A power load forecasting method and system based on spatial and temporal collaborative correction

Abstract

This invention discloses a power load forecasting method and system based on spatial and temporal collaborative correction, belonging to the field of power load forecasting technology. The method generates a basic load forecast result for a target time based on historical power load data, meteorological data, and calendar event data through a rule-based prediction model; it acquires historical power load data of multiple related load objects, and uses a spatial correlation learning model to correct the basic load forecast result to reflect the behavioral similarity between related objects and generate a first corrected forecast result; based on the historical actual load data of the target load object, a time correction amount is constructed through a temporal correlation online learning model to dynamically correct the first corrected forecast result, thereby obtaining the final power load forecast result. This invention improves the accuracy, stability, and trend consistency of load forecast results through collaborative modeling of spatial and temporal dimensions, and is applicable to scenarios such as power grid dispatching, energy management, and load monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, in particular to a power load prediction method and system based on spatial and temporal collaborative correction. BACKGROUND

[0002] Power load prediction is a key basic work in power system dispatching, operation control and planning construction, and its prediction results directly affect the safety, economy and supply-demand balance ability of the power grid. In actual operation process, power load is usually simultaneously affected by multiple factors such as meteorological changes, user behavior patterns, load linkage effects between regions and time sequence change rules, so that the load prediction presents obvious multi-dimensional and multi-time-varying characteristics.

[0003] Traditional load prediction methods mainly rely on rule models or single time sequence models based on historical statistical rules, and realize prediction through fixed logic operation on historical power load data, meteorological factors or holiday information. Such methods are simple to implement, but often have the following technical limitations: the rule model usually only corrects through single-dimensional logic such as trend, temperature or holiday, and it is difficult to accurately capture the load change characteristics under the coupling of multiple factors, resulting in large deviation of the prediction baseline in complex scenarios. In actual load changes, there are a large number of associated load objects with similar behavior patterns or affected by the same environmental factors. The traditional method is difficult to effectively mine the spatial correlation between different load objects, and the basic prediction result cannot be collaboratively compensated through the information of the spatial dimension. The load prediction error has cumulative and directional characteristics, and if there is no effective time dimension online correction mechanism, the prediction result is easy to produce continuous deviation or abnormal mutation, and the traditional method cannot track and suppress the cumulative effect of residual error in the time dimension. In periods of sharp load changes or volatility, conventional prediction often appears mutation opposite to the previous trend, which is difficult to meet the requirements of continuity, smoothness and trend consistency for power grid dispatching.

[0004] Therefore, it is urgent to propose a load prediction technical scheme that can collaboratively improve the basic prediction result in the spatial and time dimensions, so as to improve the adaptability of the prediction model to complex load changes and realize high-precision, strong-robustness and trend-continuous power load prediction. SUMMARY

[0005] The present application aims to provide a power load prediction method and system based on spatial and temporal collaborative correction, to solve the limitations of existing load prediction technology such as insufficient basic prediction accuracy, inability to effectively utilize the spatial correlation between associated loads, and lack of continuous tracking and dynamic correction ability for load time sequence deviation, and to realize comprehensive response to multi-dimensional driving factors of load changes, significantly improving the accuracy, stability and trend continuity of the prediction results.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] On the one hand, the present invention provides a power load forecasting method based on spatial and temporal collaborative correction, comprising the following steps:

[0008] S1. Based on historical power load data, meteorological data, and calendar event data at the target time, generate the basic load forecast result for the target time through a rule-based prediction model;

[0009] S2. Obtain historical power load data of multiple associated load objects, and based on the historical power load data of the associated load objects and the basic load prediction results, correct the basic load prediction results through a spatial correlation learning model to generate a first corrected prediction result.

[0010] S3. Obtain historical actual load data of the target load object. Based on the historical actual load data and the first corrected prediction result, correct the first corrected prediction result online through a time correlation online learning model to generate the final power load prediction result at the target time.

[0011] As a preferred embodiment of the present invention, the rule-based prediction model extracts trends and performs seasonal analysis on the historical power load data of the target load object, establishes a temperature response relationship based on meteorological data, and corrects the impact of holidays by combining calendar event data, thereby obtaining a structured basic prediction feature containing trend components, temperature response components, and calendar event components. The structured basic prediction feature is provided to the spatial correlation learning model to perform collaborative correction processing on the basic load prediction results.

