A short-term load forecasting method with strong robustness for power marketing business

Abstract

The application relates to a strong robustness short-term load prediction method for power marketing business, which comprises the following steps: obtaining historical power marketing business data of a target marketing area as sample data; generating an inertia prediction result based on total power consumption load data, and generating a disturbance prediction result based on marketing profile data, meteorological data, calendar event data and marketing business event data; determining a prediction scene according to corresponding data of a day to be predicted, and performing scene-based fusion on the inertia prediction result and the disturbance prediction result to obtain an original load prediction result; performing capacity constraint verification, non-negative constraint verification, ramp constraint verification and trend consistency verification on the original load prediction result, and keeping or modifying the original load prediction result according to a verification result to obtain a target load prediction result. The application can improve the accuracy, stability and business adaptability of short-term load prediction under the power marketing scene.

Description

Technical Field

[0001] This application relates to the field of power load forecasting technology, and in particular to a robust short-term load forecasting method for power marketing operations. Background Technology

[0002] With the advancement of power system reform and the construction of new power systems, electricity marketing is gradually expanding from traditional meter reading and collection management to spot market reporting, deviation assessment and control, demand response management, and refined energy services. For target marketing areas with a high proportion of industrial users, the total electricity load can directly reflect the actual electricity consumption level on the user side, and its forecast results will affect the formulation of electricity sales plans, market transaction reporting, and marketing operation risk control.

[0003] Existing short-term load forecasting methods mainly include time series forecasting methods and intelligent forecasting methods. Time series forecasting methods can depict load trends and cyclical patterns, but they are insufficient in responding to external disturbances such as holiday adjustments, extreme weather, time-of-use pricing, and demand response events. Intelligent forecasting methods can establish nonlinear relationships between multiple factors and load, but they are prone to abnormal fluctuations when marketing-side data has meter jumps, zero values, missing data, or noise interference, and the forecast results lack constraints from electricity marketing business rules.

[0004] Furthermore, existing methods typically employ uniform models or fixed parameters for forecasting, making it difficult to differentiate between stable meteorological scenarios, specific holiday scenarios, and environmentally sensitive scenarios. They also lack capacity constraints, non-negativity constraints, ramp constraints, and trend consistency verification mechanisms, leading to forecast results that may include capacity exceeding limits, abnormally low values, abrupt curve changes, or abnormal growth lacking business basis, thus affecting the reliability of marketing business decisions. Therefore, there is an urgent need for a robust short-term load forecasting method suitable for electricity marketing operations. Summary of the Invention

[0005] To overcome, to some extent, the problems in related technologies, such as the difficulty in taking into account both load inertia and external disturbance response in short-term load forecasting, insufficient adaptability to complex scenarios such as holidays and abnormal weather, and the lack of power marketing business logic verification and correction in the forecast results, this application provides a robust short-term load forecasting method for power marketing business.

[0006] The proposed solution is as follows:

[0007] A robust short-term load forecasting method for electricity marketing operations includes: Historical electricity marketing data of the target marketing area is obtained as sample data; the historical electricity marketing data includes at least: total electricity load data, marketing file data, meteorological data, calendar event data, and marketing event data; Based on the total electricity load data, load trend characteristics and load cycle characteristics are extracted to generate inertial prediction results; Based on the marketing archive data, meteorological data, calendar event data, and marketing business event data, the correspondence between external disturbance factors and total electricity load is identified, and disturbance prediction results are generated. Based on the meteorological data, calendar event data, and marketing business event data corresponding to the date to be predicted, the prediction scenario to which the date to be predicted belongs is determined; the prediction scenarios include stable weather scenarios, specific holiday scenarios, and environmentally sensitive scenarios. Based on the predicted scenario, the inertial prediction results and disturbance prediction results are fused in a scenario-based manner to generate the original load prediction results; wherein, in the stable meteorological scenario, the proportion of the inertial prediction results is increased; in the specific holiday scenario, holiday corrections are performed based on the lunar calendar alignment relationship and historical holiday load change characteristics; and in the environmentally sensitive scenario, meteorological sensitivity corrections are performed based on continuous meteorological impact characteristics. Based on the original load forecast results, perform business logic verification to adapt to the electricity marketing business rules; the business logic verification includes capacity constraint verification, non-negativity constraint verification, ramp constraint verification and trend consistency verification. The original load forecast result is maintained or modified based on the business logic verification result to obtain the target load forecast result.

[0008] Preferably, the method further includes: Obtain the actual total electricity load data corresponding to the day to be predicted, compare the target load prediction result with the actual total electricity load data, and generate prediction deviation data; Based on the prediction deviation data, determine the type and degree of deviation of the target load prediction results under different prediction scenarios; The actual total electricity load data, meteorological data corresponding to the forecast day, calendar event data, marketing business event data, and forecast deviation data are added to the sample data. Based on the type and degree of deviation, adjust the fusion parameters for scene-based fusion. Based on the correspondence between the business logic verification results and the prediction deviation data, adjust at least one of the business logic verification rules among capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification.

[0009] Preferably, historical electricity marketing data of the target marketing region is obtained as sample data, including: Obtain total electricity load data, marketing archive data, weather data, calendar event data, and marketing business event data for the target marketing region over historical periods; The total electricity load data, marketing archive data, meteorological data, calendar event data, and marketing business event data are time-aligned according to the same time granularity to obtain aligned historical electricity marketing business data. Anomaly identification is performed on the total electricity load data in the aligned historical electricity marketing business data to determine meter tripping data and zero value data; A density-based clustering outlier identification algorithm is used to identify load outliers corresponding to the meter jump data and zero value data; Based on the adjacent historical time period data and historical data of the same period corresponding to the load outlier, the weighted average of historical data is calculated, and the load outlier is filled in using the weighted average of historical data to obtain the sample data.

[0010] Preferably, generating inertial prediction results includes: The total electricity load data is sorted according to a preset time granularity to generate a historical load sequence; The historical load sequence is input into a pre-built inertial prediction model; the inertial prediction model is configured with adaptive smoothing parameters and includes a baseline load level, a load trend term, and a load seasonal term. The basic electricity consumption level of the target marketing area is determined based on the benchmark load level, and load trend characteristics are extracted based on the load trend item and load cycle characteristics are extracted based on the load seasonal item. Based on the prediction step size and cycle length, the baseline load level, load trend term, and load seasonal term are combined to generate an initial inertial prediction result; Based on the deviation between historical prediction results and historical actual total electricity load data, cumulative error data and average absolute deviation data are generated, and an adaptive tracking signal is generated based on the cumulative error data and average absolute deviation data. When the adaptive tracking signal exceeds the preset offset threshold, the adaptive smoothing parameter is increased to increase the weight corresponding to the recent total electricity load data; when the adaptive tracking signal does not exceed the preset offset threshold, the adaptive smoothing parameter is decreased to decrease the weight corresponding to the recent total electricity load data. The initial inertial prediction result is updated based on the adjusted adaptive smoothing parameters to generate the inertial prediction result.

[0011] Preferably, generating disturbance prediction results includes: Time-of-use electricity rates and demand response event identifiers are extracted from the marketing archive data and marketing business event data as marketing features; Temperature and rainfall information are extracted from the meteorological data as meteorological features, and continuous temperature impact features and human comfort features are generated as meteorological derivative features based on the temperature and rainfall information. Extract time period attribute information from the calendar event data, and perform One-Hot encoding, sine embedding, or cosine embedding on the time period attribute information to generate time features; The marketing characteristics, meteorological characteristics, meteorological derivative characteristics, and time characteristics are combined into external disturbance characteristics; The external disturbance characteristics and total electricity load data are input into a pre-built disturbance prediction model, and the nonlinear correspondence between the external disturbance characteristics and the total electricity load is identified through the disturbance prediction model. The disturbance prediction model includes a temporal convolutional processing layer and a meteorological sensitivity attention processing layer. The temporal convolutional processing layer is used to extract long-term correlation features between the external disturbance features and the total electricity load data. The meteorological sensitivity attention processing layer is used to adjust the influence weights of different meteorological features according to the degree of influence of meteorological factors on the total electricity load in different seasons. Based on the nonlinear correspondence, the disturbances to the total electricity load caused by weather changes, time-of-use electricity price changes, and demand response events are identified, and a load disturbance component relative to the inertial prediction result is generated based on the disturbances; the load disturbance component is used to represent the increase or decrease relative to the inertial prediction result. The disturbance prediction result is generated based on the load disturbance component.

[0012] Preferably, determining the prediction scenario to which the date to be predicted belongs includes: Based on the calendar event data, identify the Gregorian calendar date, lunar calendar date, holiday attributes, and work-rest adjustment attributes of the day to be predicted; If the holiday attribute indicates that the date to be predicted belongs to a holiday or the period before or after a holiday, then the prediction scenario to which the date to be predicted belongs is determined to be a specific holiday scenario. If the predicted day does not belong to the specific holiday scenario, then the meteorological data is used to determine whether the predicted day meets the preset abnormal meteorological conditions, and if it meets the preset abnormal meteorological conditions, the predicted scenario to which the predicted day belongs is determined to be an environmentally sensitive scenario. If the predicted day does not belong to the specific holiday scenario and does not meet the preset abnormal weather conditions, then the marketing business event data is used to identify whether there is a demand response event or marketing activity period on the predicted day; If the demand response event or marketing activity period does not exist on the day to be predicted, then the prediction scenario to which the day to be predicted belongs is determined to be a stable weather scenario. If the demand response event or marketing activity period exists on the date to be predicted, then the demand response event or marketing activity period is identified as an event correction marker. The event correction marker is used to correct the degree of impact of the disturbance prediction result in the contextual fusion.

