Load prediction method and device for power spot market agent enterprise set, equipment and storage medium

Abstract

This invention provides a load forecasting method, apparatus, equipment, and storage medium for a set of electricity spot market agent companies, relating to the field of artificial intelligence technology. The method includes: performing time-shifting and alignment processing on original market load data, original meteorological data, and original actual total load data based on calendar data to obtain market load data, meteorological data, and actual total load data aligned with the forecast date; extracting features from the market load data, meteorological data, calendar data, and actual total load data to obtain derived feature combinations; generating the average agent load forecast value for the set of agent companies on the forecast date using a load forecasting model based on the derived feature combinations; and determining the total load forecast value for the set of agent companies on the forecast date based on the average agent load forecast value and the total number of agent companies. This approach reduces the impact of feature misalignment caused by differences in the timeliness of multi-source data disclosure, improving the accuracy of load forecasting results.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a load forecasting method, apparatus, equipment, and storage medium for a group of electricity spot market agents. Background Technology

[0002] In the electricity spot market, business models such as agency electricity sales, agency aggregation, and agency declaration are developing rapidly. For agencies, they need to predict the total load of the group of companies they represent on a future day before the declaration date, so as to complete the electricity purchase declaration, deviation control, electricity trading optimization, and risk management on behalf of each agency in the group of companies.

[0003] In some related technologies, the actual total load data and market load data of a group of agent companies can be collected before the application date. A pre-trained model can then be used to directly predict the total load of the agent company group on a future day, allowing the agency to submit future electricity purchase demand for the group on the application date. However, this load forecasting method suffers from insufficient prediction accuracy: the actual total load data and market load data of the agent company group are often subject to data disclosure lags, making it difficult for agencies to collect complete load data for the most recent day or even several days before the application date for forecasting. Even if the actual total load data and market load data for the most recent days are collected, load data from different sources may have timeliness differences. If the actual total load data and market load data with timeliness differences are directly used for future load forecasting, the model may suffer from insufficient prediction accuracy due to the timeliness differences between different features. For example, before the application date, the agency may only collect the actual total load data of the agent company group from one week ago, but the latest market load data is from one day ago. In this case, there is a significant timeliness difference between the actual total load data collected by the agency and the market load data, which will affect the prediction accuracy of the model.

[0004] In summary, existing load forecasting methods suffer from insufficient forecasting accuracy, leading to inaccurate load forecasting results. Summary of the Invention

[0005] This invention provides a load forecasting method, apparatus, equipment, and storage medium for a group of electricity spot market agents, in order to solve the defects of insufficient forecasting accuracy in existing load forecasting methods, which leads to inaccurate load forecasting results.

[0006] This invention provides a load forecasting method for a set of agents in the electricity spot market, comprising: acquiring original market load data, original meteorological data, calendar data, original true total load data of the agent set, and the total number of agents in the agent set prior to the forecast date; the agent set includes at least two agents; based on the calendar data, performing time-shifting alignment processing on the original market load data, original meteorological data, and original true total load data to obtain market load data, meteorological data, and true total load data aligned with the forecast date; extracting features from the market load data, meteorological data, calendar data, and true total load data to obtain derived feature combinations; based on the derived feature combinations, performing load forecasting through a pre-trained load forecasting model to generate the average agent load forecast value of the agent set on the forecast date; and determining the total load forecast value of the agent set on the forecast date based on the average agent load forecast value and the total number of agents.

[0007] According to the present invention, a load forecasting method for a set of electricity spot market agents is provided. The original true total load data includes the true total load values ​​of the agent set over multiple known time periods. Before obtaining market load data, meteorological data, and true total load data aligned with the forecast date by performing time-shifting and alignment processing on the original market load data, original meteorological data, and original true total load data based on calendar data, the method further includes: determining whether the original true total load data meets the timeliness requirements of load forecasting; if the original true total load data does not meet the timeliness requirements of load forecasting, then obtaining the historical true total load data of the agent set over several recent historical dates; a historical date includes multiple historical time periods, each historical time period corresponds to a known time period, and the historical true total load data within a historical date includes the historical true total load value of the agent set in each historical time period; based on the historical true total load data of several recent historical dates, the original true total load data is supplemented and updated; wherein, the true total load value of a known time period is the average of the historical true total load values ​​of the historical time periods corresponding to the several recent historical dates.

[0008] According to the present invention, a load forecasting method for a set of electricity spot market agents is provided, which extracts features from market load data, meteorological data, calendar data, and actual total load data to obtain a derived feature combination. This includes: extracting features from market load data, meteorological data, calendar data, and actual total load data to obtain lag features of the actual total load data, market load features of the market load data, meteorological features of the meteorological data, index features of the calendar data, lunar calendar features, and load ratio features of the agent set on lunar dates; the index features include the holiday identifier of the date to be forecasted; if the date to be forecasted falls within a holiday window, a derived feature combination is constructed based on the lunar calendar features, load ratio features, holiday identifiers, and market load features; if the date to be forecasted does not fall within a holiday window, a derived feature combination is constructed based on the market load features, meteorological features, index features, and lag features.

[0009] According to the present invention, a load forecasting method for a set of agents in the electricity spot market is provided. The load forecasting model includes a first load forecasting sub-model and a second load forecasting sub-model. Based on a combination of derived features, load forecasting is performed through a pre-trained load forecasting model to generate the average agent load forecast value of the agent set on the forecast date. The method includes: dividing the forecast date into time periods to obtain a first time period set and a second time period set; the first time period set and the second time period set include different time periods of the forecast date; inputting the combination of derived features into the first load forecasting sub-model to obtain the first average agent load forecast value of the agent set in the first time period set output by the first load forecasting sub-model; inputting the combination of derived features into the second load forecasting sub-model to obtain the second average agent load forecast value of the agent set in the second time period set output by the second load forecasting sub-model; wherein, the average agent load forecast value includes the first average agent load forecast value and the second average agent load forecast value.

[0010] The load forecasting method for a set of electricity spot market agents provided by the present invention further includes: a first time period set being a daytime time period set, and a second time period set being a non-daytime time period set; or, a first time period set being a peak time period set, and a second time period set being a non-peak time period set.

[0011] According to the present invention, a load forecasting method for a set of agents in the electricity spot market determines the total load forecast value of the set of agents on the forecast date based on the average agent load forecast value of each agent and the total number of agents. The method includes: multiplying the first average agent load forecast value of each agent by the total number of agents as the first total load forecast value of the set of agents in a first time period; and multiplying the second average agent load forecast value of each agent by the total number of agents as the second total load forecast value of the set of agents in a second time period. The total load forecast value includes the first total load forecast value and the second total load forecast value.