[0012] As a preferred embodiment of the present invention, the execution process of the spatial correlation learning model includes: constructing spatial correlation features based on load data of multiple associated load objects; fusing the spatial correlation features with structured basic prediction features to obtain fused features; determining the spatial correlation weights between the target load object and each associated load object based on the fused features; and performing collaborative calculation processing on the load features reflected by the load data of the associated load objects and the basic load prediction results based on the spatial correlation weights to generate the first corrected prediction result.

[0013] As a preferred embodiment of the present invention, the plurality of associated load objects are determined by the following method: from the load object pool, load objects that are similar to the target load object in terms of geographical location, load curve shape characteristics, or user type are selected; the load data of each selected associated load object are subjected to time alignment and numerical normalization processing to form the associated load dataset required by the spatial correlation learning model.

[0014] As a preferred embodiment of the present invention, the spatial correlation learning model generates spatial correction quantities based on structured basic prediction features and associated load datasets, and the feature distances between the target load object and each associated load object. Similarity weight Calculated using the following formula:

[0015] ;

[0016] in: , This represents the trend components of the target load object and its associated load objects; , This represents the temperature response components of the target load object and the associated load object; , Represents the calendar event components of the target load object and its associated load objects; , , This represents a non-negative coefficient used to adjust the relative weights of the three types of components; This represents the weighted feature distance obtained based on the differences between the three types of components; This represents the similarity weight obtained from the feature distance mapping;

[0017] The residual signals of the associated load objects are weighted based on the similarity weights to generate a spatial correction amount, which is then applied to the basic load prediction results to obtain the first corrected prediction result.

[0018] Residual signal: The difference between the actual load of the associated load object and the predicted basic load;

[0019] Spatial correction: The result of aggregated residual signals weighted by similarity weights.

[0020] As a preferred embodiment of the present invention, the time-correlation online learning model generates a time correction amount based on historical actual load data and the first corrected prediction result. Compared with the final power load forecast results Satisfy the following formula:

[0021] ;

[0022] in: Indicates the target load object at time... The first revised forecast result; Indicates the target load object at time... The final power load forecast results; This represents the time accumulation amount obtained by weighting the residual signal according to a preset attenuation coefficient; This represents the time correction amount obtained by applying the aforementioned nonlinear mapping to the time accumulation; the time accumulation amount It is obtained by weighting and accumulating the residual signals at each time step according to a preset attenuation coefficient.

[0023] As a preferred embodiment of the present invention, when calculating the time accumulation, the spatial correction amount is used as the initial value of the time accumulation, and the update amplitude of the time accumulation is dynamically adjusted based on the directional consistency between the spatial correction amount and the current residual signal, so that the correction direction formed in the spatial correction stage is maintained or weakened in the time dimension, thus constituting a cross-stage dynamic closed-loop adjustment process.

[0024] As a preferred embodiment of the present invention, before outputting the final power load prediction result, the rate of change of the final power load prediction result at adjacent time points is detected; if the rate of change of the current prediction result is opposite to the trend change direction formed at the previous time point, and the absolute value exceeds the preset change magnitude threshold, then a constraint correction process is performed on the final power load prediction result to ensure that the final power load prediction result maintains continuous trend response characteristics in the time dimension.

[0025] On the other hand, the present invention also provides a power load forecasting system based on spatial and temporal co-correction, applied to the power load forecasting method based on spatial and temporal co-correction as described above, the system comprising:

[0026] The basic forecasting module is used to acquire historical power load data, meteorological data, and calendar event data for the target time, and generate basic load forecast results for the target time through a rule-based forecasting model based on the historical power load data, meteorological data, and calendar event data.

[0027] The spatial correction module is used to acquire historical power load data of multiple associated load objects, and based on the historical power load data of the associated load objects and the basic load prediction results, correct the basic load prediction results through a spatial correlation learning model to generate a first corrected prediction result.

[0028] The time correction module is used to acquire historical actual load data of the target load object, and based on the historical actual load data and the first corrected prediction result, to correct the first corrected prediction result online through a time correlation online learning model, so as to generate the final power load prediction result at the target time.

