. Finally, based on the credibility assessment results and pre-defined constraints (such as time stability and feature consistency), candidate credible anchors are further screened. Only those data points that pass all tests and meet the constraints are officially selected as credible anchors, forming a credible anchor group. These anchors will serve as benchmarks in subsequent processing to identify and correct distortion features, ensuring the accuracy and stability of predictions. Through this verification and screening process, the high quality of the credible anchor group is ensured, enhancing the ability to identify anomalous features.

[0034] As an optional implementation, the top twenty frame sequences with the highest rate of change (i.e., extreme dynamic segments) are selected, and frame-by-frame replay is performed to reconstruct the entire process of the evolution of features within each frame over time. Specifically, when replaying each frame of the extreme dynamic segment, the load and meteorological data for that hour are traced back point by point along the time axis to reconstruct the original change path of the load response and meteorological variable fluctuations. Using the load change curve as the main axis, the synchronous change trajectories of each meteorological dimension are superimposed to form a multi-dimensional time surface. Phase difference analysis is performed on this time surface between adjacent feature sequences to determine the synchronous response relationship between meteorological variables and load variables at the micro-period. To quantify the degree of synchronous change, a phase-locked loop consistency index is introduced. By measuring the relative phase stability between features of each dimension, it is determined whether there are frequent mismatches or drastic shifts. Particular attention is paid to the local minimum points of the phase-locked loop consistency index during the replay process, i.e., the moments when the meteorological parameter response significantly weakens or moves in the opposite direction to the load. These points, as potential synchronous breakpoints, will be marked as candidate anomalous segments for further screening.

[0035] After identifying synchronization breakpoints, the data structures of these high-mismatch frame segments are rearranged and organized into a multidimensional dynamic sample set according to time series. A low-dimensional feature space mapping is then performed on this set to observe its overall structural distribution. First, the statistical characteristics of each frame in each feature dimension are extracted, including mean, variance, slope, third moment, and frequency domain power density, forming a high-dimensional feature vector set. Then, an embedding method is used to perform nonlinear dimensionality reduction on this set, mapping it to a two-dimensional or three-dimensional space. This mapping process does not involve function fitting or model training; instead, it constructs a locally nested structure based on the distance relationships between all samples. In the low-dimensional space, the relative positional relationships of all frame segments are observed, and feature segments that are isolated in space, separated from the main group, or exhibit abnormal extension trajectories are marked. These segments represent some unexplained structural distortions or feature breaks, corresponding to critical abrupt change regions in the system behavior in the multidimensional space. These low-dimensional embedding points are used as candidate structural anomalies, and synchronization trend analysis is performed again to verify their instability during temporal evolution and to eliminate false anomalies caused by noise or unrelated interference.

[0036] Finally, for high-confidence anomaly segments that have passed structural and phase verification, representative points in the multidimensional feature space are extracted as a group of reliable anchor points by further combining the location data from the aforementioned sliding playback and low-dimensional mapping. Each anchor point must simultaneously meet four constraints: it must have a clear location on the event timeline; it must be the minimum fluctuation extreme point in the phase-locked loop consistency index; it must be located within a high-density cluster structure rather than at the edge of anomalies in the low-dimensional space mapping; and it must exhibit recurring dynamic behavior characteristics in historical prediction periods. Anchor points that pass these four criteria are extracted as the final group of reliable anchor points, which serves as a reference for disturbance response probing and residual measurement in subsequent steps. This group of anchor points possesses temporal stability, structural consistency, and response repeatability, effectively isolating the initial state of the anomaly propagation source.

[0037] Step S104: Under the constraint of the trusted anchor point group, the feature data corresponding to the adjacent sub-regions of the target region are perturbed to determine the distortion control boundary. The distortion control boundary is obtained based on the distortion propagation path and amplification critical point of the distortion features among multiple sub-regions.

[0038] Optionally, a group of trusted anchor points can be used as a reference to impose constraints on the perturbation process, ensuring that it does not affect the stability and representativeness of normal data points. Small variations are introduced into feature data (such as load data and meteorological data) in adjacent sub-regions of the target region to test how these variations affect the prediction results and identify which variations lead to prediction distortion. By identifying the propagation paths of distortion features and the amplification thresholds that may cause significant prediction bias through perturbation processing, a boundary range is defined within which changes in feature data could potentially lead to prediction distortion.

[0039] In an optional embodiment, under the constraint of a trusted anchor group, the feature data corresponding to adjacent sub-regions of the target region are perturbed to determine the distortion control boundary. This includes: injecting a small-amplitude temporal perturbation into the feature data of adjacent sub-regions of the target region under the constraint of the trusted anchor group, wherein the small-amplitude temporal perturbation is applying a perturbation of no more than a preset perturbation amplitude to the original feature data; determining the residual value of the power load prediction results corresponding to the adjacent sub-regions, wherein the residual value represents the difference in the power load prediction results of the adjacent sub-regions before and after the perturbation, and is used to quantify the degree of influence of the perturbation on the accuracy of power load prediction; and based on the cross-regional residual gradient... The residual variation characteristics are determined by the field, where the cross-regional residual gradient field is a multi-dimensional residual feature constructed based on the changes in power load forecast results of adjacent sub-regions. It is used to show the diffusion trend and intensity change of residual values ​​between different sub-regions. The residual variation characteristics are used to indicate the propagation path and amplification law of residual values ​​over time. Based on the residual variation characteristics, the distortion propagation path and amplification critical point of distortion features between multiple sub-regions are identified as distortion control boundaries. The distortion propagation path is used to describe the path of distortion features spreading from the target region to other regions, and the amplification critical point is the threshold of the amount of feature data change that causes the deviation of power load forecast results under specific disturbances.

[0040] Optionally, perturbation injection refers to applying subtle, controllable changes to the feature data of adjacent sub-regions surrounding the target region, based on a defined group of reliable anchor points. Micro-amplitude temporal perturbations mean the changes will not be too large, not exceeding a preset perturbation amplitude. This ensures the perturbation data remains within the acceptable range of normal data, but is sufficient to observe its impact on the prediction results. This perturbation strategy helps analyze how changes in feature data affect prediction results, providing a basis for subsequent distortion control. The residual value reflects the difference between the predicted and actual values ​​before and after the perturbation; it is an important indicator for quantifying the impact of the perturbation on prediction accuracy. By comparing the prediction results before and after the perturbation, the prediction residual for each sub-region can be calculated, thus understanding the extent of the perturbation's impact on the prediction accuracy of each region. These residual values ​​will be used in subsequent analysis to identify the propagation path and scope of distortion features. The cross-regional residual gradient field is a visualization tool that can be characterized through a propagation path map. It shows the distribution and changes of the residual values ​​after perturbation across different sub-regions, including how the residual values ​​spread across regions over time and the changes in intensity. Based on this gradient field, the propagation pattern and laws of the residuals can be determined, i.e., the residual variation characteristics, including how the residuals propagate over time and space, and their amplification or reduction trends during propagation. These residual characteristics can be characterized by a propagation path map. By analyzing the residual variation characteristics, it is possible to identify how distortion characteristics spread from the target area to other areas, i.e., the distortion propagation path. Furthermore, the threshold point at which distortion characteristics begin to significantly affect the prediction results under specific disturbances can be found, i.e., the amplification critical point. These paths and thresholds together constitute the distortion control boundary, used to define which areas and to what extent changes may lead to prediction distortion, thus providing precise guidance for developing strategies to suppress and eliminate distortion characteristics. This approach not only enhances the ability to identify and control distortion characteristics in power load forecasting systems but also provides a method for dynamically monitoring and responding to distortion propagation, which is crucial for improving the accuracy, stability, and robustness of predictions. By precisely defining the distortion control boundary, this method can effectively prevent the cascading amplification effect of distortion characteristics in the power network, thereby avoiding negative impacts on the overall system's prediction capability.