[0012] According to the present invention, a load forecasting method for a set of electricity spot market agents is provided. Based on derived feature combinations, load forecasting is performed using a pre-trained load forecasting model. Before generating the average agent load forecast value for the set of agents on the forecast date, the method further includes: determining a preset cutoff date and obtaining historical data before the preset cutoff date as a training sample set; the preset cutoff date is before the forecast date, and the training sample set includes sample market load data, sample meteorological data, sample calendar data, sample total load data of the sample agent set, and the total number of sample agent enterprises in the sample agent set; based on the sample total load data and the total number of sample agent enterprises, a label for the average agent load value of the sample enterprises is determined; time-shifting and alignment processing and feature extraction are performed on the sample market load data, sample meteorological data, sample calendar data, and sample total load data to obtain a sample derived feature combination; based on the sample derived feature combination and the label for the average agent load value of the sample enterprises, a first initial load forecasting sub-model and a second initial load forecasting sub-model are pre-trained to obtain the first load forecasting sub-model and the second load forecasting sub-model; wherein, the first load forecasting sub-model and the second load forecasting sub-model are LightGBM models.

[0013] This invention also provides a load forecasting device for a set of electricity spot market agents, comprising: an acquisition module for acquiring original market load data, original meteorological data, calendar data, original true total load data of the agent set, and the total number of agents in the agent set prior to the forecast date; the agent set includes at least two agents; a time-shifting alignment module for performing time-shifting alignment processing on the original market load data, original meteorological data, and original true total load data based on calendar data to obtain market load data, meteorological data, and true total load data aligned with the forecast date; a feature extraction module for extracting features from the market load data, meteorological data, calendar data, and true total load data to obtain derived feature combinations; a load forecasting module for performing load forecasting based on the derived feature combinations using a pre-trained load forecasting model to generate the average agent load forecast value for the agent set on the forecast date; and a result recovery module for determining the total load forecast value for the agent set on the forecast date based on the average agent load forecast value and the total number of agents.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-described load forecasting methods for a set of electricity spot market agents.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a load forecasting method for a set of electricity spot market agents as described above.

[0016] The load forecasting method, apparatus, equipment, and storage medium for aggregators of electricity spot market agents provided by this invention introduces a time-shifting alignment mechanism. After acquiring the original market load data, original meteorological data, calendar data, original true total load data of the agency aggregator, and the total number of agencies prior to the forecast date, the original market load data, original meteorological data, and original true total load data are time-shifted and aligned based on the calendar data. This unifies and aligns these multi-source data, which may have time-sensorship discrepancies, to the forecast date, obtaining market load data, meteorological data, and true total load data aligned with the forecast date, thereby reducing time-sensorship discrepancies in the disclosure of multi-source data. To mitigate the impact of feature misalignment, and avoid insufficient prediction accuracy in subsequent models due to feature misalignment, feature extraction is performed on market load data, meteorological data, calendar data, and actual total load data after eliminating the influence of feature misalignment between data. Derived feature combinations are obtained, and load forecasting is performed based on these derived feature combinations using a load forecasting model. This generates the average agent load forecast value for the agent enterprise set on the forecast date. Then, based on the average agent load forecast value and the total number of agent enterprises, the total load forecast value for the agent enterprise set on the forecast date is determined, thus realizing the prediction of the total load of the agent enterprise set in the future and improving the accuracy of load forecasting results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts of the load forecasting method for a group of electricity spot market agents provided by the present invention.

[0019] Figure 2 This is the second flowchart of the load forecasting method for a group of electricity spot market agents provided by the present invention.

[0020] Figure 3 This is the third flowchart of the load forecasting method for a group of electricity spot market agents provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the prediction process of the load prediction model provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the load forecasting device for a group of electricity spot market agents provided by the present invention.

[0023] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] Please see Figures 1 to 4 , Figure 1 This is one of the flowcharts illustrating the load forecasting method for a set of electricity spot market agents provided by this invention. Figure 2 This is the second flowchart illustrating the load forecasting method for a group of electricity spot market agents provided by this invention. Figure 3 This is the third flowchart of the load forecasting method for a group of electricity spot market agents provided by this invention. Figure 4 This is a schematic diagram of the prediction process of the load prediction model provided by the present invention.

[0026] like Figure 1 As shown, the load forecasting method for the electricity spot market agent group includes steps S110 to S150, and the specific steps are as follows: S110: Obtain the raw market load data, raw meteorological data, calendar data, raw true total load data of the agent firm set, and the total number of agents in the agent firm set prior to the forecast date.

[0027] The set of agents includes at least two agents.

[0028] Specifically, such as Figure 2 and Figure 3 As shown, the raw data prior to the forecast date is collected first. The raw data includes the original market load data disclosed before the forecast date, the original meteorological data, the calendar data, the original true total load data of the agent enterprise set, the total number of agents in the agent enterprise set, or information that can reflect changes in the total number of agents.

[0029] Optionally, the raw market load data includes, but is not limited to, central dispatch load data, tie line load data, renewable energy output data, market load forecast data, actual market load data, day-ahead node price, and real-time node price.

[0030] Optionally, the raw meteorological data includes meteorological information from multiple cities in the province where each agent is located, including but not limited to temperature, precipitation, wind speed, irradiance, humidity, and air pressure.

[0031] Optionally, the original calendar data includes, but is not limited to, identification of statutory holidays, identification of rest days, weekday information, quarter information, month information, lunar month information, and lunar day information.

[0032] Preferably, the original market load data, original meteorological data, and original true total load data are subjected to data preprocessing such as data cleaning, missing value imputation, outlier handling, duplicate record cleaning, and time index correction.

[0033] Optionally, during data preprocessing, for general missing values, interpolation, nearest neighbor completion, or preset rule completion can be used to fill in the missing values.

[0034] Optionally, during data preprocessing, outlier detection, upper and lower bound truncation, or a combination of adjacent time-sharing mean correction methods can be used to handle outliers.

[0035] Optionally, during data preprocessing, abnormal time indices, duplicate collection records, and empty time slots in the original data are uniformly corrected.

[0036] S120: Based on calendar data, perform time-shifting and alignment processing on the original market load data, original meteorological data, and original true total load data to obtain market load data, meteorological data, and true total load data aligned with the date to be predicted.

[0037] Understandably, under different regional or market rules, there is often a lag in the disclosure of raw true total load data, which means that the latest disclosed raw true total load data may be data from several days ago, or even data from an even earlier date.

[0038] For example, some regions may be able to disclose raw true total load data from one day prior, while other regions may only be able to disclose raw true total load data from three days prior.

[0039] Since recent raw total load data usually better reflects the current load demand of the agency, in order to ensure the accuracy of load forecasting, it is usually necessary to use recent raw total load data for load forecasting. That is, the raw total load data collected must meet the timeliness requirements of the load forecasting business.

[0040] Therefore, after collecting the raw, true total load data, it is necessary to first determine whether the raw, true total load data meets the timeliness requirements of the load forecasting service.