[0029] The beneficial effects of this invention are as follows: By sequentially combining basic forecasting, spatial collaborative correction, and online temporal correction, this invention constructs a load forecasting system capable of comprehensively utilizing multi-source data, spatial correlation information, and temporal deviation signals. The basic forecasting stage enables the forecasting process to accurately reflect the trend changes, meteorological sensitivity, and calendar event characteristics of the target load object, providing a stable and reliable forecasting baseline for subsequent corrections. The spatial collaborative correction stage introduces historical power load data from multiple related load objects, allowing the forecasting model to obtain compensation information from the behavioral similarities between load objects, effectively compensating for the shortcomings of single-object forecasting in spatial perception, thereby improving the adaptability of the forecast results to load changes in complex regions. The online temporal correction stage constructs residual accumulation based on historical actual loads and combines nonlinear mapping and directional consistency judgment mechanisms to achieve dynamic suppression and trend maintenance of current forecast deviations, making the forecast results more continuous, stable, and adaptive in the time dimension. Through the above multi-dimensional collaborative mechanisms, this invention significantly improves the accuracy, robustness, and trend consistency of load forecasting, maintaining excellent performance under different operating scenarios, thereby enhancing the overall engineering application value of load forecasting. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 This is a schematic diagram of the modular structure of the system of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0034] Example 1

[0035] like Figure 1As shown, in one embodiment of the present invention, a basic forecasting step is first performed to obtain an initial forecast result for the target time. Specifically, S1, based on historical power load data, meteorological data, and calendar event data for the target time, a basic load forecast result for the target time is generated through a rule-based forecasting model.

[0036] In this embodiment, the historical power load data includes the actual power consumption time series of the target load object over multiple historical periods. Those skilled in the art can select a cleaned load curve as input based on the sampling period of the actual scenario (e.g., 15 minutes, 30 minutes, or 1 hour). Meteorological data may include factors affecting load changes such as temperature, humidity, and weather conditions; calendar event data may include social behavior-related factors such as holiday information, weekday and weekend markers.

[0037] To extract the structured information required for basic forecasting from the aforementioned multi-source data, this embodiment employs a rule-based forecasting model to analyze the input data. Specifically, the rule-based forecasting model extracts trends and performs seasonal analysis on the historical load sequences of the target load object, establishes temperature response relationships based on meteorological data, and corrects for the impact of holidays by incorporating calendar event data. This yields structured basic forecasting features containing trend components, temperature response components, and calendar event components. These structured basic forecasting features are then provided to a spatial correlation learning model for collaborative correction of the basic load forecasting results.

[0038] In practice, trend extraction can be achieved by calculating the moving average, regression trend, or periodic pattern of historical load sequences, thereby obtaining trend components reflecting long-term changing trends. Seasonal analysis can extract typical seasonal load patterns by comparing loads of the same season, month, or adjacent periods in history. Temperature response relationships can be established based on the statistical correlation between meteorological temperature and electricity consumption, such as using piecewise linear relationships or load change mapping based on historical temperature ranges, to characterize the impact of temperature changes on target load objects. Calendar event corrections can be applied to the base forecast by weighting or shifting it according to the differences in load patterns between holidays and non-holidays, making the forecast baseline closer to actual load behavior.

[0039] Through the above analysis process, the obtained structured basic prediction features can cover the main driving factors of load changes, forming a preliminary estimate of the load at the future target time. In subsequent steps, this invention will use this basic prediction result and other related information to further perform spatial and temporal corrections, making the prediction results more accurate and robust.

[0040] In one embodiment of the present invention, in order to fully utilize the behavioral correlation between the target load object and other load objects and improve the accuracy of the basic prediction results, this embodiment further performs a spatial collaborative correction step after generating the basic prediction results. Specifically, S2, historical power load data of multiple associated load objects are obtained, and based on the historical power load data of the associated load objects and the basic load prediction results, the basic load prediction results are corrected through a spatial correlation learning model to generate a first corrected prediction result.

[0041] In practical implementation, the data for associated load objects can come from power system databases, regional distribution system monitoring platforms, or other load-side data resources. To ensure that the selected objects reflect spatial correlations meaningful to the target load, this embodiment determines associated load objects in the following way: The multiple associated load objects are determined by: filtering load objects from the load object pool that are similar to the target load object in terms of geographical location, load curve shape characteristics, or user type; performing time alignment and numerical normalization processing on the load data of each filtered associated load object to form the associated load dataset required for the spatial correlation learning model.

[0042] Time alignment addresses the issue of inconsistent timestamps between different load objects. By standardizing sampling periods and aligning time ranges, it ensures that load sequences for all objects correspond one-to-one in time. Numerical normalization eliminates amplitude differences, ensuring that spatial correlation calculations are unaffected by different user load levels. For example, min-max normalization can be used to map all load sequences to the same scale interval.