[0013] Preferably, generating the original load forecast results includes: The corresponding prediction channel is invoked according to the predicted scenario, and the corresponding scenario-based fusion parameters are obtained; When the predicted scenario is a stable weather scenario, the proportion of the inertial prediction result in the scenario fusion is increased, and the disturbance prediction result is used for auxiliary correction to generate the original load prediction result. When the predicted scenario is a specific holiday scenario, historical holiday load data for the same period are extracted according to the lunar calendar alignment relationship, and the pre-holiday baseline load is determined according to the total electricity load data of normal working days before the holiday. Based on the variation of the historical holiday load data relative to the pre-holiday baseline load, historical holiday load variation characteristics are generated, and the scenario-based fusion results are corrected for holidays using these historical holiday load variation characteristics to generate the original load prediction results. When the prediction scenario is an environmentally sensitive scenario, the temperature forecast value for the day to be predicted and the measured temperature values ​​for consecutive historical periods before the day to be predicted are obtained. After assigning different weights to the temperature forecast value for the day to be predicted and the measured temperature values ​​for consecutive historical periods before the day to be predicted, the cumulative effect temperature characteristics are calculated. Among them, the weight corresponding to the temperature forecast value is greater than the weight corresponding to the measured temperature value, and the weight corresponding to the measured temperature value that is closer to the day to be predicted in the consecutive historical periods is greater. Based on the cumulative effect temperature characteristics, the proportion of the disturbance prediction results in the scenario fusion is increased, and the scenario fusion results are subject to meteorological sensitivity correction to generate the original load prediction results.

[0014] Preferably, based on the original load forecast results, business logic verification for adapting electricity marketing business rules is performed, including: Based on the application capacity information and transformer capacity information in the marketing file data, the original load forecast result is checked for capacity constraints to determine whether the original load forecast result exceeds the allowable capacity range. Based on the distributed power access status and historical minimum total electricity load of the target marketing area, the original load prediction results are subjected to non-negative constraint verification to determine whether the original load prediction results have negative values ​​or abnormally low values. Based on the load variation range of adjacent time periods, the allowable variation range of equipment, and the historical load fluctuation range, the original load forecast results are subjected to ramp constraint verification to determine whether there are any abnormal sudden changes in the original load forecast results. Based on the relationship between the original load forecast results and the historical total electricity load data for the same period, and in conjunction with new capacity expansion records, abnormal weather information and business event information, the original load forecast results are checked for trend consistency to determine whether there is any abnormal growth or abnormal decline in the original load forecast results due to lack of business support. Based on the results of the capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification, a business logic verification result is generated.

[0015] Preferably, the original load forecast result is maintained or corrected based on the business logic verification result to obtain the target load forecast result, including: When the business logic verification result indicates that the original load forecast result passes the capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification, the original load forecast result is taken as the target load forecast result. When the business logic verification result indicates that the original load prediction result fails the capacity constraint verification, the prediction value that exceeds the capacity allowable range will be corrected to the capacity allowable range. When the business logic verification result indicates that the original load prediction result fails the non-negative constraint verification, the prediction value corresponding to the negative value or abnormally low value period is replaced and corrected according to the inertial prediction result. When the business logic verification result indicates that the original load prediction result fails the ramp constraint verification, the prediction curve segment with abnormal abrupt changes is smoothed and corrected. When the business logic verification result indicates that the original load forecast result fails the trend consistency verification, the abnormal trend is reviewed based on the newly added expansion records, abnormal meteorological information and business event information, and the forecast value for the corresponding period is maintained or corrected based on the review result to obtain the target load forecast result. When the business logic verification result triggers a high-risk warning condition, the automatic release of the target load forecast result is suspended, and the corresponding original load forecast result, business logic verification result, and correction result are sent to the manual review end so that the target load forecast result can be confirmed or adjusted according to the manual review result.

[0016] Preferably, the target load prediction result is compared with the actual total electricity load data to generate prediction deviation data, including: Based on the prediction time granularity, the target load prediction results and the actual total electricity load data are divided into multiple corresponding evaluation periods; The time weight of each evaluation period is determined based on the electricity price period attribute or marketing value attribute corresponding to each evaluation period; among them, the time weight corresponding to the peak period is greater than the time weight corresponding to the normal period, and the time weight corresponding to the normal period is greater than the time weight corresponding to the valley period. Calculate the absolute deviation of the load forecast between the target load forecast result and the actual total electricity load data for each evaluation period; Based on the time weight of each evaluation period, the absolute deviation of load forecast, and the actual total electricity load data, a weighted accuracy evaluation result is generated. The prediction deviation data is generated based on the weighted accuracy evaluation results and the absolute deviation of load prediction for each evaluation period.

[0017] The technical solution provided in this application may include the following beneficial effects: This technical solution acquires total electricity load data, marketing record data, meteorological data, calendar event data, and marketing business event data for the target marketing area. It integrates the load's own variation patterns with external factors related to electricity marketing into the forecasting process, avoiding insufficient forecast adaptability caused by relying solely on a single load sequence or meteorological factor. By generating inertial and disturbance forecasts separately, and fusing them according to stable meteorological scenarios, specific holiday scenarios, and environmentally sensitive scenarios, the forecasting process maintains the stability of normal electricity consumption changes while allowing for differentiated corrections for holiday load changes, abnormal weather impacts, and marketing business events. This improves the accuracy and robustness of short-term load forecasting in complex marketing scenarios. Simultaneously, by performing capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification on the original load forecast results, and maintaining or correcting the forecast results based on the verification results, it reduces capacity overruns, abnormally low values, load abrupt changes, and abnormal trends lacking business support. This improves the consistency between the target load forecast results and electricity marketing business rules, enhancing the reliability of the forecast results in electricity sales planning, spot market declaration, demand response arrangements, and deviation risk control.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 This is a flowchart illustrating a robust short-term load forecasting method for electricity marketing operations, provided in one embodiment of this application. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] Example 1 Figure 1 This is a flowchart illustrating a robust short-term load forecasting method for electricity marketing operations, provided in one embodiment of this application. (Refer to...) Figure 1 A robust short-term load forecasting method for electricity marketing operations includes: S11. Obtain historical electricity marketing data for the target marketing area as sample data; historical electricity marketing data shall include at least: total electricity load data, marketing record data, meteorological data, calendar event data, and marketing event data; S12. Based on the total electricity load data, extract the load trend characteristics and load cycle characteristics to generate inertial prediction results; S13. Based on marketing archive data, meteorological data, calendar event data, and marketing business event data, identify the correspondence between external disturbance factors and total electricity load, and generate disturbance prediction results; S14. Based on the meteorological data, calendar event data, and marketing business event data corresponding to the day to be predicted, determine the prediction scenario to which the day to be predicted belongs; prediction scenarios include stable weather scenarios, specific holiday scenarios, and environmentally sensitive scenarios; S15. Based on the prediction scenario, the inertial prediction results and disturbance prediction results are fused in a scenario-based manner to generate the original load prediction results. Among them, the proportion of inertial prediction results is increased in the meteorological stable scenario, holiday correction is performed based on the lunar calendar alignment relationship and historical holiday load change characteristics in the specific holiday scenario, and meteorological sensitivity correction is performed based on the continuous meteorological impact characteristics in the environmentally sensitive scenario. S16. Based on the original load forecast results, perform business logic verification of the adapted electricity marketing business rules; business logic verification includes capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification. S17. Based on the business logic verification results, maintain or correct the original load forecast results to obtain the target load forecast results.

[0023] For ease of understanding, the following explains some key terms in this embodiment: The target marketing area refers to the specific geographical area or user group in which a power company conducts its electricity marketing business. Its electricity consumption behavior and load characteristics are the focus of predictive analysis.

[0024] Historical electricity marketing data is a collection of historical records used to train and validate predictive models, and its comprehensiveness and accuracy directly affect the predictive results.

[0025] The sample data consists of preprocessed and filtered historical electricity marketing data, which is used to build and optimize the predictive model.

[0026] Total electricity load data reflects the total electricity consumption of the target marketing area at different points in time and is the core objective of load forecasting.

[0027] Marketing profile data includes basic user information, electricity usage characteristics, and installed capacity, providing a basis for identifying structural impacts on the user side.

[0028] Meteorological data, including meteorological elements such as temperature, humidity, and rainfall, is an important external factor affecting fluctuations in electricity load.

[0029] Calendar event data covers time attribute information such as Gregorian calendar, lunar calendar, holidays, and work schedule adjustments, which has a significant impact on electricity consumption patterns.

[0030] Marketing business event data records business-level events such as time-of-use electricity price adjustments, demand response activities, and marketing promotions, which may cause short-term or localized load fluctuations.

[0031] Load trend characteristics describe the long-term trend of total electricity load over time, such as growth or decline.

[0032] The load cycle characteristics reflect the repetitive variation pattern of total electricity load at different time scales (such as day, week, year).

[0033] Inertial prediction results are forecast values ​​generated based on the trends and periodic patterns of historical loads, reflecting the inherent stability of the load.

[0034] External disturbance factors refer to external variables that affect the total electricity load, other than the inherent inertial laws of the load itself, such as weather changes and marketing policies.

[0035] Disturbance prediction results are predicted values ​​that quantify the impact of external disturbance factors on the total electricity load, and are usually expressed as increases or decreases relative to inertial prediction results.

[0036] The date to be predicted refers to a future date on which load forecasting is required.

[0037] Forecast scenarios are classifications of the environment and business characteristics of the forecast date in order to adopt differentiated forecasting strategies.

[0038] A stable weather scenario refers to a normal power consumption environment where the weather conditions on the forecast date are normal and there are no major holidays or special marketing events.

[0039] Specific holiday scenarios refer to the period before or after a statutory holiday when the predicted date falls on or is affected by a holiday.