[0041] Specifically, such as Figure 2 As shown, first determine the actual date corresponding to the original actual total load data (if the original actual total load data is data from 7 days ago, then its corresponding actual date is the date from 7 days ago, and so on). Then determine whether the difference between the actual date and the date to be predicted is within the preset time requirement (for example, the difference between the actual date and the date to be predicted must be within 3 days).

[0042] If the difference between the actual date and the date to be predicted is within the preset time requirement, then the original actual total load data is considered to meet the time requirement of the load forecasting business.

[0043] If the difference between the actual date and the date to be predicted is not within the preset time requirement, then the original actual total load data is considered not to meet the time requirement of the load forecasting business.

[0044] If the original true total load data does not meet the timeliness requirements of the load forecasting business, the historical true total load data of the agency enterprise set over the most recent historical dates (e.g., the most recent 7 days before the actual date) can be extracted first, and the average value of the historical true total load data over the most recent historical dates before the actual date can be calculated. This average value can then be used to supplement and update the currently collected original true total load data.

[0045] Specifically, in this embodiment, the original true total load data used for load forecasting may include the true total load values ​​of the agent group over multiple known time periods; if the original true total load data does not meet the timeliness requirements of the load forecasting business, the historical true total load data of the agent group over the most recent historical dates can be extracted first.

[0046] For each historical date, the historical date may include multiple historical time periods, with each historical time period corresponding to a known time period. The historical true total load data within the historical date includes the historical true total load value of the agency set for each historical time period within that historical date.

[0047] For each known time period within the actual date, first determine the historical time periods corresponding to that known time period from all historical dates, then calculate the average of the historical actual total load values ​​of these corresponding historical time periods, and update the average value to the actual total load value of that known time period.

[0048] After calculating the true total load values ​​for all known time periods, these true total load values ​​for known time periods constitute the final original true total load data, completing the supplementation and update of the original true total load data.

[0049] This design can, to some extent, compensate for the lack of timeliness in the currently collected raw true total load data. When the raw true total load data does not meet the timeliness requirements of load forecasting, the raw true total load data for that day may not necessarily reflect the current load demand of the agency enterprise set. However, the average value of historical true total load data over a recent period can reflect the average load demand of the agency enterprise set over a recent period. The average load demand of the agency enterprise set over a period of time is usually relatively stable, so the representativeness and timeliness of the average load demand are usually better. In addition, the method of supplementing and updating data by using historical true total load data over a recent period can ensure that the subsequent model still has predictability even when the data on the reporting date is incomplete.

[0050] Furthermore, such as Figure 3 As shown, considering that the original market load data, original meteorological data, and original true total load data may differ in their disclosure time, i.e., there may be timeliness differences among the three, in order to avoid the timeliness differences affecting subsequent model predictions, it is necessary to perform timeliness shifting and alignment processing on the original market load data, original meteorological data, and original true total load data based on calendar data to eliminate the timeliness differences between different data.

[0051] Specifically, for the original market load data, the date difference between the corresponding actual date and the date to be predicted is determined based on calendar data, and the original market load data is shifted over time based on this date difference to the date to be predicted.

[0052] Similarly, for the raw meteorological data, the date difference between the actual date and the date to be predicted is determined based on the calendar data, and the raw meteorological data is shifted over time based on this date difference to the date to be predicted.

[0053] Similarly, for the original true total load data, the date difference between the corresponding true date and the date to be predicted is determined based on the calendar data, and the original true total load data is shifted according to the date difference to the date to be predicted.

[0054] After completing the above processing, the original market load data, original meteorological data, and original true total load data can be unified to the caliber of the forecast date, resulting in market load data, meteorological data, and true total load data aligned with the forecast date.

[0055] For example, assuming the current date is t, and the forecast date is t+1, if the actual date corresponding to the raw market load data is t-2, then the raw market load data can be shifted forward by 3 days to participate in subsequent load forecasting.

[0056] This method of shifting and aligning data from different time scales to the date to be predicted can effectively improve the utilization rate of the latest and most valid information and reduce the impact of data misalignment and timeliness differences.

[0057] S130: Extract features from market load data, meteorological data, calendar data, and actual total load data to obtain derived feature combinations.

[0058] S140: Based on the combination of derived features, load forecasting is performed through a pre-trained load forecasting model to generate the average agency load forecast value of the agency enterprise set on the forecast date.

[0059] The average agent load forecast value refers to the average agent load forecast value of all agents in the agent enterprise set.

[0060] S150: Based on the average agent load forecast value and the total number of agent enterprises, determine the total load forecast value of the agent enterprise set on the forecast date.

[0061] Specifically, the product of the average agent load forecast of each enterprise and the total number of agent enterprises on the forecast date is calculated as the total load forecast of the agent enterprise set on the forecast date.

[0062] The load forecasting method for a set of electricity spot market agents provided in this embodiment introduces a time-shifting alignment mechanism. After acquiring the original market load data, original meteorological data, calendar data, original true total load data of the agent set, and the total number of agents prior to the forecast date, the original market load data, original meteorological data, and original true total load data are time-shifted and aligned based on the calendar data. This unifies and aligns these multi-source data, which may have time-sensorship differences, to the forecast date, obtaining market load data, meteorological data, and true total load data aligned with the forecast date. This reduces the feature loss caused by time-sensorship differences in multi-source data disclosure. To mitigate the impact of feature misalignment and avoid insufficient prediction accuracy in subsequent models, feature extraction is performed on market load data, meteorological data, calendar data, and actual total load data after eliminating the influence of feature misalignment between data. This yields derived feature combinations, which are then used in the load forecasting model to perform load forecasting. This generates the average agent load forecast value for the agent enterprise set on the forecast date. Finally, based on the average agent load forecast value and the total number of agent enterprises, the total load forecast value for the agent enterprise set on the forecast date is determined, thus enabling the prediction of the total load of the agent enterprise set in the future and improving the accuracy of load forecasting results.

[0063] In some embodiments, the original true total load data includes the true total load values ​​of the agency set over multiple known time periods. Before performing time-shifting and alignment processing on the original market load data, original meteorological data, and original true total load data based on calendar data to obtain market load data, meteorological data, and true total load data aligned with the forecast date, the method further includes: determining whether the original true total load data meets the timeliness requirements of the load forecasting business; if the original true total load data does not meet the timeliness requirements of the load forecasting business, then obtaining the historical true total load data of the agency set over the most recent historical dates; a historical date includes multiple historical time periods, a historical time period corresponds to a known time period, and the historical true total load data within a historical date includes the historical true total load value of the agency set in each historical time period; the original true total load data is supplemented and updated based on the historical true total load data of the most recent historical dates; wherein, the true total load value of a known time period is the average of the historical true total load values ​​of the historical time periods corresponding to the most recent historical dates.