[0043] After obtaining the associated load dataset, this embodiment uses a spatial correlation learning model to calculate the spatial correlation strength between the target object and associated objects. Specifically, the execution process of the spatial correlation learning model includes: constructing spatial correlation features based on the load data of multiple associated load objects; fusing the spatial correlation features with structured basic prediction features to obtain fused features; determining the spatial correlation weights between the target load object and each associated load object based on the fused features; and performing collaborative calculation processing on the load features reflected by the load data of the associated load objects and the basic load prediction results based on the spatial correlation weights to generate the first corrected prediction result.

[0044] Specifically, spatial correlation characteristics may include load change trend characteristics, diurnal periodic patterns, temperature sensitivity characteristics, etc. By integrating these characteristics with basic forecast characteristics, the degree of similarity between different load objects can be more comprehensively characterized.

[0045] Furthermore, the spatial correlation learning model generates spatial corrections based on structured basic prediction features and associated load datasets. The feature distance and similarity weights between the target load object and each associated load object are calculated using the following formula:

[0046] ;

[0047] in: , This represents the trend components of the target load object and its associated load objects; , This represents the temperature response components of the target load object and the associated load object; , Represents the calendar event components of the target load object and its associated load objects; , , This represents a non-negative coefficient used to adjust the relative weights of the three types of components; This represents the weighted feature distance obtained based on the differences between the three types of components; This represents the similarity weight obtained from the feature distance mapping;

[0048] In practical operation, those skilled in the art can select different weighting coefficients according to specific scenarios, making the model pay more attention to trend changes, temperature response differences, or holiday characteristics. Smaller feature distances indicate that the behaviors between load objects are more similar, and the corresponding similarity weights are higher, and vice versa.

[0049] The spatial correction is obtained by weighted aggregation of the residuals of the associated load objects relative to the basic forecast results. Specifically, the residual signals of the associated load objects are weighted based on the similarity weights to generate the spatial correction, which is then applied to the basic load forecast results to obtain the first corrected forecast result. Residual signal: the difference between the actual load of the associated load object and the basic load forecast result; Spatial correction: the aggregated result of the residual signals weighted by the similarity weights.

[0050] Through the above process, the changing characteristics of the associated load objects can effectively compensate for the basic prediction results of the target load objects, correcting potential deviations in the basic prediction in the spatial dimension, thereby significantly improving prediction accuracy. In the following steps, this invention further incorporates online correction in the time dimension, achieving a synergistic fusion of spatial and temporal corrections to obtain even higher quality prediction results.

[0051] In one embodiment of the present invention, after completing the spatial dimension collaborative correction and obtaining the first corrected prediction result, in order to further improve the responsiveness of the prediction result to changes in the time series, this embodiment continues to perform the time collaborative online correction step. Specifically, S3, acquire the historical actual load data of the target load object, and based on the historical actual load data and the first corrected prediction result, correct the first corrected prediction result online through a time correlation online learning model to generate the final power load prediction result at the target time.

[0052] In actual operation, historical load data is used to reflect the true trajectory of the target load in the most recent period, and this data can provide real-time feedback information on prediction deviations. By combining this information with the first revised prediction result, dynamic adjustment of prediction deviations can be achieved, enabling the model to have "self-correcting capability" as it evolves over time.

[0053] To achieve the aforementioned online correction, this embodiment constructs a time correction amount and generates the final prediction result through nonlinear mapping. Specifically, the time correlation online learning model generates the time correction amount based on historical actual load data and the first correction prediction result. The time correction amount and the final power load prediction result satisfy the following formula:

[0054] ;

[0055] in: Indicates the target load object at time... The first revised forecast result; Indicates the target load object at time... The final power load forecast results; This represents the time accumulation amount obtained by weighting the residual signal according to a preset attenuation coefficient; This represents the time correction amount obtained by applying the aforementioned nonlinear mapping to the time accumulation; the time accumulation amount It is obtained by weighting and accumulating the residual signals at each time step according to a preset attenuation coefficient.

[0056] Among them, residual signal It can be represented as:

[0057] ;

[0058] in: This represents the actual load. The preset attenuation coefficient can be understood as a weighting factor used to limit the influence range of historical residuals, making newer residuals less influential. The contribution of older residuals is greater, while the contribution of older residuals gradually weakens, thus avoiding the infinite growth of time accumulation terms.