[0040] Environmentally sensitive scenarios refer to power consumption environments that are affected by abnormal meteorological conditions (such as high temperatures or cold waves) or major environmental events on the predicted date.

[0041] Scenario-based fusion is a process of weighting and combining or correcting inertial prediction results and disturbance prediction results based on the prediction scenario to which the date to be predicted belongs.

[0042] The original load forecast results are preliminary forecasts obtained after scenario-based fusion.

[0043] Business logic verification is the process of checking the original load forecast results to ensure they comply with electricity marketing business rules.

[0044] Capacity constraint verification checks whether the predicted load exceeds the installed capacity limit of the user or transformer.

[0045] Non-negative constraint verification checks whether the predicted load has negative values ​​or abnormally low electricity consumption.

[0046] The ramp constraint check is to check whether the load change range between adjacent time periods is within the reasonable range of equipment operation or load change.

[0047] Trend consistency verification checks whether the long-term trend of the forecast load is consistent with historical data from the same period and business development plans.

[0048] The target load forecast result is the final load forecast value that meets the business requirements after business logic verification and correction.

[0049] This application provides a robust short-term load forecasting method for electricity marketing operations.

[0050] First, to obtain historical electricity marketing data for the target marketing area as sample data, this historical electricity marketing data should include at least total electricity load data, marketing record data, meteorological data, calendar event data, and marketing event data. Specifically, this data can be collected from different data sources. For example, total electricity load data can be obtained from the electricity metering system, marketing record data from the customer relationship management system or marketing business system, meteorological data from a third-party meteorological service platform, calendar event data from public calendar services or manually maintained holiday tables, and marketing event data from the marketing activity management system or demand response platform. This data can be stored according to its respective recording method, such as in different file formats or database tables. During the acquisition process, methods such as periodic export, manual entry, or batch synchronization via an interface can be used.

[0051] Furthermore, based on the total electricity load data, load trend characteristics and load cycle characteristics are extracted to generate inertial prediction results. For example, load trend characteristics can be obtained by performing linear regression or polynomial fitting on historical load data to reflect the long-term growth or decline trend of the load. Load cycle characteristics can be identified through Fourier transform, wavelet analysis, or simple periodic averaging methods to capture the periodic changes of the load on a daily, weekly, monthly, or yearly basis. These extracted characteristics can then be input into a basic prediction model, such as a simple exponential smoothing model or an autoregressive moving average model, to generate preliminary inertial prediction results.

[0052] Simultaneously, based on marketing archive data, meteorological data, calendar event data, and marketing business event data, the correlation between external disturbance factors and total electricity load is identified, generating disturbance prediction results. For example, correlation analysis can be used to calculate the correlation coefficient between external factors such as temperature, humidity, holiday type, and time-of-use pricing and total electricity load. Alternatively, a simple machine learning model, such as a support vector machine or decision tree model, can be constructed, using these external factors as input and total electricity load as output, to learn the mapping relationship between them. In this way, the potential impact of different external factors on the load can be quantified, and based on this, a load disturbance component relative to the inertial prediction result can be generated, thus obtaining the disturbance prediction result.

[0053] Based on this, the forecast scenario for the predicted day is determined according to the meteorological data, calendar event data, and marketing event data corresponding to that day. These forecast scenarios include stable weather scenarios, specific holiday scenarios, and environmentally sensitive scenarios. Specifically, determining the forecast scenario for a predicted day can be based on its meteorological data, calendar event data, and marketing event data. For example, calendar event data can be checked first; if the predicted day is a statutory holiday, it is classified as a specific holiday scenario. If it is not a holiday, meteorological data is further checked, such as whether the temperature exceeds a fixed threshold of the historical normal range; if it does, it is classified as an environmentally sensitive scenario. If none of the above conditions are met, it is classified as a stable weather scenario. This classification can be implemented using a series of preset rules or a simple logical decision tree.

[0054] Subsequently, based on the predicted scenario, the inertial prediction results and disturbance prediction results are fused in a scenario-based manner to generate the original load prediction results. Specifically, in the stable meteorological scenario, the proportion of inertial prediction results is increased; in the specific holiday scenario, holiday corrections are made based on lunar calendar alignment and historical holiday load variation characteristics; and in the environmentally sensitive scenario, meteorological sensitivity corrections are made based on continuous meteorological impact characteristics. In detail, during the scenario-based fusion stage, the inertial and disturbance prediction results can be weighted and combined according to the determined prediction scenario. For example, in the stable meteorological scenario, the inertial prediction results can be given a higher weight, while the disturbance prediction results can be given a relatively lower weight to reflect load stability. In the specific holiday scenario, a fixed correction factor based on historical holiday load variations can be introduced to adjust the fused results. In the environmentally sensitive scenario, the disturbance prediction results can be amplified or reduced based on empirical values ​​of the impact of meteorological factors on the load, and then combined with the inertial prediction results. This fusion can be achieved through simple linear weighted summation or a piecewise function.

[0055] Furthermore, based on the original load forecast results, business logic checks adapted to electricity marketing business rules are performed. These checks include capacity constraint checks, non-negativity constraint checks, ramp constraint checks, and trend consistency checks. Specifically, the original load forecast results can be examined in multiple ways during the business logic checks. For example, capacity constraint checks can simply compare the forecast load with a preset maximum capacity threshold to determine if it exceeds the limit. Non-negativity constraint checks can check if the forecast load has a value less than zero. Ramp constraint checks can calculate the load change rate between adjacent time periods and compare it with a fixed maximum allowable change rate. Trend consistency checks can compare the overall trend of the forecast load with the average trend of the historical load for the same period, for example, by using a simple percentage deviation. These checks can be based on preset fixed rules or empirical thresholds.

[0056] Finally, the original load forecast results are maintained or corrected based on the business logic verification results to obtain the target load forecast results. Specifically, the original load forecast results can be processed accordingly based on the business logic verification results. If the original load forecast results pass all verifications, they can be directly used as the target load forecast results. If the capacity constraint verification fails, forecast values ​​exceeding the capacity limit can be truncated to the capacity limit. If the non-negativity constraint verification fails, negative values ​​or abnormally low values ​​can be simply replaced with zero or a preset minimum value. If the ramp constraint verification fails, abnormally abrupt curve segments can be smoothed using simple linear interpolation. If the trend consistency verification fails, forecast values ​​with abnormal trends can be manually reviewed or simply adjusted to the historical average level to obtain the final target load forecast results.

[0057] This technical solution acquires total electricity load data, marketing record data, meteorological data, calendar event data, and marketing business event data for the target marketing area. It integrates the load's own variation patterns with external factors related to electricity marketing into the forecasting process, avoiding insufficient forecast adaptability caused by relying solely on a single load sequence or meteorological factor. By generating inertial and disturbance forecasts separately, and fusing them according to stable meteorological scenarios, specific holiday scenarios, and environmentally sensitive scenarios, the forecasting process maintains the stability of normal electricity consumption changes while allowing for differentiated corrections for holiday load changes, abnormal weather impacts, and marketing business events. This improves the accuracy and robustness of short-term load forecasting in complex marketing scenarios. Simultaneously, by performing capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification on the original load forecast results, and maintaining or correcting the forecast results based on the verification results, it reduces capacity overruns, abnormally low values, load abrupt changes, and abnormal trends lacking business support. This improves the consistency between the target load forecast results and electricity marketing business rules, enhancing the reliability of the forecast results in electricity sales planning, spot market declaration, demand response arrangements, and deviation risk control.

[0058] Example 2 In some embodiments, the method further includes: Obtain the actual total electricity load data corresponding to the day to be predicted, compare the target load prediction result with the actual total electricity load data, and generate prediction deviation data; Based on the prediction deviation data, determine the type and degree of deviation of the target load prediction results under different prediction scenarios; The actual total electricity load data, meteorological data corresponding to the forecast date, calendar event data, marketing business event data, and forecast deviation data were added to the sample data. Adjust the fusion parameters for scenario-based fusion based on the type and degree of deviation; Based on the correspondence between the business logic verification results and the prediction deviation data, adjust at least one of the business logic verification rules among capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification.

[0059] Specifically, after the predicted day ends, the system acquires the actual total electricity load data for that day. Then, it compares the previously generated target load forecast with this actual total electricity load data on a time-period or day-by-day basis, calculating the differences between the two to generate prediction deviation data. Prediction deviation data can include various forms such as absolute error, relative error, and mean square error, used to quantify the degree of deviation between the prediction result and the actual situation. This process forms the basis for subsequent adaptive adjustments, ensuring that the system can learn and improve from actual operation.

[0060] Furthermore, forecast bias data not only needs to be recorded but also analyzed. This step involves in-depth analysis of the forecast bias data to identify the inherent patterns of the bias. For example, it can be analyzed whether forecasts are generally higher or lower under stable weather conditions; whether forecast biases exhibit a specific pattern under specific holiday scenarios; and whether forecast biases are strongly correlated with certain meteorological factors under environmentally sensitive scenarios. Bias types can include systematic biases, random biases, and periodic biases, while the degree of bias quantifies their magnitude. This meticulous analysis provides precise guidance for subsequent parameter adjustments.

[0061] To enable continuous learning and optimization of the model, newly acquired actual data and analyzed deviation information need to be integrated into historical sample data. Specifically, after the forecast date, the corresponding actual total electricity load data, meteorological data, calendar event data, marketing business event data, and forecast deviation data for that day will all be added as new valid samples to the historical electricity marketing business data sample set used for model training and parameter optimization. This continuous data replenishment mechanism ensures the timeliness and completeness of the sample data, enabling the forecasting model to continuously learn from the latest actual operating conditions, thereby improving its adaptability to future changes.

