Methods, equipment and storage media for electricity load forecasting

By using a multi-model prediction system that combines features such as date and weekday attributes to deeply integrate electricity load data, the problem of accuracy in predicting electricity load in different areas has been solved, thereby improving the supply and demand balance of the power system and the stability of the power grid.

CN122869305BActive Publication Date: 2026-11-10SHENZHEN GUORUIXIE CHUANG ENERGY STORAGE TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611348943.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-10
Estimated Expiration
2046-09-02

AI Technical Summary

Technical Problem

Existing technologies cannot accurately predict the electricity load of a region, resulting in insufficient accuracy in power system supply and demand balance, grid stability, and power dispatch decisions.

Method used

By constructing a multi-model prediction system, the system utilizes features such as date attributes, weekday attributes, historical electricity load data, rate of change sequences, and mean sequences, and combines models such as support vector machines and deep neural networks to perform deep fusion and weighted processing of multi-source feature information to predict the electricity load of the target day.

Benefits of technology

It improves the accuracy and stability of electricity load forecasting, provides reliable data support for power dispatching and demand-side management, and adapts to complex and ever-changing electricity consumption scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122869305B_ABST
    Figure CN122869305B_ABST
Patent Text Reader

Abstract

This application relates to the field of power system technology and discloses a method, device, and storage medium for predicting electricity load. The method is used to predict the electricity load of an area on a target day, including: determining a first sub-feature based on the date and weekday attributes of the target day; determining a second sub-feature based on the target day and a statutory holiday earlier than the target day and closest to the target day; determining a rate of change sequence, a mean sequence, and a third sub-feature based on multiple sets of historical electricity load data sequences of the area; determining a first feature based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, and the third sub-feature; determining the second feature corresponding to each sub-time period; inputting the first feature into a first model and a second model respectively, inputting each second feature into a corresponding third model, and finally determining the electricity load prediction data based on the model output. This application can accurately predict the electricity load of an area.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power system technology, specifically to a method, device, and storage medium for predicting electricity load. Background Technology

[0002] Electricity load forecasting is a prerequisite for optimizing the operation and scientific management of the power system. In a power system, accurate forecasting of electricity load in a given area not only effectively balances supply and demand and ensures the stability of grid frequency and voltage, thereby maintaining the safe and reliable operation of the system, but also serves as a crucial basis for power supply companies to formulate economically sound power generation plans, implement power dispatch decisions, and plan the construction of power grid facilities.

[0003] Electricity load is influenced by a variety of factors, including social factors and holidays, resulting in fluctuations in its patterns. This makes electricity load forecasting complex, as it is difficult to accurately capture the dynamic mapping relationship between various factors and electricity load using a single model, thus affecting the accuracy of the forecast results. How to accurately predict the electricity load of a region is a problem that needs to be solved. Summary of the Invention

[0004] In view of the above problems, this application provides an electricity load prediction method, device and storage medium to solve the problem that the prior art cannot accurately predict the electricity load of an electricity consumption area.

[0005] According to one aspect of the embodiments of this application, an electricity load forecasting method is provided for forecasting the electricity load of an electricity consumption area on a target day. The method includes: determining a date attribute and a weekday attribute of the target day, and determining a first sub-feature based on the date attribute and the weekday attribute of the target day, wherein the date attribute includes weekdays, weekends, and public holidays, and the weekday attribute includes Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday; determining the start date and end date of a public holiday earlier than the target day and closest to the target day, and determining the first sub-feature based on a first interval number of days between the holiday start date and the target day and the holiday... The second sub-feature is determined by the second interval number of days between the end date and the target date; multiple sets of historical electricity load data sequences corresponding to multiple historical days of the electricity consumption area are obtained, wherein each set of historical electricity load data sequences includes historical electricity load data corresponding to multiple sub-periods; multiple sets of change rate sequences and one set of mean sequence are determined based on the multiple sets of historical electricity load data sequences, wherein each set of change rate sequences includes the change rate between the historical electricity load data corresponding to the same sub-period of two sets of historical electricity load data sequences, and the mean sequence includes the mean of the historical electricity load data corresponding to the same sub-period of multiple sets of historical electricity load data sequences; based on the multiple sets of historical The third sub-feature is determined by the mean, peak, and trough values ​​of the electricity load data sequence; based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, and the third sub-feature, a first feature is determined; the first feature is input into a trained first model and a trained second model corresponding to the date attribute of the target day, respectively, to obtain a first electricity load prediction data sequence output by the trained first model and a second electricity load prediction data sequence output by the trained second model, wherein the first electricity load prediction data sequence includes multiple first electricity loads corresponding to multiple sub-time periods of the target day. The second electricity load forecast data sequence includes multiple second electricity load forecast data corresponding to multiple sub-time periods of the target day. The training data used to train the trained second model is determined based on the electricity load data of the electricity consumption area for multiple days that are earlier than the day with the same date attribute and adjacent to the day with the same date attribute. The day with the same date attribute is a day with the same date attribute as the target day. Based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-time period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature, the second feature corresponding to each sub-time period is determined.The second feature corresponding to each of the sub-time periods is input into the trained third model corresponding to each sub-time period to obtain a third electricity load prediction data sequence. The third electricity load prediction data sequence includes multiple third electricity load prediction data corresponding to multiple sub-time periods of the target day, and each third electricity load prediction data is output by the trained third model corresponding to its corresponding sub-time period. The first electricity load prediction data sequence, the second electricity load prediction data sequence, and the third electricity load prediction data sequence corresponding to the same sub-time period are weighted and fused to obtain an electricity load prediction data sequence. The electricity load prediction data sequence includes electricity load prediction data corresponding to multiple sub-time periods of the target day.