[0059] To prevent abnormal reverse residuals from causing the time accumulation to deviate from the normal trend, this embodiment further introduces a residual direction consistency control rule. Specifically, when constructing the time accumulation, the residual signal at each time step is processed based on the continuity of the residual direction. When it is detected that the sign of the current residual signal is inconsistent with the signs of the residual signals at previous consecutive time steps, the growth of the time accumulation is suppressed.

[0060] The aforementioned processing mechanism enables the time correction amount to stably respond to load change trends, avoiding unreasonable jumps in the predicted value over time due to short-term anomalies. Those skilled in the art will understand that residual direction consistency is equivalent to determining the trend of long-term deviations to suppress the influence of noise points.

[0061] Furthermore, to achieve synergy between spatial and temporal corrections, this embodiment introduces a spatial correction as an initial state when constructing the temporal accumulator, allowing the deviation compensation obtained in the spatial dimension to be naturally transferred to the temporal dimension. Specifically, when calculating the temporal accumulator, the spatial correction is used as the initial value of the temporal accumulator, and the update amplitude of the temporal accumulator is dynamically adjusted based on the directional consistency between the spatial correction and the current residual signal. This ensures that the correction direction formed in the spatial correction stage is maintained or weakened in the temporal dimension, constituting a cross-stage dynamic closed-loop adjustment process.

[0062] Through the above design, the spatial correction amount not only affects the prediction result at the current moment, but also provides directional guidance for the time correction mechanism at subsequent moments, realizing the natural integration and linkage control of the spatial and temporal dimensions.

[0063] To further avoid abrupt changes in the direction of the final prediction result over time, this embodiment also performs trend stability correction before outputting the final prediction value. Specifically, before outputting the final power load prediction result, the rate of change of the final power load prediction result at adjacent time points is detected; if the rate of change of the current prediction result is opposite to the trend change direction formed at previous time points, and the absolute value exceeds a preset change amplitude threshold, then a constraint correction process is performed on the final power load prediction result to ensure that the final power load prediction result maintains continuous trend response characteristics in the time dimension.

[0064] The aforementioned trend correction mechanism can effectively avoid unreasonable "reverse jumps" in the prediction results, making the prediction curve more consistent with the natural continuity of users' electricity consumption behavior, and facilitating stable and reliable decision-making by the power dispatching system.

[0065] Through the above steps, this invention achieves an online correction mechanism with controllable deviation, adaptive direction, and continuous trend in the time dimension, solving the problem that traditional load forecasting cannot continuously track and dynamically suppress time-series deviations, thus making the final forecast results more accurate and robust.

[0066] like Figure 2 As shown, in another embodiment of the present invention, the present invention also provides a system for executing the above-described power load forecasting method based on spatial and temporal collaborative correction. This system can be deployed in a power dispatching platform, a regional load forecasting device, an energy management system (EMS), or a cloud-based forecasting service to enable online operation of the forecasting method. Specifically, in another embodiment of the present invention, a power load forecasting system based on spatial and temporal collaborative correction is also provided, applied to the power load forecasting method based on spatial and temporal collaborative correction as described above, the system comprising:

[0067] (1) Basic prediction module

[0068] The basic forecasting module is used to acquire historical power load data, meteorological data, and calendar event data for the target time, and generate basic load forecast results for the target time through a rule-based forecasting model based on the historical power load data, meteorological data, and calendar event data.

[0069] In practical applications, the basic forecasting module can acquire relevant data from the power data platform, meteorological service system, and holiday information database through a data acquisition interface. Internally, this module can analyze the trend changes, seasonal cycle characteristics, temperature sensitivity, and holiday patterns of historical load sequences based on fixed rule logic, statistical characteristics, or simple trend models, thereby generating preliminary forecast results for the target time. These results serve as the input baseline for subsequent spatial and temporal correction models, thus providing a stable initial estimate for the forecasting process.

[0070] (2) Spatial Correction Module

[0071] The spatial correction module is used to acquire historical power load data of multiple associated load objects, and based on the historical power load data of the associated load objects and the basic load prediction results, correct the basic load prediction results through a spatial correlation learning model to generate a first corrected prediction result.