[0062] Building upon this foundation, scenario-based fusion is a core component of this method, with its fusion parameters directly influencing the weighting of inertial and perturbation prediction results. When analysis reveals systematic biases in specific prediction scenarios—for example, an excessively high proportion of inertial prediction results leading to prediction deviations in stable weather scenarios, or insufficient correction of perturbation prediction results in environmentally sensitive scenarios—adjustments to the corresponding fusion parameters are necessary based on the type and severity of the bias. For instance, if predictions are generally lower than expected during specific holiday scenarios, the weight of holiday corrections can be appropriately increased or the correction factor adjusted; conversely, if weather-sensitive corrections are insufficient in environmentally sensitive scenarios, the proportion of perturbation prediction results in the fusion process can be increased. This adaptive parameter adjustment mechanism allows the scenario-based fusion process to be dynamically optimized to adapt to constantly changing external environments and business needs.

[0063] Meanwhile, business logic verification is a crucial step in ensuring that forecast results conform to actual business rules. This step aims to dynamically optimize verification rules through correlation analysis of historical verification results and forecast deviation data. For example, if capacity constraint verification is frequently triggered in a certain area, but the actual electricity load has not truly exceeded the capacity, and forecast deviation data shows that forecast values ​​are generally too high, it may be necessary to reassess or adjust the allowable capacity range or verification threshold for that area. Similarly, if ramp constraint verification frequently misjudges, or trend consistency verification fails to effectively identify actual abnormal load changes, and forecast deviation data confirms these problems, then the allowable range of ramp constraints and the judgment criteria for trend consistency verification can be finely adjusted based on these correspondences. This adaptive adjustment of rules helps improve the accuracy and flexibility of business logic verification, avoiding over-correction or under-correction, thus making the final target load forecast results closer to actual business needs.

[0064] Example 3 In some embodiments, historical electricity marketing data of the target marketing area is obtained as sample data, including: Obtain total electricity load data, marketing archive data, weather data, calendar event data, and marketing business event data for the target marketing region over historical periods; The total electricity load data, marketing archive data, meteorological data, calendar event data, and marketing business event data are aligned with the same time granularity to obtain aligned historical electricity marketing business data. Anomalies are identified in the total electricity load data in the aligned historical electricity marketing business data to determine meter tripping data and zero value data. A density-based clustering outlier identification algorithm is used to identify load outliers corresponding to meter jump data and zero value data; Based on the adjacent historical time period data and historical data of the same period corresponding to the load outlier, the weighted average of historical data is calculated, and the load outlier is filled in using the weighted average of historical data to obtain sample data.

[0065] Specifically, the first step is to acquire historical data on total electricity load, marketing records, weather, calendar events, and marketing business events for the target marketing region. This data forms the foundation for building an electricity load forecasting model, comprehensively reflecting various factors influencing electricity load changes, including actual electricity consumption, user attributes, environmental conditions, time periods, and electricity marketing activities. To ensure effective integration and analysis of this multi-source, heterogeneous data, all data needs to be time-aligned at a predetermined time granularity (e.g., every hour, every 15 minutes, or every 30 minutes). This alignment process precisely maps data points from different time series onto a unified time axis, forming a complete and time-consistent historical electricity marketing business data set, providing reliable and structured input for subsequent data analysis and model training.

[0066] Building upon this foundation, identifying anomalies in the total electricity load data within the aligned historical electricity marketing data is a crucial step. Anomalies typically manifest as meter jumps and zero-value data. Meter jumps refer to sudden, non-physical spikes or drops in load within a very short period, often caused by metering equipment malfunctions, communication errors, or abnormal data transmission. Zero-value data refers to instances where the load is recorded as zero during normal electricity consumption periods, which may stem from equipment outages, data acquisition interruptions, or system errors. These anomalies do not represent actual electricity consumption behavior and, if left unaddressed, will severely mislead the predictive model's learning process.

[0067] To accurately identify these anomalous data, this application employs a density-based clustering outlier detection algorithm to identify load outliers corresponding to meter fluctuations and zero-value data. For example, algorithms such as DBSCAN (Density-Based Spatial Clustering of Applications with Noise) or OPTICS (Ordering Points to Identify the Clustering Structure) can be used. These algorithms analyze the density distribution of data points to discover clusters in the data and mark data points with densities below a preset threshold or that do not belong to any high-density clusters as outliers. Compared to simple threshold judgment or distance-based outlier detection methods, density-based algorithms exhibit stronger robustness and accuracy when handling non-spherical clusters, datasets with varying densities, and complex anomaly patterns, effectively distinguishing normal load fluctuations from genuine anomalous outliers.

[0068] After identifying load outliers, these outliers need to be appropriately filled in to restore the integrity and accuracy of the data. Specifically, this application calculates a weighted average of historical data based on adjacent historical time period data and historical data from the same period last year, corresponding to the load outlier. Adjacent historical time period data can include load data from one hour before, one hour after, or several time points before the outlier, while historical data from the same period can refer to load data with similar periodic characteristics, such as the same day and time last year, or the same day and time last week. When calculating the weighted average, different weights can be assigned based on factors such as the time proximity and similarity between the data and the outlier. For example, adjacent historical data closer to the outlier and more similar historical data from the same period have greater weights. Finally, the calculated weighted average of historical data is used to fill in the load outliers, thereby obtaining high-quality sample data.

[0069] Example 4 In some embodiments, generating inertial prediction results includes: The total electricity load data is sorted according to a preset time granularity to generate a historical load sequence; Historical load sequences are input into a pre-built inertial forecasting model; the inertial forecasting model is configured with adaptive smoothing parameters and includes a baseline load level, a load trend term, and a load seasonal term. Determine the basic electricity consumption level of the target marketing area based on the baseline load level, extract load trend characteristics based on the load trend item, and extract load cycle characteristics based on the load seasonal item; Based on the prediction step size and cycle length, the baseline load level, load trend term, and load seasonal term are combined to generate the initial inertial prediction result; Based on the deviation between historical forecast results and historical actual total electricity load data, cumulative error data and mean absolute deviation data are generated, and an adaptive tracking signal is generated based on the cumulative error data and mean absolute deviation data. When the adaptive tracking signal exceeds the preset offset threshold, the adaptive smoothing parameter is increased to increase the weight corresponding to the recent total electricity load data; when the adaptive tracking signal does not exceed the preset offset threshold, the adaptive smoothing parameter is decreased to decrease the weight corresponding to the recent total electricity load data. The initial inertial prediction results are updated based on the adjusted adaptive smoothing parameters to generate the inertial prediction results.

[0070] First, to effectively utilize historical data, the total electricity load data needs to be sorted according to a preset time granularity to generate a historical load sequence. This preset time granularity can be set to an hour, half-hour, or finer granularity according to actual needs, ensuring the temporal sequence and continuity of the data, laying the foundation for subsequent analysis. For example, if the hour is chosen as the time granularity, the total electricity load data for each historical 24-hour period is arranged in chronological order to form a continuous load sequence. Subsequently, this historical load sequence is input into a pre-built inertial prediction model. This inertial prediction model is the core of this method; it is configured with adaptive smoothing parameters and includes a baseline load level, a load trend term, and a load seasonal term. The baseline load level is used to determine the basic electricity consumption level of the target marketing area, for example, by averaging or calculating the median from long-term historical load data. The load trend term is used to extract long-term trend characteristics of the load, for example, by identifying the direction of load growth or decline through linear regression or moving average. The load seasonal term is used to capture the periodic variation characteristics of the load, for example, by identifying daily, weekly, or monthly periodic patterns through Fourier transform or seasonal decomposition. These components together constitute a comprehensive characterization of the load pattern.

[0071] Let the load sequence be {yt}, and the model contain three components: a horizontal term Lt, a trend term Tt, and a seasonal term St. Predict the load value at the h-th time in the future from the current time t. : ; Wherein, Lt is the level term, representing the baseline load level at the current moment after eliminating seasonal effects; h is the forecast step size, for example, if the time granularity is 1 hour, then h=24 for forecasting the same time tomorrow; Tt is the trend term, representing the slope of load change over time (e.g., a natural increase of 5kW per day); St-m+h is the seasonal term, representing periodic fluctuations (e.g., the load at 2 PM is usually higher than at 8 AM). m is the period length, which is usually m=24 (hours) or m=96 (15-minute frequency) for daily load curves.

[0072] After extracting the baseline load level, load trend term, and seasonal load term, these components are combined according to the prediction step size and cycle length to generate an initial inertial prediction result. The prediction step size defines the prediction time range, such as the next 24 hours, while the cycle length matches the periodicity of the seasonal term. The combination method can be an additive or multiplicative model to form a preliminary estimate of future load. To continuously evaluate and optimize the model's performance, this method generates cumulative error data and mean absolute deviation data based on the deviation between historical prediction results and historical actual total electricity load data. The cumulative error data reflects the systematic deviation of the prediction, while the mean absolute deviation data quantifies the average magnitude of the prediction error. Based on these error data, an adaptive tracking signal is further generated, which serves as a key indicator for measuring the model's prediction stability and the degree of deviation.

[0073] Unlike traditional fixed parameters, this embodiment introduces an adaptive tracking signal (TS).

[0074] ; Where TSt is the tracking signal value, used to determine whether the model is out of sync. Et is the cumulative error, and MADt is the mean absolute deviation.

[0075] When the adaptive tracking signal TSt exceeds the preset offset threshold, it indicates a large or persistent deviation between the model prediction and the actual load. In this case, the model needs to respond to the latest load changes more quickly. Therefore, this method increases the adaptive smoothing parameter. This increases the weight of recent total electricity load data in model updates, enabling the model to learn and adapt to new load patterns more quickly. Conversely, when the adaptive tracking signal does not exceed a preset offset threshold, it indicates that the model prediction is relatively stable, and the adaptive smoothing parameter will be reduced. To reduce the weight of recent total electricity load data, the model is made smoother, reducing overreaction to short-term fluctuations and improving prediction stability. Finally, the initial inertial prediction results are updated based on the adjusted adaptive smoothing parameters to generate the final inertial prediction results.