[0006] In an optional manner, before determining the first feature based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, and the third sub-feature, the method further includes: performing one or more steps of steps S11, S12, S13, and S14 to obtain relevant data, wherein step S11 is: determining multiple ratio sequences based on the multiple sets of historical electricity load data sequences, each set of ratio sequences including the ratio of historical electricity load data corresponding to the same sub-period in two sets of historical electricity load data sequences; step S12 is: from the... From multiple sets of historical electricity load data sequences, a target historical electricity load data sequence is determined that corresponds to the day closest to the target day and has the same date attribute as the target day. A third feature is determined based on this target historical electricity load data sequence. Step S13 involves determining a fourth feature based on the date attributes of the n days before and m days after the target day, where n and m are both positive integers. Step S14 involves determining a fifth feature based on meteorological forecast data for the target day and the month to which the target day belongs. The relevant data includes one or more of the multiple ratio sequences, the third feature, the fourth feature, and the fifth feature. The step based on the first sub-feature... The determination of a first feature, based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, and the third sub-feature, includes: determining the first feature based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, the third sub-feature, and the relevant data; and determining a second feature corresponding to each sub-period based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature, includes: If the relevant data includes the multiple ratio sequences, then the second feature corresponding to each sub-period is determined based on the historical electricity load data in the multiple historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple rate of change sequences, the mean in the mean sequence, the third sub-feature, and the multiple ratio sequences; if the relevant data does not include the multiple ratio sequences, then the second feature corresponding to each sub-period is determined based on the historical electricity load data in the multiple historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple rate of change sequences, the mean in the mean sequence, and the third sub-feature.

[0007] In one optional embodiment, the multiple sets of ratio sequences include a first set of ratio sequences and a second set of ratio sequences. Step S11 includes: determining from the multiple sets of historical electricity load data sequences a first historical electricity load data sequence corresponding to the day before the target date, a second historical electricity load data sequence corresponding to the two days before the target date, a seventh historical electricity load data sequence corresponding to the seven days before the target date, and a fourteenth historical electricity load data sequence corresponding to the 14 days before the target date; determining the ratio of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the second historical electricity load data sequence to obtain the first set of ratio sequences; and determining the ratio of the historical electricity load data corresponding to the same sub-period of the seventh historical electricity load data sequence and the fourteenth historical electricity load data sequence to obtain the second set of ratio sequences.

[0008] In one optional manner, the multiple sets of rate of change sequences include a first set of rate of change sequences and a second set of rate of change sequences; the step of determining multiple sets of rate of change sequences and a set of mean sequences based on the multiple sets of historical electricity load data sequences includes: determining from the multiple sets of historical electricity load data sequences a first historical electricity load data sequence corresponding to the day before the target day, a second historical electricity load data sequence corresponding to the two days before the target day, a third historical electricity load data sequence corresponding to the three days before the target day, a fourth historical electricity load data sequence corresponding to the four days before the target day, a fifth historical electricity load data sequence corresponding to the five days before the target day, a sixth historical electricity load data sequence corresponding to the six days before the target day, and a seventh historical electricity load data sequence corresponding to the seven days before the target day; determining The rate of change of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the seventh historical electricity load data sequence is used to obtain the first set of rate of change sequences; the rate of change of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the second historical electricity load data sequence is determined to obtain the second set of rate of change sequences; the average value of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence, the second historical electricity load data sequence, the third historical electricity load data sequence, the fourth historical electricity load data sequence, the fifth historical electricity load data sequence, the sixth historical electricity load data sequence, and the seventh historical electricity load data sequence is determined to obtain the mean value sequence.