[0041] As an optional implementation, under the constraint of a trusted anchor point group, micro-amplitude temporal perturbations are injected into sub-regions adjacent to the distortion source region. The cross-regional residual gradient field is measured in real time to identify the propagation path and amplification threshold (i.e., propagation critical point) of distortion features across multiple regions, serving as boundary conditions for distortion control (i.e., distortion control boundaries). Specifically, the perturbation injection start time is selected centered on several anchor points with the strongest dynamic structural stability and highest phase consistency score in the trusted anchor point group. Each anchor point contains a precise event time index, corresponding load behavior change segments, meteorological synchronization characteristic sequences, and their cross-response curves. Within the time period corresponding to the selected anchor points, perturbation synthesis operations are performed on the power load data and highly correlated meteorological data. The perturbation design employs a fine-grained variation approach; that is, on the original load curve, following a symmetrical expansion method centered on the perturbation center, the power value is increased or decreased by no more than 1% of its previous 24-hour average at 15-minute intervals, and the meteorological variables most strongly coupled with it, such as temperature, wind speed, and relative humidity, are simultaneously adjusted to cause a phase shift within a range of 5 to 10 minutes at the peak occurrence time. After the perturbation is generated, the data sequence is integrated into a complete input stream and reconnected to the prediction model in the distortion source region. This ensures that the prediction mechanism itself remains unchanged, and only the input driving features are changed to observe its external propagation effects.

[0042] After the perturbation data is formally input, the prediction results of adjacent sub-regions are captured and compared in real time. To ensure temporal consistency, a unified event time baseline is used within each sub-region to re-align the prediction output, ensuring that the prediction result of each perturbation frame corresponds to the injected frame. During the comparative analysis, a baseline prediction curve without perturbation and a prediction output curve with perturbation are constructed separately, and the two are differentially processed to obtain the residual value for each prediction point. The unit of residual is prediction power offset in kilowatts. Point-by-point interpolation is performed on the prediction results of each region within two hours before and after perturbation injection at a minute-level resolution to construct a complete temporal residual sequence. This sequence shows the immediate response of each adjacent region under the influence of perturbation, and records the specific time points, fluctuation directions, and amplitude trends of residual value mutations. With each region as a column and each time point as a row, a residual matrix is ​​formed with the horizontal dimension representing the sub-region and the vertical dimension representing time, which serves as the basic carrier of propagation response characteristics.

[0043] The internal structure of the residual matrix is ​​analyzed to identify the temporal path of perturbation propagation and its coupling strength changes. The initial response time of a region to a perturbation is determined by calculating the time position where the residual in each column first shows a significant change (e.g., more than twice the average of the previous hour). Simultaneously, the evolution rate of the residual value after this time point is tracked, and the period with the maximum slope is extracted as the amplification acceleration segment of that region. Combining the time response delay and amplitude growth trend, the paths of each sub-region are sorted. Using the starting region (i.e., the perturbation injection region) as the source, adjacent regions are connected sequentially according to their response time, and the propagation duration (in minutes) and amplification intensity (maximum residual / perturbation amplitude ratio) are labeled for each connection. This constructs a multi-node, directed, and weighted propagation chain, where each node records its position number, response delay, response intensity, residual waveform change trend, and perturbation propagation direction. To enhance structural stability, the perturbation injection operation is repeated three times, with the perturbation data maintaining a consistent direction but slightly varying amplitude. The link with the highest path structure consistency among the three rounds is selected as the final propagation path map and used as the cross-regional residual gradient field.

[0044] After obtaining the complete propagation path diagram, to establish the boundary conditions for distortion control, it is necessary to identify the amplification threshold of each propagation node. Specifically, for each non-starting node in the propagation path diagram, under multiple rounds of input with slightly varying disturbance amplitudes, the corresponding disturbance input amplitude at the point where the maximum residual occurs is extracted, and the time distance between this amplitude and the input signal is calculated. If a nonlinear surge is observed in the residual response curve when the disturbance amplitude increases slightly, i.e., abruptly changing from stable fluctuation to a short-term steep rise, this disturbance amplitude is determined to be the amplification critical point. The amplification critical disturbance amplitude value for each sub-region is recorded, and its corresponding start and end time points are marked, forming a regional disturbance amplification model. Based on this, a disturbance response sensitivity level and tolerance threshold are set for each region, and the residual change trend is classified into three modes: linear enhancement, hysteresis amplification, and nonlinear burst. The propagation chain, response delay, sensitivity level, critical disturbance value, and amplification trend model between regions are integrated to construct a cross-regional disturbance control boundary diagram. This diagram is used in subsequent steps to determine which areas should enter the proactive feature purification process, whether an isolation mechanism needs to be triggered in advance, and within which prediction period intervention is most effective.

[0045] Step S106: Based on the distortion control boundary, generate regional calibration prior data, wherein the regional calibration prior data is used to indicate the power load forecast trend and feature weight adjustment criteria of each region after eliminating distortion features, as well as the power load forecast offset compensation amount under specific conditions.

[0046] Optionally, based on the distortion control boundary, a dataset (i.e., regional calibration prior data) is generated to guide how each sub-region adjusts its forecast trends and feature weights. These datasets also include estimated power load forecast offset compensation amounts to correct forecasts in special cases, ensuring accurate predictions even under conditions of feature distortion.

[0047] In one optional embodiment, generating regional calibration prior data based on the distortion control boundary includes: constructing an adaptive isolation valve based on the distortion control boundary to actively suppress the propagation of distortion features between sub-regions, wherein the adaptive isolation valve is used to dynamically adjust the corresponding isolation or release strategy according to the real-time detected feature data status; under the control of the adaptive isolation valve, initiating a feature cleanup chain along the distortion propagation path in the distortion control boundary, and using heterogeneous redundancy comparison and trusted summary replacement methods to process the distortion features in the feature data corresponding to each of the multiple sub-regions to obtain regional calibration prior data, wherein heterogeneous redundancy comparison refers to detecting anomalies by comparing feature data from different sources or types, and the trusted summary replacement mechanism is used to replace the distortion features with summary data from a trusted source after detecting distortion features.