[0072] The spatial correction module can access a stored list of associated load objects and select load objects from the load object pool that share geographical proximity, similarity in load curve shape, or similarity in user type with the target load object. Internally, the module performs data preprocessing (including time alignment and normalization) to construct spatial feature vectors reflecting the behavioral patterns of associated load objects. After combining these with the base prediction results to form fused features, the spatial correction module uses a spatial correlation learning model to quantify the degree of association between the target load object and each associated object. As described in S2 above, this module uses a feature distance and similarity weight calculation method to obtain a spatial correction amount and applies it to the base prediction value, thereby generating a first corrected prediction result that reflects the consistency of spatial behavior.

[0073] The output of this module includes not only the predicted values ​​themselves, but also directional information for the time correction module, providing starting data for the subsequent construction of the time accumulation.

[0074] (3) Time correction module

[0075] The time correction module is used to acquire historical actual load data of the target load object, and based on the historical actual load data and the first corrected prediction result, to correct the first corrected prediction result online through a time correlation online learning model, so as to generate the final power load prediction result at the target time.

[0076] During operation, the time correction module continuously receives the latest actual load values ​​and compares them with the first corrected prediction result in each prediction period. Internally, the module constructs a residual time accumulation, suppresses the impact of abnormal reverse jumps on the accumulation term based on the residual direction consistency mechanism, and uses a nonlinear mapping function to smoothly convert the accumulated residuals into a time correction value. Subsequently, the module superimposes the time correction value with the first corrected prediction result to obtain the final prediction value.

[0077] In addition, the module re-evaluates the trend of the results before outputting the predicted values. When a strong reverse change in the predicted trend is detected and the magnitude exceeds the preset threshold, trend stability correction will be triggered to make the final predicted values ​​smoother and more continuous, avoiding peaks or reverse jumps from interfering with the scheduling strategy.

[0078] The aforementioned basic prediction module, spatial correction module, and time correction module form a complete prediction closed loop in the order of data flow. In actual use, the system can run periodically (e.g., triggered every 15 minutes, every 30 minutes, or every hour), or dynamically adjust as data updates in real-time mode. The entire system can be deployed on a single server, cloud computing node, or embedded device; all prediction logic of this invention can be completed simply by executing the corresponding program instructions through the processor.

[0079] Through the coordinated operation of these three modules, the prediction system of this invention can: provide a stable and reliable basic prediction baseline; comprehensively utilize the spatial behavior characteristics of other load objects to improve prediction accuracy; and achieve fine-grained correction in the time dimension through online residual learning and trend analysis. Ultimately, this results in more accurate predictions, greater adaptability, and good trend continuity.

[0080] Example 2

[0081] Based on the above embodiments, in order to enable those skilled in the art to more clearly understand the actual construction method of the rule-based prediction model, in some embodiments of the present invention, the rule-based prediction model can use one or more combinations of various prediction strategies to predict the base load based on the sensitivity of the load to temperature and holidays in different months. Several exemplary implementations are given below, but the present invention is not limited thereto.

[0082] In one implementation, the rule-based prediction model can use a similar temperature rolling method to calculate the baseline load prediction results for the target time. Specifically, at the beginning of the prediction month, historical power load data and corresponding meteorological data from the two months prior to the prediction month to the prediction date are selected, and the daily minimum temperature, maximum temperature, and actual daily electricity consumption corresponding to each temperature value within the temperature range are extracted. Based on the above data, the ratio of the difference in actual daily electricity consumption at both ends of the temperature range to the difference between the maximum and minimum temperatures is calculated, and this ratio is rounded to obtain a step difference value used to characterize the degree of influence of temperature changes on daily electricity consumption. Subsequently, the average temperature is calculated using effective temperature data from the most recent few days (e.g., the most recent five days), and the anchor value is calculated using actual daily electricity consumption from the most recent few days (e.g., the most recent five days). The anchor value is rounded and used as the baseline reference level. For the target prediction day, the difference between the predicted temperature and the average temperature is weighted by the step difference and added to the anchor value to obtain the baseline predicted daily electricity consumption for that prediction day. Throughout the month's operation, the above calculation process can be continuously updated with the daily addition of meteorological data and actual electricity consumption data, enabling the baseline prediction results to dynamically track temperature change trends. When there is insufficient valid data for the most recent few days, the historical data window can be extended forward until a sufficient sample is obtained; when the highest temperature and the lowest temperature are the same, making it impossible to calculate the step difference, the step difference can be set to zero or other basic prediction methods can be switched to ensure the feasibility and stability of the prediction process.