[0009] In one optional approach, determining the third sub-feature based on the mean, peak, and valley values ​​of multiple sets of historical electricity load data sequences includes: determining a first historical electricity load data sequence corresponding to the day before the target day, a second historical electricity load data sequence corresponding to the two days before the target day, and a third historical electricity load data sequence corresponding to the three days before the target day from the multiple sets of historical electricity load data sequences; determining the average value of all historical electricity load data in the first, second, and third historical electricity load data sequences to obtain the recent average electricity load; determining the peak and valley values ​​from all historical electricity load data in the first, second, and third historical electricity load data sequences to obtain the recent peak and valley values ​​of electricity load; and using the recent average electricity load, the recent peak electricity load, and the recent valley values ​​of electricity load as the third sub-feature.

[0010] In one optional approach, the first weight corresponding to the first electricity load forecast data, the second weight corresponding to the second electricity load forecast data, and the third weight corresponding to the third electricity load forecast data are predetermined through the following steps: obtaining multiple sets of electricity load training data sequences corresponding to multiple historical days of the electricity consumption area, wherein each set of electricity load training data sequences includes electricity load training data corresponding to multiple sub-time periods; dividing the multiple sets of electricity load training data sequences into multiple datasets, each dataset including multiple sets of electricity load training data sequences; determining k sets of first training features and k sets of second training features based on the multiple sets of datasets, and inputting the k sets of first training features into k first models and k second models respectively, and inputting the k sets of second training features into k third models respectively, to obtain k sets of first prediction data, k sets of second prediction data, and k sets of third prediction data, wherein k is a positive integer and k > 1; determining the first weight, the second weight, and the third weight based on the k sets of first prediction data, the k sets of second prediction data, the k sets of third prediction data, and the multiple sets of electricity load training data sequences.

[0011] In one optional approach, the step of determining k groups of first training features and k groups of second training features based on multiple sets of the datasets, and inputting the k groups of first training features into k first models and k second models respectively, and inputting the k groups of second training features into k third models respectively, to obtain k groups of first prediction data, k groups of second prediction data, and k groups of third prediction data, includes: Step S21: Determining the r-th group of first training features and the r-th group of second training features in the k groups of first training features based on the first r sets of the datasets in the multiple sets of the datasets, and inputting the k groups of second training features into k first models and k second models respectively, to obtain k groups of first prediction data, k groups of second prediction data, and k groups of third prediction data, includes: Step S21: Determining the r-th group of first training features in the k groups of first training features and the r-th group of second training features in the k groups of second training features based on the first r sets of the datasets in the multiple sets of the datasets, and inputting the k groups of first training features into k first models and k second models respectively, to obtain k groups of first prediction data, k groups of second prediction data, and k groups of third prediction data, including the first prediction data, k groups of second prediction data, and the second ... training features and k groups of second training features based on the first r sets of the datasets in the multiple sets of the datasets, and inputting the first prediction data into k first models and k second models respectively, to obtain k groups of first prediction data, k groups of second prediction data, and the second prediction data, including the first prediction data The r-th group of first training features is input into the r-th first model among the k first models and the r-th second model among the k second models, and the r-th group of second training features is input into the r-th third model among the k third models, to obtain the r-th group of first predicted data among the k first predicted data, the r-th group of second predicted data among the k second predicted data, and the r-th group of third predicted data among the k third predicted data, where r is a positive integer and the initial value of r is 1; Step S22: If r < k, then increment r and go to step S21.

[0012] In one optional approach, determining the first weight, the second weight, and the third weight based on the k sets of first predicted data, the k sets of second predicted data, the k sets of third predicted data, and the multiple sets of electricity load training data sequences includes: constructing a matrix H, wherein the first column of matrix H represents the k sets of first predicted data, the second column of matrix H represents the k sets of second predicted data, and the third column of matrix H represents the k sets of third predicted data; constructing a vector Y based on the multiple sets of electricity load training data sequences, wherein vector Y includes one column of data, and each row of vector Y represents the electricity load training data for the sub-period corresponding to each row of predicted data in matrix H; and determining the weights that satisfy the formula... The matrix w at time, where, ;Will , and The weights are respectively determined as the first weight, the second weight, and the third weight.