[0048] Optionally, the adaptive isolation valve is a dynamic control mechanism designed to address the propagation of distortion features in power load forecasting. Based on the distortion control boundary—the distortion propagation path and amplification critical point—it dynamically adjusts its isolation (preventing data propagation) or release (allowing normal data flow) strategy when anomalies in feature data are detected, effectively suppressing the spread of distortion features across different sub-regions. This valve mechanism intelligently responds to data changes, avoiding the problem of inadvertently isolating valid information due to rigid, static rules. The feature purification chain is a crucial process for processing distortion features. Guided by the adaptive isolation valve, it purifies feature data in each sub-region along the distortion propagation path defined in the distortion control boundary. This process aims to repair or replace data affected by distortion, preventing the spread of anomalous features in the system. In the process of processing distortion features in the feature data corresponding to multiple sub-regions using heterogeneous redundancy comparison and trusted digest replacement methods, heterogeneous redundancy comparison is an effective means of detecting and identifying anomalous features by comparing feature data (redundancy) from different data sources (heterogeneous) or different types of feature data (redundancy). For example, if wind speed data in a certain area suddenly changes abnormally, while data from other areas and historical data at the same time point do not show similar changes, this inconsistency may indicate data distortion. Once distorted features are identified in the characteristic data, the trusted summary replacement mechanism searches for relevant and stable summary data from trusted data sources for replacement. Trusted data sources refer to historical data or data from neighboring areas that have shown stable performance in the past and have high correlation with the current data. The replacement process aims to replace distorted data with validated and reliable characteristic data to correct the input of the prediction model. Regional calibration prior data is generated after the characteristic data has been cleaned. It not only includes the cleaned characteristic data but also information about the data source, cleansing method, and changes before and after cleansing, providing important reference for subsequent global prediction correction. Through this data, the central coordination node can understand the status of characteristic data in each sub-region and make more accurate prediction corrections accordingly.

[0049] Through the above feature data purification process, not only can the propagation of distorted features be actively controlled, but the feature data can also be repaired and improved in a targeted manner, thereby enhancing the accuracy and stability of the overall power load forecasting system. By employing intelligent control of the adaptive isolation valve and dynamic processing of the feature purification chain, the method in this embodiment can effectively address common data distortion problems in power load forecasting, enhancing the self-repair capability and robustness of the power system.

[0050] As a more alternative implementation, under boundary condition constraints, an adaptive isolation valve is constructed, and a feature purification chain is initiated along the propagation path. Heterogeneous redundancy comparison and trusted summary replacement mechanisms are used to actively mitigate the influence of distorted features, generating regional calibration prior data as reference information for subsequent corrections. Specifically, based on the established disturbance propagation path diagram and the amplification threshold of each region, an adaptive identification procedure is initiated for each response region in the path to determine whether it has entered a distortion isolation state. The judgment criterion is that the prediction residual of the region exceeds twice its historical stability threshold within three consecutive prediction cycles, accompanied by a residual slope increasing more than four times within three minutes. When this condition is met, a local feature freezing mechanism is initiated within the current prediction window. The freezing operation includes stopping the self-updating of features in the region, temporarily shielding the region's input contribution to the prediction results of other regions, and locking its current original input data for analysis. Freezing does not affect its reception of correction feedback from the central node, but it prohibits the region from actively generating globally shared features. The freezing process does not involve structural stripping or interfere with the communication channel; it only affects the data usage logic within the current prediction process, ensuring that local anomalies no longer propagate.

[0051] The feature cleansing operation is initiated to structurally clean all predicted input data within the currently frozen area, reconstructing the usable feature space. The operation begins with anomaly feature identification, systematically checking whether the values ​​of all input variables in the area have drifted more than twice the standard deviation of the mean in the most recent three prediction periods, whether the volatility exceeds 30% of its monthly average, and whether there is a significant decrease in the rate of co-change with similar features. Any feature meeting any of these conditions is marked as an anomaly. For marked features, the system automatically extracts their historical values, current values ​​from neighboring areas, and redundant values ​​from multiple measurement channels. For example, regarding wind speed: if the wind speed value in the area suddenly increases, while its corresponding historical wind speed shows no similar change, the wind speed in neighboring geographical areas remains stable, and a sudden deviation between multiple measurement points exceeds the tolerance limit of 3%, then the wind speed data is confirmed to be distorted. Based on this, the three most stable window segments in the historical normal value sequence are extracted, and their mean and trend structure are calculated as historical redundancy references; the current data of two areas in the neighboring area wind speed data that match the climate type of this area are fused using the mean as a spatial redundancy reference; through redundancy comparison, it is determined whether the current feature needs to be replaced.

[0052] When a certain feature dimension is confirmed to be necessary for replacement, the reliable summary replacement stage begins, constructing a highly consistent and structurally complete feature replacement sequence. Replacement data sources are divided into two categories: one is anchor point samples from the anchor point group data that have high structural similarity to the current anomalous segment; the other is fused feature trajectories from multi-view redundancy that have undergone weighted averaging and phase calibration. Anchor points are identified using a structural matching method, mapping the current anomalous feature segment to four types of indicator feature vectors: slope, volatility, periodicity, and amplitude stability. Three anchor point segments with the smallest Euclidean distance are retrieved from the anchor point library, and the best anchor point summary data group is selected based on the criterion of controlling the average response difference to within 5%. If anchor point matching fails, redundant fused data is selected as a fallback, and trend smoothing is performed on it to ensure that the temporal continuity of the replaced value is not disrupted after insertion. All replacement operations must meet the following two constraints: first, the change in the coefficient of coordination between variables does not exceed 0.15; second, the change in the prediction residual before and after replacement does not exceed 50% of the original deviation. After the replacement is complete, the complete prediction input data frame is reassembled, and the input vector is updated for use in subsequent model prediction stages.

[0053] Based on the input sequence after feature purification, a regional calibration prior is constructed. This prior data includes not only the latest purified feature data frame itself, but also historical error weights associated with the data, spatial redundancy structure identifiers, anchor point corresponding scores, replacement impact factors, and time window labels. All data are organized into time-series vector groups, with each feature in the vector group carrying a source label (historical, spatial, anchor point), confidence level score (0-1), intervention record (whether it was replaced, replacement type, replacement ratio), time span, and linkage coefficient with upstream and downstream regions. This calibration prior serves as a reference for correcting regional prediction results and will be invoked by the central coordination node during the global correction phase.