[0064] Understandably, under different regional or market rules, there is often a lag in the disclosure of raw true total load data, which means that the latest disclosed raw true total load data may be data from several days ago, or even data from an even earlier date.

[0065] For example, some regions may be able to disclose raw true total load data from one day prior, while other regions may only be able to disclose raw true total load data from three days prior.

[0066] Since recent raw total load data usually better reflects the current load demand of the agency, in order to ensure the accuracy of load forecasting, it is usually necessary to use recent raw total load data for load forecasting. That is, the raw total load data collected must meet the timeliness requirements of the load forecasting business.

[0067] Therefore, after collecting the raw, true total load data, it is necessary to first determine whether the raw, true total load data meets the timeliness requirements of the load forecasting service.

[0068] Specifically, such as Figure 2 As shown, first determine the actual date corresponding to the original actual total load data (if the original actual total load data is data from 7 days ago, then its corresponding actual date is the date from 7 days ago, and so on). Then determine whether the difference between the actual date and the date to be predicted is within the preset time requirement (for example, the difference between the actual date and the date to be predicted must be within 3 days).

[0069] If the difference between the actual date and the date to be predicted is within the preset time requirement, then the original actual total load data is considered to meet the time requirement of the load forecasting business.

[0070] If the difference between the actual date and the date to be predicted is not within the preset time requirement, then the original actual total load data is considered not to meet the time requirement of the load forecasting business.

[0071] If the original true total load data does not meet the timeliness requirements of the load forecasting business, the historical true total load data of the agency enterprise set over the most recent historical dates (e.g., the most recent 7 days before the actual date) can be extracted first, and the average value of the historical true total load data over the most recent historical dates before the actual date can be calculated. This average value can then be used to supplement and update the currently collected original true total load data.

[0072] Specifically, in this embodiment, the original true total load data used for load forecasting may include the true total load values ​​of the agent group over multiple known time periods; if the original true total load data does not meet the timeliness requirements of the load forecasting business, the historical true total load data of the agent group over the most recent historical dates can be extracted first.

[0073] For each historical date, the historical date may include multiple historical time periods, with each historical time period corresponding to a known time period. The historical true total load data within the historical date includes the historical true total load value of the agency set for each historical time period within that historical date.

[0074] For each known time period within the actual date, first determine the historical time periods corresponding to that known time period from all historical dates, then calculate the average of the historical actual total load values ​​of these corresponding historical time periods, and update the average value to the actual total load value of that known time period.

[0075] After calculating the true total load values ​​for all known time periods, these true total load values ​​for known time periods constitute the final original true total load data, completing the supplementation and update of the original true total load data.

[0076] This design can, to some extent, compensate for the lack of timeliness in the currently collected raw true total load data. When the raw true total load data does not meet the timeliness requirements of load forecasting, the raw true total load data for that day may not necessarily reflect the current load demand of the agency enterprise set. However, the average value of historical true total load data over a recent period can reflect the average load demand of the agency enterprise set over a recent period. The average load demand of the agency enterprise set over a period of time is usually relatively stable, so the representativeness and timeliness of the average load demand are usually better. In addition, the method of supplementing and updating data by using historical true total load data over a recent period can ensure that the subsequent model still has predictability even when the data on the reporting date is incomplete.

[0077] In some embodiments, feature extraction is performed on market load data, meteorological data, calendar data, and actual total load data to obtain a derived feature combination, including: extracting features from market load data, meteorological data, calendar data, and actual total load data to obtain lag features of actual total load data, market load features of market load data, meteorological features of meteorological data, index features of calendar data, lunar calendar features, and load ratio features of the agent enterprise set under the lunar calendar date; the index features include the holiday identifier of the date to be predicted; if the date to be predicted is within a holiday window, a derived feature combination is constructed based on the lunar calendar features, load ratio features, holiday identifier, and market load features; if the date to be predicted is not within a holiday window, a derived feature combination is constructed based on the market load features, meteorological features, index features, and lag features.

[0078] Specifically, feature extraction is performed on market load data to obtain market load characteristics, which include, but are not limited to, time-shift characteristics of market load forecasts and time-shift characteristics of measured market loads.

[0079] Similarly, feature extraction is performed on meteorological data to obtain meteorological features, which include, but are not limited to, aggregated features of meteorological data from multiple cities.

[0080] Optionally, the aggregation feature of meteorological data from multiple cities can be the mean of meteorological feature values ​​corresponding to multiple cities at the same time, or it can be a weighted average, median, or other statistics.

[0081] Similarly, feature extraction is performed on the actual total load data to obtain the lag characteristics of the actual total load data.

[0082] Similarly, feature extraction is performed on the calendar data to obtain the index features, lunar calendar features, and load ratio features of the agent enterprise set under the lunar calendar date.

[0083] Optionally, the indexing features of the calendar data include, but are not limited to, statutory holiday indexes, rest day indexes, quarterly indexes, monthly indexes, weekday indexes, and time-period indexes. The statutory holiday index includes the holiday identifiers for the days to be predicted.

[0084] Optionally, the lunar calendar features of the calendar data include, but are not limited to, lunar month features and lunar day features.

[0085] Considering that the company's power load is strongly correlated with holidays, in order to ensure the accuracy of subsequent model predictions, it is necessary to select representative features based on whether the predicted date falls before or after a holiday, and form a combination of derived features.

[0086] Specifically, if the forecast date is determined to be within a holiday window based on the holiday identifier of the forecast date, i.e., the forecast date is before or after a holiday, then the lunar calendar characteristics, load ratio characteristics, holiday identifiers, and market load characteristics are combined to form a derived feature combination.

[0087] If it is determined that the date to be predicted is not within a holiday window based on the holiday identifier of the date to be predicted, that is, the date to be predicted is not before or after a holiday, then the market load characteristics, meteorological characteristics, index characteristics, and lag characteristics are combined into a derived characteristic combination.

[0088] This feature combination method can avoid the strong holiday effect being diluted by conventional meteorological features or market load features, so that subsequent models can fully consider the impact of holidays. In addition, by introducing lunar calendar features and holiday windows, the model's load forecasting ability for special dates can be improved when the forecast date falls before or after a holiday.

[0089] In some embodiments, the load forecasting model includes a first load forecasting sub-model and a second load forecasting sub-model. Based on a combination of derived features, load forecasting is performed using a pre-trained load forecasting model to generate an average agent load forecast value for the agent set on the forecast date. This includes: dividing the forecast date into time periods to obtain a first time period set and a second time period set; the first and second time period sets include different time periods of the forecast date; inputting the combination of derived features into the first load forecasting sub-model to obtain a first average agent load forecast value for the agent set in the first time period set; inputting the combination of derived features into the second load forecasting sub-model to obtain a second average agent load forecast value for the agent set in the second time period set; wherein the average agent load forecast value includes both the first and second average agent load forecast values.