[0013] According to another aspect of the embodiments of this application, an electricity load forecasting device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the electricity load forecasting method as described above.

[0014] According to another aspect of the embodiments of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the power load prediction method as described above.

[0015] In this embodiment, the above-described method fully leverages the global trend capture capability of the first model, the specific date type adaptability of the second model, and the refined time-period fitting capability of the third model, achieving deep fusion and complementarity of multi-source feature information and multi-scale change patterns. Ultimately, the electricity load forecast data sequence obtained through weighted fusion effectively overcomes the limitations of a single model in the face of complex and ever-changing electricity consumption scenarios, thereby improving the prediction accuracy and stability of the target day's electricity load and providing reliable data support for power dispatching, demand-side management, and electricity cost control in the electricity consumption area.

[0016] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0017] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the electricity load forecasting method provided in an embodiment of this application is shown. Figure 2 The embodiments provided in this application are shown. Figure 1 A flowchart illustrating the sub-steps of step S104; Figure 3 The embodiments provided in this application are shown. Figure 1 A flowchart illustrating the sub-steps of step S105. Figure 4 A flowchart illustrating the method for determining the first weight, the second weight, and the third weight provided in an embodiment of this application is shown. Figure 5 A schematic diagram of the electrical load prediction device provided in an embodiment of this application is shown. Detailed Implementation

[0018] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein.

[0019] Electricity load is influenced by a variety of factors, including social conditions and holidays, resulting in fluctuations in its patterns. For example, during holidays, industrial production stagnates, and electricity load data is typically lower than on weekdays. Traditional methods often use only one model to predict electricity load, but this approach cannot adapt to complex and ever-changing scenarios, struggles to capture the nonlinear characteristics of electricity load data, and ultimately fails to meet prediction accuracy requirements.

[0020] Other traditional methods extrapolate the next day's curve using statistical models such as historical averages, exponential smoothing, and autoregressive moving averages (ARIMA); or they first select similar days (with similar temperatures and weekday types) from a historical database, and then shift or weight the load curves of these similar days as the prediction result. These methods are simple to implement and highly interpretable, but they are difficult to simultaneously characterize nonlinear factors such as holiday work schedule adjustments, production scheduling changes, and shift changes, and they are less adaptable to sudden changes in the curve shape of industrial and commercial users.

[0021] Since electricity load is influenced by multiple factors, using different models to capture both local variations and overall trends in the data for prediction can effectively improve prediction accuracy. For example, considering that weekdays, holidays, and weekends are key factors affecting electricity load, training models with differentiated data to specifically predict electricity load for different types of days will further improve the accuracy of prediction results.

[0022] Based on this, this application proposes an electricity load forecasting method for predicting the electricity load of a target day in an electricity consumption area. The method determines whether the target day is a weekday, weekend, or public holiday, and which day of the week it falls on. Based on this determination, a first sub-feature is identified. Then, a second sub-feature is identified based on the target day and a public holiday earlier than or closest to the target day. Furthermore, a change rate sequence, a mean sequence, and a third sub-feature are determined based on multiple sets of historical electricity load data sequences corresponding to several historical days in the target area. Based on the first sub-feature, the second sub-feature, the historical electricity load data sequences, the change rate sequence, the mean sequence, and the third sub-feature, the first feature is determined, and the second feature corresponding to each sub-time period of the day is determined. Finally, the first feature is input into a first model and a second model, and each second feature is input into a third model corresponding to each sub-time period. The electricity load forecast data is determined based on the output results of all models.

[0023] The electricity load forecasting method provided in this application will be described in detail below with reference to the accompanying drawings. Figure 1A flowchart illustrating the electricity load forecasting method provided in this application embodiment is shown. This method is executed by an electricity load forecasting device, which may be a device including one or more processors, such as a server, computer, touchscreen phone, smartphone, tablet computer, or other electronic device. The following description will use the forecasting of the electricity load of a target day in an electricity consumption area as an example. Here, the electricity consumption area refers to the area where electricity load forecasting needs to be performed, such as an industrial park. Figure 1 As shown, the method includes the following steps S101 to S110.