[0054] Step S108: Based on the regional calibration prior data, the power load prediction results corresponding to each of the multiple sub-regions are corrected to obtain the corrected power load prediction results corresponding to each of the multiple sub-regions.

[0055] Optionally, correction processing is a remedial process used to adjust the power load forecast results for sub-regions using information from prior data of regional calibration, eliminating or mitigating the effects of distortion characteristics. The corrected forecast results are closer to actual power load changes, exhibiting higher accuracy and reliability, and contributing to the effective scheduling and management of the power system.

[0056] In one optional embodiment, based on regional calibration prior data, the power load forecast results corresponding to each of the multiple sub-regions are corrected to obtain corrected power load forecast results for each of the multiple sub-regions. This includes: constructing a regional interaction graph based on the regional calibration prior data and the power grid physical topology relationship between the multiple sub-regions, wherein each sub-region is used as a graph node to indicate the dynamic coupling relationship and information flow path between the sub-regions, and a prediction error weight evaluated based on historical data, wherein the prediction error weight is used to reflect the reliability and historical volatility of the load forecast results of each sub-region; and based on the regional interaction graph, the coupling is consistent. The coupling consistency regularization term and the power flow feasible region constraint are used to globally correct the power load forecast results corresponding to each of the multiple sub-regions, resulting in corrected power load forecast results for each of the multiple sub-regions. Among them, the coupling consistency regularization term is used to control the deviation of the forecast output of highly correlated sub-regions so as not to exceed the upper limit of the historical fluctuation range. Highly correlated sub-regions refer to sub-regions with a correlation greater than a preset correlation threshold. The correlation can be obtained based on the load level correlation between regions, geographical location relationship, etc. For example, regions with high load level correlation and similar geographical locations are regarded as highly correlated sub-regions. The power flow feasible region constraint is used to control the load operation of multiple sub-regions to conform to the power grid operation safety boundary.

[0057] Optionally, the regional interaction graph is a graphical model constructed by combining the characteristic data status of each sub-region with the physical connectivity of the power grid, reflecting the dynamic coupling relationships and information flow paths between regions. In the graph, each sub-region is represented as a node, and the connections between nodes represent the mutual influence of power grid physical connectivity or load forecast results. This model specifically considers the prediction error weights based on historical data evaluation to reflect the reliability and historical volatility of the prediction results for each sub-region, providing a structured basis for subsequent optimization and correction. Based on the regional interaction graph, graph consensus optimization, combined with coupling consistency regularization terms and power flow feasible region constraints, is used to uniformly correct the prediction results of all sub-regions. The coupling consistency regularization term ensures that highly correlated sub-regions maintain consistency in their prediction results; that is, their prediction output deviations cannot exceed the upper limit of the historical fluctuation range, which helps improve the coordination and reliability of the prediction results. The power flow feasible region constraint ensures that the load forecast results of all sub-regions do not violate the safety boundaries of power grid operation; that is, the prediction results must be feasible considering the physical characteristics and operational limitations of the power grid, preventing actual operational risks caused by prediction errors. Through the aforementioned global correction process, the prediction results for each sub-region are adjusted to meet the dual standards of coupling consistency and power flow feasibility in the regional interaction diagram. This not only improves the accuracy of the prediction results but also ensures their physical feasibility, avoiding contradictions between the prediction results and actual power grid operating conditions, thus enhancing the practical value and security of the predictions.

[0058] The above method comprehensively considers the mutual influence between sub-regions and the constraints of the power grid's physical topology by constructing a regional interaction graph, thereby achieving global optimization of the prediction results. The coupling consistency regularization term ensures the logical consistency of the prediction results, while the power flow feasible region constraint guarantees the physical feasibility of the prediction results. This dual guarantee makes the prediction results more stable and reliable, providing crucial support for the stable operation and dispatch decisions of the power system. In this way, even in complex and ever-changing power load forecasting environments, the accuracy and security of the predictions can be ensured, providing strong technical support for power system management.

[0059] As an optional implementation, supported by regional calibration priors, the central coordinating node performs graph consensus optimization. Combining coupling consistency regularization terms and power flow feasible region constraints, it globally corrects the prediction results of each sub-region and decomposes and writes the precise correction amount back to the corresponding sub-region. Specifically, the calibration priors submitted by each region are uniformly collected, and a regional interaction graph is constructed based on the physical topology of the power grid. Each region corresponds to a graph node, and their adjacency is determined by the transmission line connection situation and load coupling strength. Before constructing the graph structure, the contents contained in the calibration priors of each region need to be extracted, including the purified prediction input, feature confidence score, prediction residual variation trend, historical response coefficient, and cooperative correlation coefficient with adjacent regions. Using features with high load correlation as weights, connection strength values ​​are assigned to the edges in the graph, thereby generating a weighted graph. This graph not only reflects the physical transmission relationship but also embodies the structural correlation between prediction features. The feature data attached to all nodes will serve as input parameters for subsequent consensus optimization operations. Compared to traditional graph construction methods that rely solely on topology or historical power flow, this method introduces multivariate calibration priors, which improves the graph structure's ability to perceive abnormal data distributions.

[0060] Based on the graph structure, a cross-regional forecast consensus optimization operation is performed. This process does not aim to minimize the error of a single region, but rather pursues the overall coordination of the forecast behavior of all nodes in the graph across spatial and temporal dimensions. One of the optimization objectives is the difference in predicted power load values ​​between adjacent nodes in the graph. Simultaneously, it incorporates the confidence score from each node's calibration prior as a confidence adjustment factor to establish an optimization path based on structural coupling. To prevent abnormal states in one region from affecting the forecast results of neighboring regions, a coupling consistency regularization term is set to control the deviation between forecast outputs of highly correlated regions to not exceed the upper limit of their historical fluctuation range, ensuring trend synchronization and amplitude matching within regions with strong coupling relationships. Furthermore, to prevent local overfitting or excessive reverse adjustment, the offset between the forecast change rate of each node and its historical change trend is dynamically evaluated during the optimization process, and a maximum allowable slope change boundary is set. Through these mechanisms, graph consensus optimization not only strengthens the consistency of forecast outputs between regions but also ensures the stability of the forecast structure. This step differs from the approach of only employing local adaptive strategies in distributed forecast structures in related technologies; while ensuring local accuracy, it achieves proactive adjustment of forecast coupling between regions for the first time.