[0083] In another embodiment, the rule-based prediction model can use the average displacement method to generate the base load prediction result for the target time. Specifically, when making a base forecast for a certain prediction day, the actual daily electricity consumption data of the most recent several working days (e.g., the most recent five non-weekend and non-statutory holidays) before the prediction date are selected, and the arithmetic mean of these valid samples is calculated. The resulting average value is used as the base forecast result for that prediction day. If the number of valid samples is insufficient due to consecutive holidays or other reasons, the search range can be automatically expanded forward to supplement the required number of samples, or the similar temperature rolling method can be switched to complete the base forecast for that prediction day. For historical daily electricity consumption data containing obvious outliers, preprocessing methods such as median substitution, outlier removal, or smoothing can be used to improve the robustness of the base forecast result.

[0084] In another implementation, for periods of significant load fluctuations around the Spring Festival, the rule-based forecasting model can employ a special Spring Festival forecasting method to calculate the base load. Specifically, at the beginning of the forecasting month, the initial base forecast values ​​for each forecast day of the month are first obtained using the average displacement method. During the month's operation, the reference data is updated in a rolling manner, for example, using the most recent day's actual daily electricity consumption as the rolling reference baseline. When constructing the load change trend during the Spring Festival, a time interval corresponding to the Spring Festival in the previous year (e.g., several consecutive days before and after the Spring Festival) is selected as a reference window. The daily growth rate is calculated based on the actual daily electricity consumption of two adjacent days within the window, and the arithmetic mean of the daily growth rates within the window is calculated to obtain the ten-day growth rate used to characterize the load change trend during the holiday. During the forecasting process, the initial base forecast value can be used as a benchmark, and the forecast can be iteratively calculated day by day using the method of "the current forecast value plus the product of the benchmark value and the ten-day growth rate" to obtain a base forecast result with holiday inertia characteristics. When the ten-day growth rate is abnormally large, reverses its sign, or deviates significantly from historical holiday patterns, it can be corrected by setting an upper limit threshold, smoothing the data, or using the moving average growth rate of adjacent periods. If the data for the corresponding time interval of the previous year is missing, the growth rate of nearby years or similar periods in the recent period can be used as a substitute reference to ensure the feasibility of this method in real-world scenarios.

[0085] It should be noted that the aforementioned similar temperature rolling method, average displacement method, and Spring Festival special forecasting method, when used individually or in combination, are merely specific ways to implement the rule-based forecasting model. They are used to generate the basic load forecast results for the target time during the basic forecasting stage, providing an input baseline for subsequent spatial correlation learning models and temporal correlation online learning models. Those skilled in the art can make appropriate adjustments to the time window length, holiday division method, growth rate smoothing strategy, and abnormal data handling method according to different regions, seasons, and business needs. All equivalent substitution schemes based on the above basic ideas fall within the protection scope of this invention.

[0086] In summary, this invention demonstrates higher model robustness and practical deployment value in complex power operation scenarios. It can be widely used in key business processes such as power grid dispatch optimization, demand response strategy formulation, energy efficiency management, and load monitoring. It provides important technical support for building a high-precision, intelligent power system prediction and control system and has good engineering application prospects and promotion value.

[0087] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0088] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functionality involved.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A power load forecasting method based on spatial and temporal collaborative correction, characterized in that, The method comprises: Based on the historical power load data, meteorological data and calendar event data of the target time, a basic load prediction result of the target time is generated through a rule prediction model; Obtain historical power load data of a plurality of associated load objects, and based on the historical power load data of the associated load objects and the basic load prediction result, correct the basic load prediction result through a spatial correlation learning model to generate a first corrected prediction result; Obtain historical actual load data of the target load object, and based on the historical actual load data and the first corrected prediction result, correct the first corrected prediction result online through a time correlation online learning model to generate a final power load prediction result of the target time; The space correlation learning model generates a space correction amount based on a structured basic prediction feature and an associated load data set, a feature distance between a target load object and each associated load object and a similarity weight is calculated by the following equation: ； wherein: , represents a trend component of the target load object and the associated load object; , represents a temperature response component of the target load object and the associated load object; , represents a calendar event component of the target load object and the associated load object; , represents a trend component of the target load object and the associated load object; , represents a calendar event component of the target load object and the associated load object; , represents a calendar event component of the target load object and the associated load object; , represents a calendar event component of the target load object and the associated load object; , represents a calendar event component of the target load object and the associated load object; represents non-negative coefficients for adjusting relative weights of the three types of components; represents a weighted feature distance based on difference values of the three types of components; represents a similarity weight mapped from the feature distance. And based on the similarity weight, the residual signal of the associated load object is weighted to generate a spatial correction amount, and the spatial correction amount is applied to the basic load prediction result to obtain the first corrected prediction result.