[0024] Step S101: Determine the date attribute and weekday attribute of the target day, and determine the first sub-feature based on the date attribute and weekday attribute of the target day.

[0025] The date attribute includes weekdays, weekends, and public holidays, while the weekday attribute includes Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday. Taking a target day as a weekday, Thursday, as an example, the target day's date attribute is weekday, and its weekday attribute is Thursday. It should be noted that if the target day is a day off in lieu of a workday, the corresponding date attribute will be weekday.

[0026] In this embodiment, the first sub-feature is a 10-dimensional vector, with its first to tenth dimensions corresponding to: whether it is a weekday, whether it is a weekend, whether it is a public holiday, Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday, respectively. When determining the first sub-feature, the values ​​of the dimensions corresponding to the date and weekday attributes of the target day are specifically set to 1, while the values ​​of the remaining dimensions are all set to 0. For example, if the target day is a weekday or Thursday, then the values ​​corresponding to the weekday and Thursday dimensions are both 1, and the values ​​of the remaining dimensions are all 0, resulting in the first sub-feature being [1, 0, 0, 0, 0, 0, 1, 0, 0, 0].

[0027] As mentioned earlier, electricity consumption patterns differ between weekdays, weekends, and public holidays. For example, industrial parks experience higher electricity loads on weekdays and significantly lower loads on holidays. In this step, the date and weekday attributes of the target day are determined, and a first sub-feature is constructed based on these attributes. This allows for accurate prediction of electricity load based on the first sub-feature.

[0028] Step S102: Determine the start date and end date of the holiday that is earlier than the target date and closest to the target date, and determine the second sub-feature based on the first interval between the holiday start date and the target date and the second interval between the holiday end date and the target date.

[0029] Considering that there are usually special electricity consumption patterns during and around statutory holidays (such as reduced production during holidays leading to lower load, or resumption of work after holidays leading to a rebound in load), this step determines the second sub-feature based on the first interval number of days and the second interval number of days. This allows the model to identify the potential impact of the most recent statutory holiday on the target day when the second sub-feature is subsequently input into the model to predict electricity load, thereby further improving the accuracy of electricity load prediction.

[0030] The second sub-feature is a two-dimensional vector. In this step, the first interval number of days and the second interval number of days are directly used as the two dimensions of this two-dimensional vector to obtain the second sub-feature. For example, if the target date is November 6, 2025, then the start date of the statutory holiday earlier than the target date and closest to the target date is October 1, 2025, and the end date of the statutory holiday is October 8, 2025. Therefore, the first interval number of days is 36, the second interval number of days is 29, and the corresponding second sub-feature is [36, 29].

[0031] Step S103: Obtain multiple sets of historical electricity load data sequences corresponding to multiple historical days of the electricity consumption area.

[0032] A significant component of electricity bills for industrial and commercial users is the "demand charge," which is based on the user's highest average power consumption within a billing cycle (usually one month). The calculation window for this "average power consumption" is explicitly defined as 15 minutes in the power grid company's rules. In other words, the power grid company continuously monitors users' electricity consumption, dividing the month's electricity usage into numerous consecutive, non-overlapping 15-minute segments. It calculates the average power consumption within each segment, then identifies the highest average power consumption among all these calculations. This highest average power consumption is the "maximum demand" for that month, and it serves as the basis for charging the demand charge. Furthermore, modern smart meters typically calculate users' electricity consumption in 15-minute cycles.

[0033] Therefore, in this step, a set of historical electricity load data sequences corresponds to the actual electricity load data of a certain historical day earlier than the target day, and each set of historical electricity load data sequences includes historical electricity load data corresponding to multiple sub-time periods. Preferably, each sub-time period is 15 minutes long. If each sub-time period is 15 minutes long, then a 24-hour day can be divided into 96 sub-time periods. Correspondingly, each set of historical electricity load data sequences includes 96 historical electricity load data corresponding to 96 sub-time periods, and the sub-time periods corresponding to each set of historical electricity load data sequences are consistent. By acquiring electricity load data every 15 minutes within each day of historical multiple days, the short-term fluctuation characteristics of electricity load can be captured, thereby providing data support for subsequent feature extraction and higher-precision electricity load prediction.