[0061] During the graph consensus optimization process, a power flow feasible region constraint based on the physical laws of power operation is introduced to ensure that the optimization results meet the actual power grid operation safety requirements. Specifically, when optimizing the predicted output value of each node, its adjustment result is mapped in real time to the corresponding power flow impact parameters, including node voltage changes, line power distribution, network power loss, and power flow direction stability indicators. For any predicted adjustment value, it is determined whether the mapping will cause the voltage deviation of the access node to exceed the technical standard limit, or cause load distribution exceeding limits or insufficient thermal stability margin in its downstream lines. Once it is detected that the adjusted predicted value will cause the line power flow to exceed the limit, the adjustment range in that area is immediately rolled back, and the adjustment amount is redistributed to adjacent areas with power flow acceptance capabilities, achieving local reconstruction of power flow adjustment. The entire constraint process relies on the power grid topology matrix and the sensitivity of power flow transmission between regions, forming an integrated iterative judgment mechanism of "prediction-power flow". This method avoids the fragmented approach of "prediction optimization first, power flow analysis later" in related technologies, controlling potential risks from the data source and ensuring that the prediction optimization results can be used for actual scheduling execution.

[0062] After consensus optimization and power flow verification, the predicted output of each node is differentially calculated from the original predicted value to obtain the precise correction amount for that node. This correction amount is then propagated upwards to its affected region according to the weight ratio of each edge in the graph structure. The correction amount contains two parts: one is the predicted trend correction value generated by the consensus mechanism, used to adjust the overall trend of the predicted curve; the other is the mandatory correction value generated by the power flow feasible region constraint, used to correct structural overshoots. The two correction amounts are added to form the final predicted adjustment vector, which is then decomposed into corresponding time periods for each sub-region in time series format, replacing the original predicted output. The write-back operation uses a bidirectional update method: on the one hand, it updates the output of the current prediction process in the sub-region; on the other hand, it updates its calibration prior history library, incorporating this correction action into the cumulative correction record, including the adjustment magnitude, source region, response time period, propagation path, and whether it is subject to power flow constraints. All correction amounts are encoded in a unified format and returned to each region for model iteration reference in the next prediction cycle. This process not only ensures a high degree of consistency in the predicted outputs of each region at the data level, but also allows for the tracing, analysis, and continuous optimization of the correction history, making it engineering reusable and evolvable.

[0063] As an optional implementation, in the system environment after the correction is written back, a time-reversal phase-gated control mechanism is activated to drive the conjugate energy injector to establish a reverse information channel. A pulse-level quenching operation is then performed in the distortion source region, synchronously feeding the calibration matrix back to the unified event time baseline. This achieves reversible closed-loop steady-state control of the cross-regional prediction system, suppressing the cascading amplification effect of the distortion signal. Specifically, in the time environment after the correction is written back to the corresponding regions, using the unified event time baseline as a reference, structural time-reversal processing is performed on all prediction input and output data of the distortion source region and its connected regions. The specific operations include: starting from the moment the distortion first appears, tracing back twelve complete prediction cycles in chronological order, extracting the load input data, prediction output sequence, feature selection state, residual response curve, and anomaly judgment label for each time period. Subsequently, based on the causal relationship between prediction errors and feature disturbances in each time period, a regional-level time-reversal mapping model is established. The disturbance direction and data increase trend are logically reversed to recover the entire path of evolution from a stable state to a distorted state. To ensure the accuracy and dynamic adaptability of data backtracking, a multi-point interpolation correction mechanism is embedded within each prediction window to maintain a balance between feature continuity and perturbation spectrum matching in the inversion results. Through this operation, the system obtains a complete "distortion formation trajectory," identifies the phase shift points of the distortion source region across multiple time domains, and locks the optimal phase intervention window accordingly.

[0064] Based on retrospective information, three highly sensitive time domains with the greatest intervention potential are selected as the core segments for phase adjustment, initiating a phase gating control strategy under time inversion. This strategy uses multi-period phase volatility as the control variable, applying small-amplitude intervention perturbations layer by layer within the gating window. The goal of each perturbation is to gradually adjust the temporal order and feature weight distribution in the feature inputs, bringing them closer to the phase structure of the stable interval. During implementation, a prediction output trend constraint coefficient is introduced to limit output discontinuities caused by over-adjustment; simultaneously, the maximum amplitude of each phase perturbation is set not to exceed one-tenth of the historical mean change range, thereby controlling the stability of the prediction results. After each phase adjustment, the effectiveness of the perturbation is judged by calculating three indicators: multi-region prediction synchronicity, residual average value, and load growth slope consistency. If the adjusted system state is detected to gradually converge to the previously stable prediction form, the gating perturbation is stopped, and the current time window is locked as the conjugate intervention benchmark window. This strategy, while maintaining the continuity of the prediction system, completes the active rearrangement of the asynchronous temporal chain, possessing high accuracy and strong retrospective capability.

[0065] Within the window after phase modulation stabilization, a conjugate energy sequence is constructed to establish a distortion cancellation path and drive the energy injection process to achieve regional reverse information reconstruction. This sequence is generated by constructing a dataset with a mirror relationship to the current predicted state of the distortion source region at the time-series level. The specific contents of the conjugate sequence include: the residual pullback vector after time inversion, the inverse vector of the prediction trend, and the inverse mapping channel of the perturbation. After fusing these three types of vector data, based on the continuous response characteristics of the prediction behavior in the perturbation sequence, low-amplitude, short-period conjugate perturbation signals are gradually injected into the original distortion path. After each injection, the changes in the root mean square error of the system prediction output, the spectral coupling degree of adjacent regions, and the prediction residual density are measured in real time to ensure that the energy injection does not cause prediction structure oscillation or coupling instability. The injection process adopts a dynamic step size control method that adjusts with the feedback gain. When the prediction structure response tends to be stable and the energy distribution in the conjugate channel shows a mean decay trend, the injection operation is considered to have achieved the suppression effect.

[0066] After the conjugate channel is closed, a local pulse quenching operation is performed in the distortion source region to completely shield the feature feedback that still has potential instability. The response matrix formed by the entire prediction correction process is then structured, encapsulated, and written back to the unified event time baseline. The quenching operation forcibly injects a zero-residual shielding sequence with a duration of less than half a cycle into the prediction path of the target region, while simultaneously shielding the feature-shared index between this region and adjacent regions, causing it to enter a prediction isolation state for a short time, thereby clearing any possible residual prediction resonance factors. Based on this, the endpoint state of the conjugate energy path, the adjusted prediction output, the time phase compensation amount, and the corrected responses of each region are combined into a complete calibration matrix, which is then uniformly written back to the unified event time baseline. All prediction tasks in future cycles will use this matrix as the historical background for prediction offset judgment and path selection, achieving early blocking and prevention of abnormal prediction behavior. The entire process constructs a reversible closed-loop control structure guided by state inversion, bridged by phase gating, mediated by energy cancellation, and terminated by calibration matrix closure, effectively blocking the risk of backflow and amplification of distortion signals across multiple regions and cycles.

[0067] Through the above steps S102 to S108, the goal of using a group of trusted anchor points to guide anomaly detection and control, constructing a distortion control boundary to prevent the propagation of distortion signals, and achieving accurate calibration of power load forecast results can be achieved. This improves the accuracy of power load forecast results and effectively suppresses the cascading effect of distortion signals, thereby solving the technical problem of distortion feature propagation and amplification in the forecasting of multi-regional distributed loads in power systems.