[0076] In some embodiments, generating perturbation prediction results includes: Time-of-use electricity rates and demand response event identifiers are extracted from marketing archive data and marketing business event data as marketing features; Temperature and rainfall information are extracted from meteorological data as meteorological features, and continuous temperature impact features and human comfort features are generated as meteorological derivative features based on temperature and rainfall information. Extract time period attribute information from calendar event data, and perform One-Hot encoding, sine embedding, or cosine embedding on the time period attribute information to generate time features; Marketing characteristics, meteorological characteristics, meteorological derivative characteristics, and time characteristics are combined into external disturbance characteristics; External disturbance characteristics and total electricity load data are input into a pre-built disturbance prediction model, and the nonlinear correspondence between external disturbance characteristics and total electricity load is identified through the disturbance prediction model. The disturbance prediction model includes a temporal convolutional processing layer and a meteorological sensitivity attention processing layer. The temporal convolutional processing layer is used to extract long-term correlation features between external disturbance features and total electricity load data. The meteorological sensitivity attention processing layer is used to adjust the influence weights of different meteorological features according to the degree of influence of meteorological factors on total electricity load in different seasons. Based on the nonlinear correspondence, the disturbances to the total electricity load caused by weather changes, time-of-use electricity price changes, and demand response events are identified, and load disturbance components relative to the inertial forecast results are generated based on the disturbances. The load disturbance components are used to represent the increase or decrease relative to the inertial forecast results. Based on the load disturbance components, disturbance prediction results are generated.

[0077] Specifically, when extracting marketing features, time-of-use electricity rates directly reflect the cost of electricity consumption at different times, directly guiding users' electricity consumption behavior; demand response event identifiers quantify the power company's activities in adjusting load through market mechanisms. These features can effectively capture the immediate impact of market and policy levels on load. Regarding meteorological features, in addition to basic temperature and rainfall information, this application further generates continuous temperature impact features and human comfort features as meteorological derivative features. Continuous temperature impact features reflect the cumulative effect of continuous temperature changes on load, such as the continuous increase in air conditioning load caused by several consecutive days of high temperatures; human comfort features comprehensively consider the impact of various meteorological factors such as temperature and humidity on human perception and behavior, thus more precisely depicting the comprehensive effect of meteorological conditions on electricity load. For time features, by performing One-Hot encoding, sine embedding, or cosine embedding on time period attribute information, discrete or periodic time information can be transformed into numerical representations that are easy for the model to process, thereby effectively capturing the daily, weekly, and seasonal periodic variation patterns of load.

[0078] By combining the aforementioned multi-source heterogeneous marketing features, meteorological features, meteorological derived features, and temporal features, a comprehensive external disturbance feature vector is formed, providing rich and multi-dimensional input for the subsequent disturbance prediction model. This disturbance prediction model is structurally optimized, comprising a temporal convolutional processing layer and a meteorological sensitivity attention processing layer. The temporal convolutional processing layer, through its unique convolutional kernel design, effectively extracts long-term correlated features from the external disturbance feature sequence, such as seasonal load patterns or long-term trends, thereby capturing the deep impact of external factors on the load over a longer timescale. The meteorological sensitivity attention processing layer further enhances the model's adaptability, dynamically adjusting the weights of different meteorological features (such as temperature, rainfall, and continuous temperature effects) on the total electricity load based on the season of the current prediction period. For example, in summer, the model assigns higher weights to high-temperature features, while in winter it focuses more on low-temperature features. This dynamic adjustment mechanism allows the model to more accurately reflect the time-varying impact of meteorological factors.

[0079] Example 5 In some embodiments, determining the prediction scenario to which the date to be predicted belongs includes: Based on calendar event data, identify the Gregorian calendar date, lunar calendar date, holiday attributes, and work-rest adjustment attributes of the day to be predicted; If the holiday attribute indicates that the date to be predicted belongs to a holiday or the period before or after a holiday, then the prediction scenario to which the date to be predicted belongs is determined to be a specific holiday scenario. If the forecast date does not fall under a specific holiday scenario, then the meteorological data will be used to determine whether the forecast date meets the preset abnormal meteorological conditions. If it meets the preset abnormal meteorological conditions, the forecast scenario to which the forecast date belongs will be determined as an environmentally sensitive scenario. If the forecast date does not fall under a specific holiday scenario and does not meet the preset abnormal weather conditions, then the marketing business event data will be used to identify whether there is a demand response event or marketing activity period on the forecast date. If there are no demand response events or marketing activities on the forecast date, then the forecast scenario for the forecast date is determined to be a stable weather scenario. If there is a demand response event or marketing activity period on the forecast date, the demand response event or marketing activity period will be identified as an event correction marker. The event correction marker is used to correct the degree of impact of the disturbance forecast result in the contextual integration.

[0080] Specifically, based on the calendar event data, the Gregorian calendar date, lunar calendar date, holiday attributes, and workday adjustment attributes of the day to be predicted are identified. This can be achieved by querying a preset calendar database or calling a calendar service interface. The Gregorian calendar date provides basic time information, while the lunar calendar date is crucial for identifying traditional Chinese holidays (such as the Spring Festival and Mid-Autumn Festival), as the impact of these holidays on power load is often closely related to the lunar cycle. Holiday attributes and workday adjustment attributes directly indicate whether the day is a statutory holiday or a special workday / rest day formed due to workday adjustments; these attributes are key basis for judging load behavior patterns.

[0081] If the holiday attribute indicates that the date to be predicted belongs to a holiday or the period before or after a holiday, then the prediction scenario to which the date to be predicted belongs is determined to be a specific holiday scenario. For example, the period before or after a holiday can be defined as 1-3 days before the start of the holiday or 1-2 days after the end of the holiday. The load behavior during these periods is often significantly affected by factors such as holiday travel, stocking up, or return trips, and is significantly different from ordinary working days or rest days.

[0082] If the predicted date does not fall under the specific holiday scenario, then based on the meteorological data, it is determined whether the predicted date meets preset abnormal meteorological conditions. If it does meet the preset abnormal meteorological conditions, the prediction scenario to which the predicted date belongs is determined to be an environmentally sensitive scenario. Preset abnormal meteorological conditions may include, but are not limited to, extreme high temperatures, extreme low temperatures, heavy rain, heavy snow, and strong winds. These conditions typically lead to significant abnormal fluctuations in electricity load. When these conditions are met, the predicted date is determined to be an environmentally sensitive scenario, indicating that meteorological factors will have a dominant impact on the load.

[0083] If the predicted date does not fall under the specific holiday scenario and does not meet the preset abnormal weather conditions, then the presence of a demand response event or marketing activity period on the predicted date is identified based on the marketing business event data. If the predicted date does not have a demand response event or marketing activity period, then the prediction scenario to which the predicted date belongs is determined to be a stable weather scenario. This represents a daily, relatively stable electricity consumption environment, where load behavior is mainly affected by inertial trends and periodic factors. If the predicted date has a demand response event or marketing activity period, then the demand response event or marketing activity period is identified as an event correction marker. The event correction marker is used to correct the degree of impact of the disturbance prediction result in scenario-based fusion. Demand response events may include peak shaving and valley filling activities initiated by power companies to balance supply and demand, while marketing activity periods may refer to specific promotions or electricity discount activities. Although these events do not constitute independent prediction scenarios, their load disturbance to local time periods cannot be ignored. Therefore, it is necessary to finely adjust the degree of impact of the disturbance prediction result in scenario-based fusion through event correction markers.

[0084] Example 6 In some embodiments, generating raw load forecast results includes: The corresponding prediction channel is invoked based on the predicted scenario, and the corresponding scenario-based fusion parameters are obtained; When the predicted scenario is a stable weather scenario, the proportion of inertial prediction results in the scenario fusion is increased, and disturbance prediction results are used for auxiliary correction to generate the original load prediction results. When the predicted scenario is a specific holiday scenario, historical holiday load data for the same period are extracted based on the lunar calendar alignment relationship, and the pre-holiday baseline load is determined based on the total electricity load data of normal working days before the holiday. Based on the variation of historical holiday load data relative to the pre-holiday baseline load, historical holiday load variation characteristics are generated, and the scenario-based fusion results are corrected for holidays using these historical holiday load variation characteristics to generate the original load forecast results. When the prediction scenario is an environmentally sensitive scenario, the temperature forecast value for the day to be predicted and the measured temperature values ​​for consecutive historical periods before the day to be predicted are obtained. The cumulative effect temperature characteristics are calculated after assigning different weights to the temperature forecast value for the day to be predicted and the measured temperature values ​​for consecutive historical periods before the day to be predicted. Among them, the weight corresponding to the temperature forecast value is greater than the weight corresponding to the measured temperature value, and the weight corresponding to the measured temperature value that is closer to the day to be predicted in the consecutive historical periods is greater. The proportion of disturbance prediction results in scenario-based fusion is increased based on the cumulative effect temperature characteristics, and the scenario-based fusion results are modified with meteorological sensitivity to generate the original load prediction results.

[0085] Specifically, when generating the raw load forecast results, the system first selects and activates the corresponding forecast processing path based on the determined forecast scenario (e.g., a stable weather scenario, a specific holiday scenario, or an environmentally sensitive scenario). This involves calling the corresponding forecast channel and obtaining the corresponding scenario-specific fusion parameters. Each forecast channel has a pre-set set of fusion strategies and parameters optimized for a specific scenario. These parameters include, but are not limited to, the weighting of inertial and disturbance forecast results, correction factors, and specific correction algorithm configurations. This approach ensures that the subsequent fusion process accurately adapts to the characteristics of the current forecast scenario, improving the targeting and effectiveness of the forecast.