[0090] Specifically, such as Figure 4 As shown, the day to be predicted is divided into time periods by hours to obtain multiple time periods, and these time periods are further divided into two sets of time periods, namely the first time period set and the second time period set. The first time period set and the second time period set include different time periods of the day to be predicted.

[0091] Optionally, if the first time period set is a daytime time period set, then the second time period set is a non-daytime time period set; the daytime time period set includes multiple time periods during the daytime of the day to be predicted, and the non-daytime time period set includes multiple time periods at night of the day to be predicted.

[0092] Optionally, if the first time period set is a peak time period set, then the second time period set is an off-peak time period set; the peak time period set includes multiple time periods during the peak load period of the day to be predicted, and the off-peak time period set includes multiple time periods during the off-peak load period of the day to be predicted.

[0093] Furthermore, the derived feature combination is input into the first load forecasting sub-model to obtain the first average agency load forecast value of the first agency enterprise set in the first time period set output by the first load forecasting sub-model.

[0094] Similarly, the derived feature combination is input into the second load forecasting sub-model to obtain the second average agency load forecast value of the agency enterprise set in the second time period set.

[0095] The enterprise average agent load forecast includes the first enterprise average agent load forecast and the second enterprise average agent load forecast.

[0096] Optionally, the first enterprise average agent load forecast may include the enterprise average agent load of the agent set in each time period of the first time period set, and the second enterprise average agent load forecast may include the enterprise average agent load of the agent set in each time period of the second time period set.

[0097] Optionally, the day to be predicted can be divided into more time periods, but more load forecasting sub-models need to be trained at the same time to participate in load forecasting.

[0098] This time-segmented modeling and forecasting method allows for the separate forecasting of average agency loads for enterprises in different time periods with different load formation mechanisms. This helps agencies to anticipate the load situation of the agency group at different times on the forecast date.

[0099] In some embodiments, the method further includes: the first time period set is a daytime time period set, and the second time period set is a non-daytime time period set; or, the first time period set is a peak time period set, and the second time period set is a non-peak time period set.

[0100] In some embodiments, determining the total load forecast value of the agent set on the forecast date based on the enterprise's average agent load forecast value and the total number of agent enterprises includes: multiplying the first enterprise's average agent load forecast value by the total number of agent enterprises as the first total load forecast value of the agent set in the first time period; multiplying the second enterprise's average agent load forecast value by the total number of agent enterprises as the second total load forecast value of the agent set in the second time period; wherein the total load forecast value includes the first total load forecast value and the second total load forecast value.

[0101] Specifically, the product of the average agent load forecast of the first enterprise and the total number of agent enterprises is used as the first total load forecast of the agent enterprise set in the first time period.

[0102] Optionally, the first enterprise average agent load forecast value may include the enterprise average agent load of the agent enterprise set in each time period of the first time period set. The first total load of the agent enterprise set in each time period of the first time period set can be obtained by multiplying the enterprise average agent load of the agent enterprise set in each time period of the first time period set by the total number of agents. The first total load forecast value includes all first total loads.

[0103] Similarly, the product of the average agent load forecast of the second enterprise and the total number of agent enterprises is used as the second total load forecast of the agent enterprise set in the second time period.

[0104] Optionally, the second enterprise average agent load forecast value may include the enterprise average agent load of the agent enterprise set in each time period of the second time period set. The second total load of the agent enterprise set in each time period of the second time period set can be obtained by multiplying the enterprise average agent load of the agent enterprise set in each time period of the second time period set by the total number of agents. The second total load forecast value includes all second total loads.

[0105] The total load forecast includes a first total load forecast and a second total load forecast.

[0106] Optionally, if it is necessary to postpone the forecast date or extend the forecast period, the same model can be used to re-forecast after simultaneously adjusting the time shift step of the forecast date and various types of raw data (i.e., raw market load data, raw meteorological data, and raw real total load data).

[0107] Optionally, the day to be predicted can be any future day, such as tomorrow, the day after tomorrow, or even a day a week from now; multiple consecutive days to be predicted can form a prediction cycle, for example, the next three days can be regarded as a prediction cycle. The load prediction method of this embodiment is also applicable to load prediction of prediction cycles.

[0108] In some embodiments, before generating the average agent load forecast value for the agent enterprise set on the forecast date by performing load forecasting through a pre-trained load forecasting model based on derived feature combinations, the method further includes: determining a preset cutoff date and obtaining historical data before the preset cutoff date as a training sample set; the preset cutoff date is before the forecast date, and the training sample set includes sample market load data, sample meteorological data, sample calendar data, sample total load data of the sample agent enterprise set, and the total number of sample agent enterprises in the sample agent enterprise set; determining the label of the average agent load value of the sample enterprises based on the sample total load data and the total number of sample agent enterprises; performing time-shifting alignment processing and feature extraction on the sample market load data, sample meteorological data, sample calendar data, and sample total load data to obtain the sample derived feature combination; and pre-training the first initial load forecasting sub-model and the second initial load forecasting sub-model based on the sample derived feature combination and the label of the average agent load value of the sample enterprises to obtain the first load forecasting sub-model and the second load forecasting sub-model; wherein, the first load forecasting sub-model and the second load forecasting sub-model are LightGBM models.

[0109] Understandably, before using the model for load forecasting, it is necessary to train a first load forecasting sub-model and a second load forecasting sub-model.

[0110] To ensure the model's prediction accuracy and timeliness, this embodiment selects to use historical data of the agency group prior to the prediction date for model training.

[0111] Specifically, such as Figure 3 As shown, a preset deadline is selected before the date to be predicted, and historical data before the preset deadline is obtained as the training sample set.

[0112] The training sample set includes sample market load data, sample meteorological data, sample calendar data, sample total load data of the sample agent enterprise set, and the total number of sample agent enterprises in the sample agent enterprise set.

[0113] Furthermore, the ratio of the total sample load data to the total number of sample agent companies is calculated (i.e., the two are divided), and this ratio is used as the label for the average agent load value of the sample companies.

[0114] The purpose of constructing the average agent load value label for sample enterprises is to eliminate the direct impact of changes in the number of sample agent enterprises on the label size, so that the model can focus more on the average load pattern of individual sample agent enterprises, while improving the comparability between samples from different dates.

[0115] Furthermore, time-shifting alignment and feature extraction are performed on the sample market load data, sample meteorological data, sample calendar data, and sample total load data to obtain sample derived feature combinations.

[0116] Specifically, for the sample market load data, the date difference between the actual sample date and the sample forecast date is determined based on the sample calendar data, and the sample market load data is shifted over time according to the date difference to the sample forecast date.

[0117] Similarly, for sample meteorological data, the date difference between the actual date and the predicted date of the sample is determined based on the sample calendar data, and the sample meteorological data is shifted over time according to this date difference to the predicted date of the sample.