[0068] Based on the above embodiments and optional embodiments, the present invention proposes an optional implementation method. Figure 2 This is a flowchart of an optional power load forecasting processing method according to an embodiment of the present invention, such as... Figure 2 As shown, the method includes:

[0069] S001. Under a unified event time baseline, a cross-regional topology time series diagram is constructed. By processing load data and meteorological data, characteristic quality spectra are extracted, and noise fingerprints and drift slopes are calculated to locate potential distortion source areas at the system level and establish an analytical benchmark for distortion source tracing. In order to effectively identify potential distortion source areas in distributed power load forecasting, a unified event time baseline is constructed, and multi-sub-region topology data processing, feature extraction, and distortion source tracing operations are carried out. The specific implementation process is the same as the aforementioned embodiment and will not be repeated here.

[0070] S002, under the constraints of the analysis benchmark, a counterfactual playback sequence for the distortion source region is generated, and extreme dynamic segments are replayed frame by frame. Suspicious low-dimensional features are identified by combining the phase-locked consistency index and the quantile manifold embedding method, forming a group of credible anchor points for subsequent anomaly determination. After completing the localization of the distortion source in the target region and establishing a cross-regional topological time series diagram and analysis benchmark based on a unified event time baseline, the specific implementation process for further identifying extreme dynamic segments and extracting reliable reference feature points is the same as in the aforementioned embodiments, and will not be repeated here.

[0071] In step S003, under the constraint of a trusted anchor point group, micro-amplitude temporal perturbations are injected into the sub-regions adjacent to the distortion source region. The cross-regional residual gradient field is measured in real time to identify the propagation path and amplification threshold (i.e., propagation critical point) of distortion features across multiple regions, serving as the boundary conditions for distortion control (i.e., the distortion control boundary). In the preceding steps, the trusted anchor point group has been extracted, and the location of the distortion source region and its key time window have been clarified. To analyze the cross-regional propagation mechanism of anomalous features, perturbation response injection and residual gradient measurement are performed based on this, and the anomalous diffusion path and amplification critical point are identified to construct the distortion control boundary. The specific implementation process is the same as in the aforementioned embodiments and will not be repeated here.

[0072] S004, under boundary constraints, an adaptive isolation valve is constructed, and a feature purification chain is initiated along the propagation path. Heterogeneous redundancy comparison and trusted summary replacement mechanisms are used to actively mitigate the influence of distorted features, generating regional calibration prior data as reference information for subsequent corrections. Based on the identified distortion propagation path and regional amplification threshold, to prevent further cross-regional propagation of local distortion information and interference with the global prediction results, the feature purification chain needs to be triggered based on the constructed boundary constraints, generating trusted regional calibration data as input for subsequent optimization. The specific implementation process is the same as in the aforementioned embodiments and will not be repeated here.

[0073] S005, with the support of regional calibration priors, the central coordination node performs graph consensus optimization, combining coupling consistency regularization terms and power flow feasible region constraints to globally correct the prediction results of each sub-region, and decomposes the precise correction amount and writes it back to the corresponding sub-region. With the support of the constructed regional calibration priors, in order to ensure that the prediction results of each region have logical consistency and power flow feasibility in the overall network, the central coordination node needs to perform a unified correction operation oriented towards the topology based on the multi-region structured prediction data, and accurately decompose the correction results and synchronously update them to the prediction process of each region. The specific implementation process is the same as the aforementioned embodiment, and will not be repeated here.

[0074] In step S006, within the system environment after the correction amount is written back, a time-reversal phase-gated control mechanism is initiated to drive the conjugate energy injector to establish a reverse information channel. A pulse-level quenching operation is then performed in the distortion source region, synchronously feeding the calibration matrix back to the unified event time baseline. This achieves reversible closed-loop steady-state control of the cross-regional prediction system, suppressing the cascading amplification effect of the distortion signal. After the central coordination node completes global prediction correction and synchronously writes back the correction amounts from each region, the system needs to construct a reversible closed-loop structure to further eliminate residual distortion effects and prevent its re-cascading propagation. To this end, a time-reversal mechanism, a phase-gated control strategy, a conjugate energy injection method, and a distortion pulse quenching operation are introduced to complete the closed-loop feedback of the calibration matrix and the steady-state locking of the prediction channel. The specific implementation process is the same as in the aforementioned embodiments and will not be repeated here.

[0075] It should be noted that the method in this embodiment constructs a globally observable topological time series graph by unifying the event time baseline and integrates counterfactual playback and anchor point extraction to accurately lock onto the formation and propagation paths of anomalies. Based on this, it dynamically identifies and actively blocks cascading distortion propagation paths by combining residual gradient injection and feature purification chain control. Furthermore, it comprehensively improves the consistency and accuracy of multi-region prediction results through graph consensus optimization and power flow feasible region correction. Finally, it constructs a prediction input closed loop through conjugate energy injection and pulse extinction mechanisms to achieve steady-state regulation and time series memory feedback of the global prediction environment. The overall scheme has significant advantages such as high prediction accuracy, fast response speed, strong anomaly suppression capability, and strong reversible closed-loop control capability, providing highly robust and scalable technical support for multi-region distributed load prediction in large-scale complex power systems.

[0076] Based on the above embodiments and optional embodiments, the present invention also proposes another optional implementation method. Figure 3 This is a schematic diagram of an optional power load forecasting and processing system according to an embodiment of the present invention, such as... Figure 3 As shown, the system includes: a distortion localization module, an anchor point extraction module, a propagation identification module, a purification and isolation module, a global correction module, and a steady-state closed-loop control module, wherein:

[0077] The distortion localization module is used to construct cross-regional topological time series maps based on a unified event time baseline, process load and meteorological data, extract characteristic quality spectra and calculate noise fingerprints and drift slopes, locate distortion source areas and establish analysis benchmarks;

[0078] The anchor point extraction module is used to generate counterfactual playback sequences under the analysis benchmark, replay dynamic segments of the target area frame by frame, and combine phase-locked signals and manifold embedding to identify suspicious features and form a group of credible anchor points.

[0079] The propagation identification module is used to inject micro-amplitude perturbations into adjacent sub-regions under the constraint of a group of trusted anchor points, measure the cross-regional residual gradient field, identify the distortion propagation path and amplification threshold, and form boundary conditions.

[0080] The purification and isolation module is used to construct an adaptive isolation valve under boundary conditions, initiate a feature purification chain along the propagation path, suppress distortion based on heterogeneous comparison and trusted summary replacement mechanism, and generate regional calibration priors.

[0081] The global correction module is used to perform global correction on the prediction results of each sub-region based on regional calibration priors, with the central coordination node performing graph consensus optimization, and combining coupled regularization and power flow constraints, and decomposes and writes back the correction amount.