[0086] When the predicted scenario is a stable meteorological scenario, electricity load typically exhibits strong historical regularity and periodicity, and is less affected by sudden external factors. Therefore, in this scenario, inertial prediction results have high reliability due to their excellent ability to capture historical load trends and load cycle characteristics. To fully utilize this advantage, the proportion of inertial prediction results in the scenario fusion process will be significantly increased, making it a major component of the original load prediction results. Meanwhile, although disturbance prediction results have a relatively small impact in this scenario, they can still provide fine-grained corrections for minor external factors (such as daily temperature fluctuations and non-major marketing events). These disturbance prediction results can then be used for auxiliary corrections to further improve the accuracy of the predictions, ultimately generating the original load prediction results.

[0087] When the predicted scenario is a specific holiday, load behavior differs significantly from that of ordinary workdays or weekends, and the dates of many traditional holidays (such as the Spring Festival and Mid-Autumn Festival) are determined based on the lunar calendar. To accurately capture the load characteristics of these holidays, it is necessary to precisely extract historical holiday load data from historical data that is the same as or similar to the predicted date, based on the lunar calendar alignment. This historical load data reflects the overall level and fluctuation pattern of electricity load during holidays. Furthermore, to quantify the impact of holidays on load, a "pre-holiday baseline load" needs to be determined. This is typically obtained by analyzing the total electricity load data of several normal workdays before the holiday, representing the normal load level without the holiday's influence. After determining the historical holiday load data and the pre-holiday baseline load, the change in historical holiday load data relative to the pre-holiday baseline load can be calculated. This change, such as the percentage or absolute value of the load decrease, constitutes the historical holiday load change characteristic. This characteristic quantifies the typical impact pattern of holidays on electricity load. In the scenario fusion phase, this historical holiday load change characteristic will be used to correct the initial fusion results for holiday-related issues. The correction process can be to adjust the preliminary fusion results according to the change characteristics, such as lowering the overall value or adjusting the curve shape according to historical patterns, so that the prediction results are more in line with the actual electricity consumption patterns during holidays, and finally generate the original load prediction results.

[0088] For example: The "baseline load" is defined as the average load Pbase on normal working days before the Spring Festival. This means the average actual load of the week leading up to the Spring Festival holiday is selected as the baseline. This value represents the normal electricity consumption level in the region before the holiday, reflecting the basic capacity and demand of local industrial, commercial, and residential electricity consumption. The "Spring Festival factor" is calculated as the percentage decrease in load relative to the baseline load for each day during the Spring Festival period in historical years (New Year's Eve, New Year's Day...the seventh day of the Lunar New Year). It can quantify the decline of daily load relative to the benchmark during the Spring Festival in history and capture the daily variation pattern of load during the Spring Festival.

[0089] ; Where d represents the specific lunar calendar date; This represents the actual load value of the lunar date d in the historical year; This represents the baseline load value for a normal workday prior to a special holiday in a historical year.

[0090] By combining actual electricity consumption data with historical correction factors, the predicted load value for the target day can be calculated. : ; in, This represents the current baseline load, which is the average load of normal working days before the Spring Festival this year (such as the week before the festival), and represents the "basic capacity" of electricity consumption in the region for the year. This represents the Spring Festival factor corresponding to the lunar date d, calculated from historical data. It represents the economic growth coefficient, which is based on the regional economic growth rate, changes in industrial capacity, and population flow in that year.

[0091] By using lunar calendar alignment technology, the problem of pattern matching caused by the misalignment of holiday dates is solved. The relative change of load is quantified by the Spring Festival factor, and the economic growth coefficient is introduced to adapt to the differences in annual electricity consumption characteristics. This can accurately depict the load change process of the sudden drop in industrial load during the Spring Festival and the step-by-step recovery after the holiday, thus improving the accuracy of holiday load forecasting.

[0092] When the predicted scenario is an environmentally sensitive scenario, the electricity load exhibits high sensitivity to meteorological factors (especially temperature). To more accurately assess the impact of temperature on the load, this application introduces a cumulative effect temperature characteristic. This considers not only the temperature forecast value for the predicted day but also the measured temperature values ​​over consecutive historical periods prior to the predicted day. This is because human perception and equipment operating status are affected by temperatures over multiple consecutive days, not just the temperature on the current day. When calculating the cumulative effect temperature characteristic, different weights are assigned to the temperature forecast value for the predicted day and the measured temperature values ​​over consecutive historical periods prior to the predicted day. The weight corresponding to the temperature forecast value is greater than the weight corresponding to the measured temperature value, because the temperature forecast value for the predicted day directly reflects the meteorological conditions of the next day and plays a decisive role in the load forecast for that day. Simultaneously, the measured temperature values ​​closer to the predicted day within the consecutive historical periods have a greater weight, reflecting the continuous impact of recent temperatures on the current load; for example, the measured temperature value of the previous day has a higher weight than that of the two days prior. This weighting method can more realistically simulate the comprehensive and dynamic impact of temperature on the load. In environmentally sensitive scenarios, meteorological factors are a key source of load fluctuations. Therefore, to accurately capture this sensitivity, the proportion of the disturbance prediction results in the scenario-based fusion process is increased based on the cumulative effect temperature characteristics. Simultaneously, meteorological sensitivity corrections are applied to the scenario-based fusion results. This correction can involve raising or lowering the overall load curve, or fine-tuning the load for specific time periods, to reflect the additional impact of the cumulative effect temperature characteristics on the load, thereby generating original load prediction results that better reflect actual electricity consumption in environmentally sensitive scenarios.

[0093] For example: The load is related not only to the temperature of the day, but also to the cumulative temperature of the previous few days. To capture this pattern, the model introduces the cumulative effect temperature feature. (Equivalent temperature) is the core characteristic variable: ; in, This represents the cumulative effect temperature characteristics, which are the actual temperature characteristic values ​​input to the model, comprehensively reflecting the superimposed impact of three consecutive days of temperature on the load; This indicates the predicted temperature value for the day to be forecasted, representing the immediate impact of the temperature on that day. This represents the measured temperature value the day before the forecast date, with a weighting coefficient of 0.3, reflecting the lagging effect of yesterday's temperature. This represents the measured temperature values ​​for the two days prior to the forecast date, with a weighting coefficient of 0.1, reflecting the residual influence of the previous day's temperature.

[0094] Statistical analysis of historical load and temperature data revealed a significant threshold effect in the load's sensitivity to temperature: when the ambient temperature exceeds 28℃ (summer cooling load threshold) or falls below 10℃ (winter heating load threshold), the slope of the load change with temperature (i.e., temperature sensitivity) increases significantly, and even small temperature fluctuations can trigger large changes in the load; while within the comfortable range of 10℃ to 28℃, the load response to temperature is relatively gentle.

[0095] Based on this principle, this technical solution incorporates a temperature threshold activation function into the model to achieve piecewise linear regression for load forecasting. First, the cumulative effect temperature of the day to be predicted is... Compared with the critical thresholds of 28℃ and 10℃, if If the load exceeds the threshold range, a high-sensitivity prediction mode is automatically activated, using a larger response slope to capture sudden spikes in load. If the load is within the threshold range, the normal sensitivity mode is maintained. This mechanism can quickly and accurately respond to spikes in load caused by extreme weather, solving the problem of excessive prediction deviation outside the threshold range in traditional linear models.

[0096] Example 7 In some embodiments, business logic verification for adapting electricity marketing business rules is performed based on the original load forecast results, including: Based on the installed capacity information and transformer capacity information in the marketing archives, the original load forecast results are checked for capacity constraints to determine whether the original load forecast results exceed the allowable capacity range. Based on the distributed power access status and historical minimum total electricity load of the target marketing area, the original load forecast results are checked for non-negativity constraints to determine whether the original load forecast results have negative values ​​or abnormally low values. Based on the load variation range of adjacent time periods, the allowable variation range of equipment, and the historical load fluctuation range, the original load forecast results are checked for ramp constraints to determine whether there are any abnormal sudden changes in the original load forecast results. Based on the relationship between the original load forecast results and the historical total electricity load data for the same period, and in conjunction with new capacity expansion records, abnormal weather information and business event information, the original load forecast results are checked for trend consistency to determine whether there is any abnormal growth or abnormal decline in the original load forecast results that lack business support. Based on the results of capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification, business logic verification results are generated.

[0097] Specifically, during capacity constraint verification, the system checks the original load forecast results against the application capacity and transformer capacity information in the marketing archive data to determine whether the original load forecast results exceed the allowable capacity range. For example, the system can obtain information such as the contract application capacity of each user in the target marketing area, as well as the rated capacity and maximum allowable load of transformers in the area, from the marketing archive data. Then, the original load forecast results are compared with these capacity limits. If the forecast load exceeds the preset threshold (e.g., 95% or 100%) of the application capacity or transformer capacity at any time period, the capacity constraint verification is considered to have failed.

[0098] During the non-negativity constraint verification, the system assesses the original load forecast based on the distributed power supply access status and historical minimum total electricity load of the target marketing area to determine if the original load forecast contains negative or abnormally low values. For example, considering that the access of distributed power sources (such as photovoltaics and wind power) may lead to net loads approaching zero or even being negative during local periods (i.e., power being supplied to the grid), the overall total electricity load forecast should generally remain non-negative. Therefore, this verification checks whether the forecast result is less than zero or lower than a reasonable low value threshold set based on historical minimum load and distributed power supply access status. If the forecast value is negative or lower than this threshold, it is marked as failing the non-negativity constraint verification.