[0118] Similarly, for the total load data of the sample, the date difference between the actual date of the sample and the predicted date of the sample is determined based on the sample calendar data, and the total load data of the sample is shifted over time according to the date difference to the predicted date of the sample.

[0119] Then, feature extraction is performed on the sample market load data to obtain the sample market load characteristics, which include, but are not limited to, the time shift characteristics of the sample market load forecast and the time shift characteristics of the sample market load measured.

[0120] Similarly, feature extraction is performed on the sample meteorological data to obtain the sample meteorological features, which include, but are not limited to, the aggregated features of sample meteorological data from multiple cities.

[0121] Optionally, the aggregation feature of sample meteorological data from multiple cities can be the mean of the sample meteorological feature values ​​corresponding to multiple sample cities at the same time, or it can be the weighted average, median, or other statistics.

[0122] Similarly, feature extraction is performed on the total load data to obtain the sample lag features of the total load data.

[0123] Similarly, feature extraction is performed on the sample calendar data to obtain the sample index features, sample lunar calendar features, and sample load ratio features of the sample agency enterprise set under the lunar calendar date.

[0124] Optionally, the sample index features of the sample calendar data include, but are not limited to, sample statutory holiday index, sample rest day index, sample quarter index, sample month index, sample weekday index, and sample time period index. The sample statutory holiday index includes the holiday identifier of the sample prediction day.

[0125] Optionally, the sample lunar calendar features of the sample calendar data include, but are not limited to, sample lunar month features and sample lunar day features.

[0126] Optionally, the sample load ratio characteristic of the sample agency enterprise set under the lunar calendar date refers to the historical load ratio characteristic obtained from the statistics of historical samples under the same lunar calendar date.

[0127] Furthermore, if the sample prediction date is determined to be within a holiday window based on the holiday identifier of the sample prediction date, i.e., the sample prediction date is before or after a holiday, then the sample lunar calendar characteristics, sample load ratio characteristics, the holiday identifier of the sample prediction date, and sample market load characteristics are combined to form a sample derived feature combination.

[0128] If the sample prediction date is determined not to fall within a holiday window based on the holiday identifier of the sample prediction date, i.e., the sample prediction date is not before or after a holiday, then the sample market load characteristics, sample meteorological characteristics, sample index characteristics, and sample lag characteristics will be combined to form a sample derived characteristic combination.

[0129] Furthermore, the predicted days of the samples are divided into time periods to obtain the first time period set and the second time period set of the samples.

[0130] Furthermore, based on the combination of sample derived features and the label of the average agent load value of sample enterprises, the first initial load prediction sub-model is pre-trained to obtain the first load prediction sub-model.

[0131] Similarly, based on the combination of sample derived features and the label of the average agent load value of sample enterprises, the second initial load prediction sub-model is pre-trained to obtain the second load prediction sub-model.

[0132] Optionally, depending on the actual forecasting needs, more sample time periods can be further divided and more load forecasting sub-models can be trained.

[0133] Both the first and second load prediction sub-models are LightGBM models.

[0134] Optionally, for each forecast date, all historical data before its corresponding preset deadline can be selected as the training sample set, and multiple load forecasting sub-models can be retrained in a rolling manner. Rolling training is more in line with the daily reporting and updating business process of the electricity spot market.

[0135] Preferably, the preset deadline is 3 or 5 days earlier than the date to be predicted.

[0136] The load forecasting method provided in this embodiment reduces the impact of changes in the number of agent enterprises on model training and forecasting stability by constructing an average agent load value label for sample enterprises. At the same time, by training time-segmented load forecasting sub-models, the model's load forecasting accuracy and generalization ability in different time periods are improved. In addition, the rolling training and forecasting cycle extension methods improve the model's engineering feasibility in the actual electricity spot market.

[0137] For ease of understanding, a specific example of a load forecasting method is provided here for reference.

[0138] Specifically, assuming the current date is t, and the day to be predicted is t+1, the day to be predicted is divided into 48 time periods using half an hour as the basic time granularity, and these 48 time periods are further divided into a first time period set and a second time period set.

[0139] Furthermore, the latest raw market load data, raw meteorological data, calendar data, raw true total load data of the agent enterprise set, and the total number of agent enterprises in the agent enterprise set are collected.

[0140] Optionally, if the original true total load data is missing, it can be supplemented and updated using the average of the data from the most recent 7 days.

[0141] Furthermore, based on calendar data, the original market load data, original meteorological data, and original true total load data are time-shifted and aligned to obtain market load data, meteorological data, and true total load data aligned with the date to be predicted.

[0142] For example, if the latest raw market load data corresponds to the actual date t-2, the raw market load data can be shifted forward by 3 days to participate in subsequent load forecasting.

[0143] Furthermore, if the latest original true total load data is the data from day t-3, then historical data prior to day t-3 are collected to construct a training sample set.

[0144] The training sample set includes sample market load data, sample meteorological data, sample calendar data, sample total load data of the sample agent enterprise set, and the total number of sample agent enterprises in the sample agent enterprise set.

[0145] Furthermore, the ratio of the total sample load data to the total number of sample agent companies is calculated (i.e., the two are divided), and this ratio is used as the label for the average agent load value of the sample companies.

[0146] Furthermore, by utilizing sample market load data, sample meteorological data, sample calendar data, sample total load data, and sample enterprise average agent load value labels, the first load forecasting sub-model and the second load forecasting sub-model are trained.

[0147] Furthermore, feature extraction is performed on market load data, meteorological data, calendar data, and actual total load data to obtain the lag characteristics of actual total load data, market load characteristics of market load data, meteorological characteristics of meteorological data, index characteristics of calendar data, lunar calendar characteristics, and load ratio characteristics of the agency enterprise set under lunar calendar dates.

[0148] If the forecast date falls within a holiday window, a combination of derived features is constructed based on lunar calendar characteristics, load ratio characteristics, holiday identifiers, and market load characteristics.

[0149] If the forecast date does not fall within a holiday window, a combination of derived features is constructed based on market load characteristics, meteorological characteristics, index characteristics, and lag characteristics.

[0150] Furthermore, the derived feature combination is input into the first load forecasting sub-model to obtain the first average agent load forecast value of the agent enterprise set in the first time period set output by the first load forecasting sub-model; similarly, the derived feature combination is input into the second load forecasting sub-model to obtain the second average agent load forecast value of the agent enterprise set in the second time period set output by the second load forecasting sub-model.

[0151] The forecast value of the average agency load of the agency enterprises on the forecast date includes the forecast value of the first enterprise's average agency load and the forecast value of the second enterprise's average agency load.

[0152] Furthermore, the product of the average agent load forecast of the first enterprise and the total number of agent enterprises is used as the first total load forecast of the agent enterprise set in the first time period.