[0082] A steady-state closed-loop control module is used to initiate time-inversion phase gating after the correction amount is written back, drive the conjugate energy injector to establish a reverse channel, perform a shutdown operation in the target region, and feed back the calibration matrix to a unified time baseline to achieve closed-loop steady-state control of the system and suppress cascaded distortion amplification. This embodiment also provides a power load forecasting processing device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the terms "module" and "device" can refer to a combination of software and / or hardware that performs a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0083] According to an embodiment of the present invention, an apparatus embodiment for implementing the above-described power load forecasting processing method is also provided. Figure 4 This is a schematic diagram of the structure of a power load forecasting and processing device according to an embodiment of the present invention, as shown below. Figure 4 As shown, the above-mentioned power load forecasting processing device includes: an anchor point group determination module 400, a boundary determination module 402, a priori data generation module 404, and a load forecasting correction module 406, wherein:

[0084] Anchor group determination module 400, connected to anchor group determination module 400, is used to determine a reliable anchor group based on extreme dynamic segments of the target area within the target time period. The target area is an area with distortion characteristics in multiple sub-regions, and the extreme dynamic segments are feature data with a change amplitude greater than a preset amplitude. The feature data includes power load data and meteorological data. The reliable anchor group is used to indicate feature reference points under normal load change trends.

[0085] The boundary determination module 402 is connected to the anchor point group determination module 400. It is used to perform perturbation processing on the feature data corresponding to the adjacent sub-regions of the target region under the constraint of the trusted anchor point group, and determine the distortion control boundary. The distortion control boundary is obtained based on the distortion propagation path and amplification critical point of the distortion features among multiple sub-regions.

[0086] The prior data generation module 404 is connected to the boundary determination module 402 and is used to generate regional calibration prior data based on the distortion control boundary. The regional calibration prior data is used to indicate the power load forecast trend and feature weight adjustment standard of each region after eliminating distortion features, as well as the power load forecast offset compensation amount under specific conditions.

[0087] The load forecasting correction module 406 is connected to the prior data generation module 404. It is used to correct the power load forecasting results corresponding to each of the multiple sub-regions based on the regional calibration prior data, so as to obtain the corrected power load forecasting results corresponding to each of the multiple sub-regions.

[0088] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0089] It should be noted that the anchor point group determination module 400, boundary determination module 402, prior data generation module 404, and load prediction correction module 406 mentioned above correspond to steps S102 to S108 in the embodiments. The instances and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules, as part of the device, can run on a computer terminal.

[0090] It should be noted that the optional or preferred implementation methods of this embodiment can be found in the relevant descriptions in the embodiments, and will not be repeated here.

[0091] The aforementioned power load forecasting processing device may also include a processor and a memory. The aforementioned anchor point group determination module 400, boundary determination module 402, prior data generation module 404, load forecasting correction module 406, etc., are all stored in the memory as program modules, and the processor executes the aforementioned program modules stored in the memory to realize the corresponding functions.

[0092] The processor contains a core that retrieves the corresponding program modules from memory. One or more cores may be configured. Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip.

[0093] According to an embodiment of this application, an embodiment of a non-volatile storage medium is also provided. Optionally, in this embodiment, the non-volatile storage medium includes a stored program, wherein, when the program runs, it controls the device where the non-volatile storage medium is located to execute any of the above-mentioned power load forecasting processing methods.

[0094] Optionally, in this embodiment, the non-volatile storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals, and the non-volatile storage medium includes stored programs.

[0095] Optionally, a program that controls the device containing the non-volatile storage medium to execute any of the above-mentioned power load prediction processing steps during program execution.

[0096] According to an embodiment of this application, an embodiment of a processor is also provided. Optionally, in this embodiment, the processor is used to run a program, wherein the program executes any of the above-described power load forecasting processing methods.

[0097] According to an embodiment of this application, an embodiment of a computer program product is also provided, which, when executed on a data processing device, is adapted to execute a program that initializes the power load forecasting processing method steps described above.

[0098] This invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the above-described power load forecasting processing methods.

[0099] The order of the above embodiments of the present invention is merely for description and does not represent the superiority or inferiority of the embodiments.

[0100] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of modules described above can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between modules, and may be electrical or other forms.

[0102] The modules described above as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0103] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0104] If the aforementioned integrated modules are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable non-volatile storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a non-volatile storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned non-volatile storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0105] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting and processing electricity load, characterized in that, include: Based on extreme dynamic segments of the target area within the target time period, a group of reliable anchor points is determined. The target area is a region with distortion characteristics among multiple sub-regions. The extreme dynamic segments are feature data with a change amplitude greater than a preset amplitude. The feature data includes power load data and meteorological data. The group of reliable anchor points is used to indicate feature benchmark points under normal load change trends. Under the constraint of the trusted anchor point group, the feature data corresponding to the adjacent sub-regions of the target region are perturbed to determine the distortion control boundary. The distortion control boundary is obtained based on the distortion propagation path and amplification critical point of the distortion features among multiple sub-regions. Based on the distortion control boundary, regional calibration prior data is generated, wherein the regional calibration prior data is used to indicate the power load forecasting trend and feature weight adjustment criteria of each region after eliminating distortion features, as well as the power load forecasting offset compensation amount under specific conditions. Based on the regional calibration prior data, the power load prediction results corresponding to each of the multiple sub-regions are corrected to obtain the corrected power load prediction results corresponding to each of the multiple sub-regions.

2. The method according to claim 1, characterized in that, Before determining the group of trusted anchor points based on extreme dynamic segments of the target region within the target time period, the method further includes: Based on the characteristic data of each of the multiple sub-regions, a characteristic quality spectrum is constructed, wherein the characteristic quality spectrum is used to characterize the abnormal fluctuation trend in the characteristic data, as well as the degree of coordination between load data and meteorological data; Based on the characteristic quality spectrum, noise fingerprint data and drift slope corresponding to each of the multiple sub-regions are determined, wherein the noise fingerprint data is used to indicate the statistical characteristics of a predetermined abnormal pattern in the characteristic data; the drift slope is used to indicate the stability of the characteristic data over time and to identify regions where the characteristic data continuously shifts. The noise fingerprint data and drift slope corresponding to each of the multiple sub-regions are mapped to a cross-regional topology time series diagram to obtain the mapped topology time series diagram, as well as the anomaly cumulative score and cooperation relationship identifier corresponding to each of the multiple sub-regions. The cross-regional topology time series diagram is constructed based on the power grid physical topology relationship and geographical adjacency relationship between the multiple sub-regions and is used to characterize the dynamic coupling and information exchange path between the multiple sub-regions. The cooperation relationship identifier is used to indicate the degree of synchronization and spatial correlation of the feature data of each region in the time series. Multiple candidate regions are identified in the mapped topology time sequence graph. Based on whether there is an abnormal propagation path between the multiple candidate regions and their corresponding neighboring regions, the target region is determined from the multiple candidate regions. The candidate regions are the regions in the multiple sub-regions whose abnormal cumulative scores are greater than a preset score threshold, and the neighboring regions are the regions within a predetermined neighborhood of the corresponding candidate regions.