[0099] During the ramp constraint verification, the system checks the original load forecast results based on the load change amplitude of adjacent time periods, the allowable change range of equipment, and the historical load fluctuation range to determine whether there are any abnormal abrupt changes in the original load forecast results. For example, changes in power load are usually gradual and do not occur drastically in a very short period of time. This verification calculates the load change rate or amount of the original load forecast results between consecutive time points (such as adjacent 15 minutes or 1 hour) and compares it with the maximum load change rate that power equipment (such as generator sets and transformers) can withstand and the normal fluctuation range of historical loads. If the instantaneous change amplitude of the predicted load exceeds the preset ramp limit, the ramp constraint verification is considered to have failed.

[0100] During trend consistency verification, the system assesses the original load forecast based on the relationship between the original load forecast and historical total load data for the same period, combined with records of new capacity expansion, abnormal weather information, and business event information. This assessment determines whether the original load forecast exhibits abnormal growth or decline lacking business support. For example, the verification analyzes the long-term trend of the forecast load (e.g., daily, weekly, and monthly year-on-year changes) and compares it with known business developments (e.g., new users, enterprise expansion, equipment upgrades), abnormal weather events (e.g., extreme heat waves, cold waves), and major business events (e.g., large-scale events, policy adjustments). If the forecast trend does not match this supporting information—for example, a significant increase in forecast load without new capacity expansion or abnormal weather events, or a significant decrease in forecast load without major production shutdowns or energy-saving policies—then the trend consistency verification fails.

[0101] Finally, the system will generate a comprehensive business logic verification result based on the results of the above capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification.

[0102] Example 8 In some embodiments, the original load forecast result is maintained or modified based on the business logic verification result to obtain the target load forecast result, including: When the business logic verification result indicates that the original load forecast result passes the capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification, the original load forecast result will be used as the target load forecast result. When the business logic verification result indicates that the original load forecast result fails the capacity constraint verification, the forecast value that exceeds the capacity allowable range will be corrected to the capacity allowable range. When the business logic verification result indicates that the original load forecast result fails the non-negative constraint verification, the forecast value corresponding to the negative value or abnormally low value for the time period is replaced and corrected according to the inertial forecast result. When the business logic verification result indicates that the original load forecast result fails the ramp constraint verification, the forecast curve segment with abnormal abrupt changes is smoothed and corrected. When the business logic verification result indicates that the original load forecast result fails the trend consistency verification, the abnormal trend is reviewed based on the newly added expansion records, abnormal meteorological information and business event information, and the forecast value for the corresponding period is maintained or corrected based on the review result to obtain the target load forecast result. When the business logic verification result triggers a high-risk warning condition, the automatic release of the target load forecast result is suspended, and the corresponding original load forecast result, business logic verification result, and correction result are sent to the manual review end so that the target load forecast result can be confirmed or adjusted according to the manual review result.

[0103] When the business logic verification result indicates that the original load forecast result fails the capacity constraint verification, the system corrects the forecast value that exceeds the allowable capacity range to the allowable range. For example, if the predicted load value exceeds the maximum carrying capacity of the transformer or line in the target marketing area, the predicted value will be adjusted to the maximum allowable capacity value. This correction ensures that the forecast result is physically achievable and complies with the power system operation specifications, avoiding potential equipment overload or power supply risks caused by over-capacity forecasting.

[0104] When the business logic verification result indicates that the original load forecast result fails the non-negativity constraint verification, i.e., the forecast value is negative or abnormally low, close to zero, which is unreasonable in actual electricity load, the system will replace and correct the forecast value for the corresponding time period based on the inertial forecast result. The inertial forecast result is generated based on the trend and periodic characteristics of historical load, and usually has good stability and benchmark reference value. It can effectively replace unreasonable negative values ​​or abnormally low values ​​with reasonable load levels that are more in line with historical patterns, thereby improving the accuracy of the forecast result.

[0105] When the business logic verification result indicates that the original load forecast result fails the ramp constraint verification, it means that there are abnormally drastic changes in the forecast load curve between adjacent time periods, that is, the rate of load increase or decrease exceeds the reasonable range that the power equipment or system can withstand. In response, the system will perform smoothing correction on the forecast curve segments with abnormal abrupt changes. Smoothing correction can be achieved by applying various filtering algorithms, such as moving average, exponential smoothing, or spline interpolation, aiming to eliminate peaks or sudden drops in the forecast curve, making the load change trend more stable and continuous, consistent with the physical inertia of power system load changes.

[0106] When the business logic verification result indicates that the original load forecast result fails the trend consistency verification, it means that the predicted load trend contradicts the known business background information. In this case, the system will review the abnormal trend based on new expansion records, abnormal weather information, and business event information. For example, if the predicted load shows abnormal growth but there are no records of new large users or expansion upgrades, nor are there meteorological factors such as abnormal high temperatures or large-scale marketing activities to support it, then the growth trend may need to be revised. The review result will determine whether to maintain the original forecast value (if the abnormal trend has reasonable support) or revise the forecast value for the corresponding period (if the abnormal trend lacks support), ultimately obtaining the target load forecast result. This ensures that the forecast result maintains a high degree of logical consistency with actual business development and changes in the external environment.

[0107] Furthermore, to enhance the robustness of the forecasts, when the business logic verification results trigger high-risk warning conditions, the system will pause the automatic release of the target load forecast results and send the corresponding original load forecast results, business logic verification results, and correction results to the manual review end for confirmation or adjustment of the target load forecast results based on the manual review results. High-risk warning conditions can be defined according to preset rules, such as when multiple verification items fail simultaneously, or when the correction magnitude exceeds a specific threshold. This mechanism incorporates expert experience, allowing for final judgment and adjustment through manual intervention when there is significant uncertainty or potential risk in the forecast results. This effectively avoids significant business risks that automated forecasting may bring and ensures the accuracy and reliability of the final released forecast results.

[0108] Example 9 The target load prediction result is compared with the actual total electricity load data to generate prediction deviation data, including: Based on the prediction time granularity, the target load prediction results and the actual total electricity load data are divided into multiple corresponding evaluation periods; The time weight of each evaluation period is determined based on the electricity price period attribute or marketing value attribute corresponding to each evaluation period; among them, the time weight corresponding to the peak period is greater than the time weight corresponding to the normal period, and the time weight corresponding to the normal period is greater than the time weight corresponding to the valley period. Calculate the absolute deviation of the load forecast between the target load forecast result and the actual total electricity load data for each evaluation period; Based on the time weight of each evaluation period, the absolute deviation of load forecast, and the actual total electricity load data, a weighted accuracy evaluation result is generated. The prediction deviation data is generated based on the weighted accuracy evaluation results and the absolute deviation of load prediction for each evaluation period.

[0109] Traditional mean squared error (MSE) metrics fail to reflect the impact of load forecasting on marketing revenue. In this embodiment, to balance spatiotemporal characteristics with marketing value in reflecting the impact of load forecasting on marketing revenue, a weighted accuracy metric is constructed: ; Where T represents the total number of time periods within the forecast period (if a short-term load forecast with a 15-minute granularity is used, T=96 per day; if it is a 1-hour granularity, T=24 per day). This represents the time weighting coefficient for the t-th time period; This represents the load forecast value for the t-th time period (unit: kW or MW). Represents the actual load value in the t-th time period (unit: kW or MW); This represents the absolute error of load forecasting for the t-th time period.

[0110] in the formula It is a time-weighted coefficient, exhibiting a characteristic of high weight during peak hours and low weight during trough hours. That is, it sets a higher weight during peak periods when spot prices are relatively high (such as 19:00-21:00). = 1.5; Set at lower price troughs. = 0.8. This evaluation system forces the algorithm model to prioritize optimizing the prediction accuracy during high-value periods, thereby directly serving the maximization of marketing revenue.

[0111] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0112] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0113] Any process or method described 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, and 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 function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0114] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0115] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0116] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0117] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0118] 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. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0119] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A robust short-term load forecasting method for electricity marketing operations, characterized in that, include: Obtain historical electricity marketing data from the target marketing region as sample data; The historical electricity marketing data includes at least: total electricity load data, marketing record data, meteorological data, calendar event data, and marketing business event data; Based on the total electricity load data, load trend characteristics and load cycle characteristics are extracted to generate inertial prediction results; Based on the marketing archive data, meteorological data, calendar event data, and marketing business event data, the correspondence between external disturbance factors and total electricity load is identified, and disturbance prediction results are generated. Based on the meteorological data, calendar event data, and marketing business event data corresponding to the date to be predicted, the prediction scenario to which the date to be predicted belongs is determined; the prediction scenarios include stable weather scenarios, specific holiday scenarios, and environmentally sensitive scenarios. Based on the predicted scenario, the inertial prediction results and disturbance prediction results are fused in a scenario-based manner to generate the original load prediction results; wherein, in the stable meteorological scenario, the proportion of the inertial prediction results is increased; in the specific holiday scenario, holiday corrections are performed based on the lunar calendar alignment relationship and historical holiday load change characteristics; and in the environmentally sensitive scenario, meteorological sensitivity corrections are performed based on continuous meteorological impact characteristics. Based on the original load forecast results, perform business logic verification to adapt to the electricity marketing business rules; the business logic verification includes capacity constraint verification, non-negativity constraint verification, ramp constraint verification and trend consistency verification. The original load forecast result is maintained or modified based on the business logic verification result to obtain the target load forecast result.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the actual total electricity load data corresponding to the day to be predicted, compare the target load prediction result with the actual total electricity load data, and generate prediction deviation data; Based on the prediction deviation data, determine the type and degree of deviation of the target load prediction results under different prediction scenarios; The actual total electricity load data, meteorological data corresponding to the forecast day, calendar event data, marketing business event data, and forecast deviation data are added to the sample data. Based on the type and degree of deviation, adjust the fusion parameters for scene-based fusion. Based on the correspondence between the business logic verification results and the prediction deviation data, adjust at least one of the business logic verification rules among capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification.