2. The power load forecasting method based on spatial and temporal collaborative correction according to claim 1, characterized in that, The rule prediction model extracts trends and analyzes seasonality from the historical power load data of the target load object, establishes a temperature response relationship based on meteorological data, and corrects the influence of holidays in combination with calendar event data to obtain structured basic prediction features including trend components, temperature response components and calendar event components. The structured basic prediction features are provided to the spatial correlation learning model for collaborative correction processing of the basic load prediction result.

3. The power load forecasting method based on spatial and temporal collaborative correction according to claim 2, characterized in that, The execution process of the spatial correlation learning model comprises: constructing spatial correlation features based on the load data of a plurality of associated load objects; fusing the spatial correlation features and the structured basic prediction features to obtain fusion features; determining the spatial correlation weight between the target load object and each associated load object based on the fusion features; and based on the spatial correlation weight, the load features reflected by the load data of the associated load objects and the basic load prediction result are collaboratively calculated and processed to generate the first corrected prediction result.

4. The power load forecasting method based on spatial and temporal collaborative correction according to claim 3, characterized in that, The plurality of associated load objects are determined by the following method: from the load object pool, filter out load objects that have similarities in geographical location, load curve shape feature or user type with the target load object; perform time alignment and numerical normalization processing on the load data of each filtered associated load object to form the associated load data set required by the spatial correlation learning model.

5. The power load forecasting method based on spatial and temporal collaborative correction according to claim 1, characterized in that, The time correlation online learning model generates a time correction amount based on historical actual load data and the first corrected prediction result, and the time correction amount satisfies the following formula: satisfies the following formula: ； Wherein: represents the first correction prediction result of the target load object at time ; represents the final power load prediction result of the target load object at time ; represents a time accumulation quantity obtained by weighting and accumulating residual signals according to a preset attenuation coefficient; represents a time correction quantity obtained by nonlinear mapping of the time accumulation quantity; the time accumulation quantity is obtained by weighting and accumulating residual signals at each time according to a preset attenuation coefficient.

6. The power load forecasting method based on spatial and temporal collaborative correction according to claim 5, characterized in that, When constructing the time accumulation, the residual signals of each time are processed based on the continuity of the residual direction. When it is detected that the sign of the current residual signal is inconsistent with the signs of the residual signals of the previous at least two continuous times, the growth of the time accumulation is suppressed.

7. The power load forecasting method based on spatial and temporal collaborative correction according to claim 6, characterized in that, When constructing the time accumulation, the spatial correction amount is taken as the initial value of the time accumulation, and based on the direction consistency between the spatial correction amount and the current residual signal, the update amplitude of the time accumulation is dynamically adjusted, so that the correction direction formed in the spatial correction stage is maintained or weakened in the time dimension, constituting a dynamic closed-loop adjustment process across stages.

8. The power load forecasting method based on spatial and temporal collaborative correction according to claim 7, characterized in that, Before outputting the final power load prediction result, a change rate of the final power load prediction result at a neighboring time point is detected; if the change rate of the current prediction result is opposite to a trend change direction formed at a previous time point and an absolute value exceeds a preset change amplitude threshold, a constraint correction process is performed on the final power load prediction result, so that the final power load prediction result maintains a continuous trend response characteristic in a time dimension.

9. A power load forecasting system based on spatial and temporal collaborative correction, applied to the power load forecasting method based on spatial and temporal collaborative correction according to any one of claims 1-8, characterized in that, The system comprises: a basic prediction module configured to acquire historical power load data, meteorological data and calendar event data at a target time point, and generate a basic load prediction result at the target time point based on the historical power load data, the meteorological data and the calendar event data through a rule prediction model; a spatial correction module configured to acquire historical power load data of a plurality of associated load objects, and correct the basic load prediction result based on the historical power load data of the associated load objects and the basic load prediction result through a spatial correlation learning model to generate a first correction prediction result; a time correction module configured to acquire historical actual load data of a target load object, and correct the first correction prediction result based on the historical actual load data and the first correction prediction result through a time correlation online learning model to generate a final power load prediction result at the target time point.

Images