3. The method according to claim 1, characterized in that, Before determining the group of trusted anchor points based on extreme dynamic segments of the target region within the target time period, the method further includes: Obtain feature data of the target area during a predetermined time period, wherein the predetermined time period is a predetermined duration before and after the distortion of the target area occurs; The feature data of the predetermined time period is processed by framing according to a preset sliding window to obtain a multi-frame data sequence; Based on the time-varying indices corresponding to each of the multi-frame data sequences, a target number of data sequences are determined from the multi-frame data sequences as the extreme dynamic segments. The time-varying indices are used to quantify the dynamic characteristics of each frame of data over time. The time-varying indices include the instantaneous rate of change of the load curve, the curvature of the temperature change, the wind speed vector transformation rate, the atmospheric pressure gradient, the amplitude of rainfall fluctuations, and the rate of increase in solar radiation.

4. The method according to claim 1, characterized in that, The determination of a group of reliable anchor points based on extreme dynamic segments of the target region within the target time period includes: Phase verification is performed on the extreme dynamic segment to obtain phase verification results, wherein the phase verification is used to verify the degree of synchronous change between the load variable and the meteorological variable in the time-varying indicators of the corresponding data sequence; The extreme dynamic fragment is subjected to structural verification to obtain structural verification results, wherein the structural verification is used to verify the degree of clustering and isolation of the corresponding data sequence in the low-dimensional feature space mapping; Based on the phase verification results and the structure verification results, multiple candidate trusted anchor points are determined from the extreme dynamic segment; The credibility of the multiple candidate trustworthy anchor points is evaluated to obtain the credibility evaluation results; Based on the credibility assessment results, credible anchor points that meet the preset constraints are selected from the multiple candidate credible anchor points to obtain the credible anchor point group.

5. The method according to claim 1, characterized in that, The step of perturbing the feature data corresponding to adjacent sub-regions of the target region under the constraint of the trusted anchor point group to determine the distortion control boundary includes: Under the constraint of the trusted anchor group, the feature data of the adjacent sub-regions of the target region are perturbed by a micro-amplitude temporal perturbation, wherein the micro-amplitude temporal perturbation is to apply a perturbation not exceeding a preset perturbation amplitude to the original feature data; Determine the residual value of the power load forecast results corresponding to the adjacent sub-regions, wherein the residual value represents the difference between the power load forecast results of the adjacent sub-regions before and after the disturbance, and is used to quantify the degree of impact of the disturbance on the accuracy of power load forecasting; Based on the cross-regional residual gradient field, the residual variation characteristics are determined. The cross-regional residual gradient field is a multi-dimensional residual feature constructed based on the changes in the power load prediction results of the adjacent sub-regions. It is used to show the diffusion trend and intensity change of the residual value between different sub-regions. The residual variation characteristics are used to indicate the propagation path and amplification law of the residual value over time. Based on the residual change characteristics, the distortion propagation path and amplification critical point of the distortion characteristics among multiple sub-regions are identified as the distortion control boundary. The distortion propagation path is used to describe the path of the distortion characteristics spreading from the target region to other regions, and the amplification critical point is the threshold of the amount of characteristic data change that causes the deviation of the power load forecast result under a specific disturbance.

6. The method according to claim 1, characterized in that, The generation of regional calibration prior data based on the distortion control boundary includes: Based on the distortion control boundary, an adaptive isolation valve is constructed to actively suppress the propagation of distortion features between sub-regions. The adaptive isolation valve is used to dynamically adjust the corresponding isolation or release strategy according to the real-time detected feature data status. Under the control of the adaptive isolation valve, a feature cleanup chain is initiated along the distortion propagation path in the distortion control boundary. Using heterogeneous redundancy comparison and trusted summary replacement methods, the distortion features in the feature data corresponding to each of the multiple sub-regions are processed to obtain the region calibration prior data. The heterogeneous redundancy comparison refers to detecting anomalies by comparing feature data from different sources or types. The trusted summary replacement mechanism is used to replace the distortion features with summary data from a trusted source after detecting distortion features.

7. The method according to claim 1, characterized in that, The step of correcting the power load forecast results corresponding to each of the multiple sub-regions based on the regional calibration prior data to obtain the corrected power load forecast results corresponding to each of the multiple sub-regions includes: Based on the prior calibration data of the region and the power grid physical topology relationship between the multiple sub-regions, a regional interaction graph is constructed. The regional interaction graph uses each sub-region as a graph node to indicate the dynamic coupling relationship and information flow path between each sub-region, as well as the prediction error weight based on historical data evaluation. The prediction error weight is used to reflect the reliability and historical volatility of the load prediction results of each sub-region. Based on the regional interaction graph, the coupling consistency regularization term and the power flow feasible region constraint are used to globally correct the power load prediction results corresponding to each of the multiple sub-regions, thereby obtaining the corrected power load prediction results corresponding to each of the multiple sub-regions. The coupling consistency regularization term is used to control the deviation of the prediction output of highly correlated sub-regions to not exceed the upper limit of the historical fluctuation range. The highly correlated sub-regions refer to sub-regions with a correlation greater than a preset correlation threshold. The power flow feasible region constraint is used to control the load operation of the multiple sub-regions to conform to the power grid operation safety boundary.

8. A power load forecasting and processing device, characterized in that, include: Anchor point group determination module is used to determine a reliable anchor point group based on extreme dynamic segments of a target area within a target time period. The target area is an area with distortion characteristics in multiple sub-regions, and the extreme dynamic segments are feature data with a change amplitude greater than a preset amplitude. The feature data includes power load data and meteorological data. The reliable anchor point group is used to indicate feature benchmark points under normal load change trends. The boundary determination module is used to perturb the feature data corresponding to the adjacent sub-regions of the target region under the constraints of the trusted anchor point group, and determine the distortion control boundary, wherein the distortion control boundary is obtained based on the distortion propagation path and amplification critical point of the distortion features among multiple sub-regions; The prior data generation module is used to generate regional calibration prior data based on the distortion control boundary, wherein the regional calibration prior data is used to indicate the power load forecast trend and feature weight adjustment standard of each region after eliminating distortion features, as well as the power load forecast offset compensation amount under specific conditions. The load forecasting correction module is used to correct the power load forecasting results corresponding to each of the multiple sub-regions based on the regional calibration prior data, so as to obtain the corrected power load forecasting results corresponding to each of the multiple sub-regions.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions, which are adapted to be loaded by a processor and executed by the power load forecasting processing method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the power load forecasting processing method according to any one of claims 1 to 7.

Images