3. The method according to claim 1, characterized in that, Obtain historical electricity marketing data for the target marketing region as sample data, including: Obtain total electricity load data, marketing archive data, weather data, calendar event data, and marketing business event data for the target marketing region over historical periods; The total electricity load data, marketing archive data, meteorological data, calendar event data, and marketing business event data are time-aligned according to the same time granularity to obtain aligned historical electricity marketing business data. Anomaly identification is performed on the total electricity load data in the aligned historical electricity marketing business data to determine meter tripping data and zero value data; A density-based clustering outlier identification algorithm is used to identify load outliers corresponding to the meter jump data and zero value data; Based on the adjacent historical time period data and historical data of the same period corresponding to the load outlier, the weighted average of historical data is calculated, and the load outlier is filled in using the weighted average of historical data to obtain the sample data.

4. The method according to claim 1, characterized in that, Generate inertial prediction results, including: The total electricity load data is sorted according to a preset time granularity to generate a historical load sequence; The historical load sequence is input into a pre-built inertial prediction model; the inertial prediction model is configured with adaptive smoothing parameters and includes a baseline load level, a load trend term, and a load seasonal term. The basic electricity consumption level of the target marketing area is determined based on the benchmark load level, and load trend characteristics are extracted based on the load trend item and load cycle characteristics are extracted based on the load seasonal item. Based on the prediction step size and cycle length, the baseline load level, load trend term, and load seasonal term are combined to generate an initial inertial prediction result; Based on the deviation between historical prediction results and historical actual total electricity load data, cumulative error data and average absolute deviation data are generated, and an adaptive tracking signal is generated based on the cumulative error data and average absolute deviation data. When the adaptive tracking signal exceeds the preset offset threshold, the adaptive smoothing parameter is increased to increase the weight corresponding to the recent total electricity load data; when the adaptive tracking signal does not exceed the preset offset threshold, the adaptive smoothing parameter is decreased to decrease the weight corresponding to the recent total electricity load data. The initial inertial prediction result is updated based on the adjusted adaptive smoothing parameters to generate the inertial prediction result.

5. The method according to claim 1, characterized in that, Generate disturbance prediction results, including: Time-of-use electricity rates and demand response event identifiers are extracted from the marketing archive data and marketing business event data as marketing features; Temperature and rainfall information are extracted from the meteorological data as meteorological features, and continuous temperature impact features and human comfort features are generated as meteorological derivative features based on the temperature and rainfall information. Extract time period attribute information from the calendar event data, and perform One-Hot encoding, sine embedding, or cosine embedding on the time period attribute information to generate time features; The marketing characteristics, meteorological characteristics, meteorological derivative characteristics, and time characteristics are combined into external disturbance characteristics; The external disturbance characteristics and total electricity load data are input into a pre-built disturbance prediction model, and the nonlinear correspondence between the external disturbance characteristics and the total electricity load is identified through the disturbance prediction model. The disturbance prediction model includes a temporal convolutional processing layer and a meteorological sensitivity attention processing layer. The temporal convolutional processing layer is used to extract long-term correlation features between the external disturbance features and the total electricity load data. The meteorological sensitivity attention processing layer is used to adjust the influence weights of different meteorological features according to the degree of influence of meteorological factors on the total electricity load in different seasons. Based on the nonlinear correspondence, the disturbances to the total electricity load caused by weather changes, time-of-use electricity price changes, and demand response events are identified, and a load disturbance component relative to the inertial prediction result is generated based on the disturbances; the load disturbance component is used to represent the increase or decrease relative to the inertial prediction result. The disturbance prediction result is generated based on the load disturbance component.

6. The method according to claim 1, characterized in that, Determining the prediction scenario to which the date to be predicted belongs includes: Based on the calendar event data, identify the Gregorian calendar date, lunar calendar date, holiday attributes, and work-rest adjustment attributes of the day to be predicted; If the holiday attribute indicates that the date to be predicted belongs to a holiday or the period before or after a holiday, then the prediction scenario to which the date to be predicted belongs is determined to be a specific holiday scenario. If the predicted day does not belong to the specific holiday scenario, then the meteorological data is used to determine whether the predicted day meets the preset abnormal meteorological conditions, and if it meets the preset abnormal meteorological conditions, the predicted scenario to which the predicted day belongs is determined to be an environmentally sensitive scenario. If the predicted day does not belong to the specific holiday scenario and does not meet the preset abnormal weather conditions, then the marketing business event data is used to identify whether there is a demand response event or marketing activity period on the predicted day; If the demand response event or marketing activity period does not exist on the day to be predicted, then the prediction scenario to which the day to be predicted belongs is determined to be a stable weather scenario. If the demand response event or marketing activity period exists on the date to be predicted, then the demand response event or marketing activity period is identified as an event correction marker. The event correction marker is used to correct the degree of impact of the disturbance prediction result in the contextual fusion.

7. The method according to claim 1, characterized in that, Generate raw load forecast results, including: The corresponding prediction channel is invoked according to the predicted scenario, and the corresponding scenario-based fusion parameters are obtained; When the predicted scenario is a stable weather scenario, the proportion of the inertial prediction result in the scenario fusion is increased, and the disturbance prediction result is used for auxiliary correction to generate the original load prediction result. When the predicted scenario is a specific holiday scenario, historical holiday load data for the same period are extracted according to the lunar calendar alignment relationship, and the pre-holiday baseline load is determined according to the total electricity load data of normal working days before the holiday. Based on the variation of the historical holiday load data relative to the pre-holiday baseline load, historical holiday load variation characteristics are generated, and the scenario-based fusion results are corrected for holidays using these historical holiday load variation characteristics to generate the original load prediction results. When the prediction scenario is an environmentally sensitive scenario, the temperature forecast value for the day to be predicted and the measured temperature values ​​for consecutive historical periods before the day to be predicted are obtained. After assigning different weights to the temperature forecast value for the day to be predicted and the measured temperature values ​​for consecutive historical periods before the day to be predicted, the cumulative effect temperature characteristics are calculated. Among them, the weight corresponding to the temperature forecast value is greater than the weight corresponding to the measured temperature value, and the weight corresponding to the measured temperature value that is closer to the day to be predicted in the consecutive historical periods is greater. Based on the cumulative effect temperature characteristics, the proportion of the disturbance prediction results in the scenario fusion is increased, and the scenario fusion results are subject to meteorological sensitivity correction to generate the original load prediction results.

8. The method according to claim 1, characterized in that, Based on the original load forecast results, perform business logic verification to adapt to the electricity marketing business rules, including: Based on the application capacity information and transformer capacity information in the marketing file data, the original load forecast result is checked for capacity constraints to determine whether the original load forecast result exceeds the allowable capacity range. Based on the distributed power access status and historical minimum total electricity load of the target marketing area, the original load prediction results are subjected to non-negative constraint verification to determine whether the original load prediction results have negative values ​​or abnormally low values. Based on the load variation range of adjacent time periods, the allowable variation range of equipment, and the historical load fluctuation range, the original load forecast results are subjected to ramp constraint verification to determine whether there are any abnormal sudden changes in the original load forecast results. Based on the relationship between the original load forecast results and the historical total electricity load data for the same period, and in conjunction with new capacity expansion records, abnormal weather information and business event information, the original load forecast results are checked for trend consistency to determine whether there is any abnormal growth or abnormal decline in the original load forecast results due to lack of business support. Based on the results of the capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification, a business logic verification result is generated.

9. The method according to claim 1, characterized in that, Based on the business logic verification results, the original load forecast results are maintained or corrected to obtain the target load forecast results, including: When the business logic verification result indicates that the original load forecast result passes the capacity constraint verification, non-negativity constraint verification, ramp constraint verification, and trend consistency verification, the original load forecast result is taken as the target load forecast result. When the business logic verification result indicates that the original load prediction result fails the capacity constraint verification, the prediction value that exceeds the capacity allowable range will be corrected to the capacity allowable range. When the business logic verification result indicates that the original load prediction result fails the non-negative constraint verification, the prediction value corresponding to the negative value or abnormally low value period is replaced and corrected according to the inertial prediction result. When the business logic verification result indicates that the original load prediction result fails the ramp constraint verification, the prediction curve segment with abnormal abrupt changes is smoothed and corrected. When the business logic verification result indicates that the original load forecast result fails the trend consistency verification, the abnormal trend is reviewed based on the newly added expansion records, abnormal meteorological information and business event information, and the forecast value for the corresponding period is maintained or corrected based on the review result to obtain the target load forecast result. When the business logic verification result triggers a high-risk warning condition, the automatic release of the target load forecast result is suspended, and the corresponding original load forecast result, business logic verification result, and correction result are sent to the manual review end so that the target load forecast result can be confirmed or adjusted according to the manual review result.

10. The method according to claim 2, characterized in that, The target load prediction result is compared with the actual total electricity load data to generate prediction deviation data, including: Based on the prediction time granularity, the target load prediction results and the actual total electricity load data are divided into multiple corresponding evaluation periods; The time weight of each evaluation period is determined based on the electricity price period attribute or marketing value attribute corresponding to each evaluation period; among them, the time weight corresponding to the peak period is greater than the time weight corresponding to the normal period, and the time weight corresponding to the normal period is greater than the time weight corresponding to the valley period. Calculate the absolute deviation of the load forecast between the target load forecast result and the actual total electricity load data for each evaluation period; Based on the time weight of each evaluation period, the absolute deviation of load forecast, and the actual total electricity load data, a weighted accuracy evaluation result is generated. The prediction deviation data is generated based on the weighted accuracy evaluation results and the absolute deviation of load prediction for each evaluation period.

Images