[0153] Optionally, the first enterprise average agent load forecast value may include the enterprise average agent load of the agent enterprise set in each time period of the first time period set. The first total load of the agent enterprise set in each time period of the first time period set can be obtained by multiplying the enterprise average agent load of the agent enterprise set in each time period of the first time period set by the total number of agents. The first total load forecast value includes all first total loads.

[0154] Similarly, the product of the average agent load forecast of the second enterprise and the total number of agent enterprises is used as the second total load forecast of the agent enterprise set in the second time period.

[0155] Optionally, the second enterprise average agent load forecast value may include the enterprise average agent load of the agent enterprise set in each time period of the second time period set. The second total load of the agent enterprise set in each time period of the second time period set can be obtained by multiplying the enterprise average agent load of the agent enterprise set in each time period of the second time period set by the total number of agents. The second total load forecast value includes all second total loads.

[0156] The total load forecast for the agency group on the forecast date includes a first total load forecast and a second total load forecast.

[0157] Optionally, if it is necessary to extend the forecast date to t+2 days, t+3 days, etc., the same model can be used to re-forecast after simultaneously adjusting the time shift step of the forecast date and various types of raw data (i.e., raw market load data, raw meteorological data, and raw true total load data).

[0158] This invention also provides a load forecasting device for a group of electricity spot market brokerage firms. Please refer to [link / reference]. Figure 5 , Figure 5 This is a schematic diagram of the load forecasting device for a set of electricity spot market agents provided by the present invention. In this embodiment, the load forecasting device for a set of electricity spot market agents includes an acquisition module 510, a time-shifting alignment module 520, a feature extraction module 530, a load forecasting module 540, and a result recovery module 550.

[0159] The acquisition module 510 is used to acquire the original market load data, original meteorological data, calendar data, original true total load data of the agent enterprise set, and the total number of agent enterprises in the agent enterprise set before the forecast date.

[0160] The set of agents includes at least two agents.

[0161] The time-shifting alignment module 520 is used to perform time-shifting alignment processing on the original market load data, original meteorological data and original true total load data based on calendar data, so as to obtain market load data, meteorological data and true total load data aligned with the date to be predicted.

[0162] The feature extraction module 530 is used to extract features from market load data, meteorological data, calendar data and actual total load data to obtain derived feature combinations.

[0163] The load forecasting module 540 is used to perform load forecasting based on derived feature combinations and a pre-trained load forecasting model, generating the average agency load forecast value of the agency enterprise set on the forecast date.

[0164] The result recovery module 550 is used to determine the total load forecast value of the agent enterprise set on the forecast date based on the enterprise's average agent load forecast value and the total number of agent enterprises.

[0165] In some embodiments, the original true total load data includes the true total load values ​​of the agency set over multiple known time periods. Before performing time-shifting and alignment processing on the original market load data, original meteorological data, and original true total load data based on calendar data to obtain market load data, meteorological data, and true total load data aligned with the forecast date, the method further includes: determining whether the original true total load data meets the timeliness requirements of the load forecasting business; if the original true total load data does not meet the timeliness requirements of the load forecasting business, then obtaining the historical true total load data of the agency set over the most recent historical dates; a historical date includes multiple historical time periods, a historical time period corresponds to a known time period, and the historical true total load data within a historical date includes the historical true total load value of the agency set in each historical time period; the original true total load data is supplemented and updated based on the historical true total load data of the most recent historical dates; wherein, the true total load value of a known time period is the average of the historical true total load values ​​of the historical time periods corresponding to the most recent historical dates.

[0166] In some embodiments, feature extraction is performed on market load data, meteorological data, calendar data, and actual total load data to obtain a derived feature combination, including: extracting features from market load data, meteorological data, calendar data, and actual total load data to obtain lag features of actual total load data, market load features of market load data, meteorological features of meteorological data, index features of calendar data, lunar calendar features, and load ratio features of the agent enterprise set under the lunar calendar date; the index features include the holiday identifier of the date to be predicted; if the date to be predicted is within a holiday window, a derived feature combination is constructed based on the lunar calendar features, load ratio features, holiday identifier, and market load features; if the date to be predicted is not within a holiday window, a derived feature combination is constructed based on the market load features, meteorological features, index features, and lag features.

[0167] In some embodiments, the load forecasting model includes a first load forecasting sub-model and a second load forecasting sub-model. Based on a combination of derived features, load forecasting is performed using a pre-trained load forecasting model to generate an average agent load forecast value for the agent set on the forecast date. This includes: dividing the forecast date into time periods to obtain a first time period set and a second time period set; the first and second time period sets include different time periods of the forecast date; inputting the combination of derived features into the first load forecasting sub-model to obtain a first average agent load forecast value for the agent set in the first time period set; inputting the combination of derived features into the second load forecasting sub-model to obtain a second average agent load forecast value for the agent set in the second time period set; wherein the average agent load forecast value includes both the first and second average agent load forecast values.

[0168] In some embodiments, the method further includes: the first time period set is a daytime time period set, and the second time period set is a non-daytime time period set; or, the first time period set is a peak time period set, and the second time period set is a non-peak time period set.

[0169] In some embodiments, determining the total load forecast value of the agent set on the forecast date based on the enterprise's average agent load forecast value and the total number of agent enterprises includes: multiplying the first enterprise's average agent load forecast value by the total number of agent enterprises as the first total load forecast value of the agent set in the first time period; multiplying the second enterprise's average agent load forecast value by the total number of agent enterprises as the second total load forecast value of the agent set in the second time period; wherein the total load forecast value includes the first total load forecast value and the second total load forecast value.

[0170] In some embodiments, before generating the average agent load forecast value for the agent enterprise set on the forecast date by performing load forecasting through a pre-trained load forecasting model based on derived feature combinations, the method further includes: determining a preset cutoff date and obtaining historical data before the preset cutoff date as a training sample set; the preset cutoff date is before the forecast date, and the training sample set includes sample market load data, sample meteorological data, sample calendar data, sample total load data of the sample agent enterprise set, and the total number of sample agent enterprises in the sample agent enterprise set; determining the label of the average agent load value of the sample enterprises based on the sample total load data and the total number of sample agent enterprises; performing time-shifting alignment processing and feature extraction on the sample market load data, sample meteorological data, sample calendar data, and sample total load data to obtain the sample derived feature combination; and pre-training the first initial load forecasting sub-model and the second initial load forecasting sub-model based on the sample derived feature combination and the label of the average agent load value of the sample enterprises to obtain the first load forecasting sub-model and the second load forecasting sub-model; wherein, the first load forecasting sub-model and the second load forecasting sub-model are LightGBM models.

[0171] The present invention also provides an electronic device. Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logic instructions in the memory 630 to execute a load forecasting method for a set of electricity spot market agents.

[0172] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the load forecasting method for a set of electricity spot market agents provided by the above methods.

[0174] The present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the load forecasting method for a set of electricity spot market agents provided by the above methods.

[0175] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A load forecasting method for a group of electricity spot market agents, characterized in that, include: Obtain the original market load data, original meteorological data, calendar data, original true total load data of the agent enterprise set, and the total number of agents in the agent enterprise set before the forecast date; The set of agency companies includes at least two agency companies; Based on the calendar data, the original market load data, the original meteorological data, and the original true total load data are subjected to time-shifting and alignment processing to obtain market load data, meteorological data, and true total load data aligned with the date to be predicted. Feature extraction is performed on the market load data, the meteorological data, the calendar data, and the actual total load data to obtain derived feature combinations; Based on the derived feature combination, load forecasting is performed through a pre-trained load forecasting model to generate the average agency load forecast value of the agency enterprise set on the forecast date. Based on the average agent load forecast value of the enterprise and the total number of agent enterprises, the total load forecast value of the agent enterprise set on the forecast date is determined.

2. The load forecasting method for a set of electricity spot market agents according to claim 1, characterized in that, The original true total load data includes the true total load values ​​of the agency group over multiple known time periods; Before performing time-shifting and alignment processing on the original market load data, the original meteorological data, and the original actual total load data based on the calendar data to obtain market load data, meteorological data, and actual total load data aligned with the date to be predicted, the process further includes: Determine whether the original true total load data meets the timeliness requirements of the load forecasting service; If the original true total load data does not meet the timeliness requirements of the load forecasting service, then the historical true total load data of the agent enterprise set over the most recent historical dates is obtained; one historical date includes multiple historical time periods, one historical time period corresponds to one known time period, and the historical true total load data within one historical date includes the historical true total load value of the agent enterprise set in each historical time period; Based on the historical true total load data from the most recent historical dates, the original true total load data is supplemented and updated. Wherein, the true total load value for a known time period is the average of the historical true total load values ​​for the historical time periods corresponding to the most recent several historical dates.

3. The load forecasting method for a set of electricity spot market agents according to claim 1, characterized in that, The step of extracting features from the market load data, the meteorological data, the calendar data, and the actual total load data to obtain derived feature combinations includes: Feature extraction is performed on the market load data, the meteorological data, the calendar data, and the actual total load data to obtain the lag features of the actual total load data, the market load features of the market load data, the meteorological features of the meteorological data, the index features of the calendar data, the lunar calendar features, and the load ratio features of the agency enterprise set under the lunar calendar date; the index features include the holiday identifier of the date to be predicted; If the date to be predicted falls within a holiday window, then the derived feature combination is constructed based on the lunar calendar characteristics, the load ratio characteristics, the holiday identifiers, and the market load characteristics. If the predicted date does not fall within the holiday window, then the derived feature combination is constructed based on the market load characteristics, the meteorological characteristics, the index characteristics, and the lag characteristics.

4. The load forecasting method for a set of electricity spot market agents according to claim 1, characterized in that, The load forecasting model includes a first load forecasting sub-model and a second load forecasting sub-model; The process of generating the average agency load forecast value for the agency enterprise set on the forecast date based on the derived feature combination and a pre-trained load forecast model includes: The day to be predicted is divided into time periods to obtain a first time period set and a second time period set; the first time period set and the second time period set include different time periods of the day to be predicted; The derived feature combination is input into the first load forecasting sub-model to obtain the first enterprise average agent load forecast value of the agent enterprise set in the first time period set, which is output by the first load forecasting sub-model. The derived feature combination is input into the second load forecasting sub-model to obtain the second enterprise average agent load forecast value of the agent enterprise set in the second time period set, which is output by the second load forecasting sub-model. The enterprise average agent load forecast includes the first enterprise average agent load forecast and the second enterprise average agent load forecast.

5. The load forecasting method for a set of electricity spot market agents according to claim 4, characterized in that, Also includes: The first time period set is a daytime time period set, and the second time period set is a non-daytime time period set; or, the first time period set is a peak time period set, and the second time period set is a non-peak time period set.

6. The load forecasting method for a set of electricity spot market agents according to claim 4, characterized in that, The step of determining the total load forecast value of the agent group on the forecast date based on the average agent load forecast value of the enterprise and the total number of agent enterprises includes: The product of the first enterprise's average agent load forecast and the total number of agent enterprises is used as the first total load forecast of the agent enterprise set in the first time period set; The product of the second enterprise's average agent load forecast and the total number of agent enterprises is used as the second total load forecast of the agent enterprise set in the second time period set; The total load forecast includes the first total load forecast and the second total load forecast.

7. The load forecasting method for a set of electricity spot market agents according to claim 4, characterized in that, The step of generating the agent enterprise set before generating the average agent load forecast value for the enterprise set on the forecast date by performing load forecasting through a pre-trained load forecasting model based on the derived feature combination further includes: A preset deadline is determined, and historical data prior to the preset deadline is obtained as a training sample set; the preset deadline is prior to the date to be predicted, and the training sample set includes sample market load data, sample meteorological data, sample calendar data, sample total load data of the sample agent enterprise set, and the total number of sample agent enterprises in the sample agent enterprise set; Based on the total sample load data and the total number of sample agent companies, determine the average agent load value label for the sample companies; The sample market load data, the sample meteorological data, the sample calendar data, and the sample total load data are subjected to time-shifting alignment processing and feature extraction to obtain a sample derived feature combination; Based on the sample derived feature combination and the sample enterprise average agent load value label, the first initial load prediction sub-model and the second initial load prediction sub-model are pre-trained to obtain the first load prediction sub-model and the second load prediction model; The first load forecasting sub-model and the second load forecasting sub-model are LightGBM models.

8. A load forecasting device for a group of electricity spot market agents, characterized in that, include: The acquisition module is used to acquire the original market load data, original meteorological data, calendar data, original true total load data of the agent enterprise set and the total number of agent enterprises in the agent enterprise set before the forecast date. The set of agency companies includes at least two agency companies; The time-shifting alignment module is used to perform time-shifting alignment processing on the original market load data, the original meteorological data, and the original true total load data based on the calendar data, to obtain market load data, meteorological data, and true total load data aligned with the date to be predicted; The feature extraction module is used to extract features from the market load data, the meteorological data, the calendar data, and the actual total load data to obtain a derived feature combination. The load forecasting module is used to perform load forecasting based on the derived feature combination and a pre-trained load forecasting model to generate the average agency load forecast value of the agency enterprise set on the forecast date. The result recovery module is used to determine the total load forecast value of the agent enterprise set on the forecast date based on the average agent load forecast value of the enterprise and the total number of agent enterprises.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the load forecasting method for a set of electricity spot market agents as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the load forecasting method for a set of electricity spot market agents as described in any one of claims 1 to 7.

Images