[0034] It is understood that the historical multiple days in this step specifically refers to multiple days earlier than the target date, such as the day, two days, or three days before the target date. Since historical data of different time spans can reflect the changing patterns of electricity load in the electricity consumption area from different dimensions, for example, recent data (such as 1 to 3 days before the target date) reflects the short-term continuity and trend of electricity consumption behavior in the electricity consumption area, while slightly older data (such as 7 days or 14 days before the target date) reflects the fluctuation characteristics of electricity load in the electricity consumption area on a weekly cycle and the long-term electricity consumption pattern. Therefore, in this embodiment of the application, preferably, the historical multiple days refer to the day, two days, three days, seven days, and fourteen days before the target date.

[0035] It should be noted that after obtaining multiple sets of historical electricity load data sequences, the data can be cleaned first, specifically including deduplication, abnormal spike processing, data missing anomaly processing, and negative value anomaly processing. Specifically, deduplication means that if duplicate historical electricity load data corresponding to the same sub-time period exist in a set of historical electricity load data sequences, the historical electricity load data with the latest timestamp is taken, and the remaining historical electricity load data for the same sub-time period are deleted. Abnormal spike processing means that the 99.5th percentile of all data is calculated first. For example, historical electricity load on weekdays is 800–1200 kW, historical electricity load at night is 300–500 kW, historical electricity load on weekends is 200–400 kW, and historical electricity load on public holidays is 50–150 kW. The overall 99.5th percentile is approximately 1273 kW. Then, for single points exceeding this value—for example, a historical electricity load data point corresponding to a specific sub-period in a series of historical electricity load data sets with a value of 2500 kW—this value, originally higher than 1273 kW, may be due to sensor fluctuations causing data anomalies. Therefore, 2500 kW is replaced with the average of the historical electricity load data points corresponding to the two sub-periods before and after it. If the values ​​before and after this sub-period are 960 kW and 975 kW respectively, then it is replaced with (960+975) / 2≈968 kW. If the historical electricity load data corresponding to three or more consecutive sub-periods exceeds the quantile, the entire segment is marked as a missing segment. Data missing anomaly handling refers to the following: If, due to equipment maintenance, historical electricity load data for a specific time period on a certain day cannot be collected, meaning that the historical electricity load data for that specific time period is missing from the historical electricity load sequence for that day, then the average of the historical electricity load data from the preceding and following time periods is used as the historical electricity load data for that specific time period. Negative value anomaly handling refers to the following: If the historical electricity load data for a given time period is negative, it is uniformly set to 0.

[0036] Step S104: Determine multiple sets of rate of change sequences and one set of mean sequences based on multiple sets of historical electricity load data sequences.

[0037] Each set of change rate sequences includes the change rate between historical electricity load data corresponding to the same sub-period of two sets of historical electricity load data sequences, and the mean sequence includes the mean of historical electricity load data corresponding to the same sub-period of multiple sets of historical electricity load data sequences.

[0038] In this embodiment of the application, preferably, the multiple sets of rate of change sequences include a first set of rate of change sequences and a second set of rate of change sequences. The specific implementation of step S104 will be described in detail below with reference to the accompanying drawings. Figure 2 The embodiments provided in this application are shown. Figure 1A flowchart illustrating the sub-steps of step S104. (See attached flowchart.) Figure 2 As shown, the first set of rate of change sequences, the second set of rate of change sequences, and the mean sequence can be determined through the following steps S1041 to S1044.

[0039] Step S1041: Determine the first historical electricity load data sequence corresponding to the day before the target date, the second historical electricity load data sequence corresponding to the two days before the target date, the third historical electricity load data sequence corresponding to the three days before the target date, the fourth historical electricity load data sequence corresponding to the four days before the target date, the fifth historical electricity load data sequence corresponding to the five days before the target date, the sixth historical electricity load data sequence corresponding to the six days before the target date, and the seventh historical electricity load data sequence corresponding to the seven days before the target date from multiple sets of historical electricity load data sequences.

[0040] Step S1042: Determine the rate of change of historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the seventh historical electricity load data sequence to obtain the first set of rate of change sequences.

[0041] For example, if the electricity load corresponding to a certain sub-period in the first historical electricity load data sequence is 910 kW, and the electricity load corresponding to the same sub-period in the seventh historical electricity load data sequence is 880 kW, then the rate of change corresponding to that sub-period in the first set of rate of change sequence is (910-880) / 880=3.4%.

[0042] In this step, the rate of change of electricity load data for each of the same sub-periods is calculated for two specific historical days (i.e., the day before the target day and the seven days before), so that the first set of rate of change sequence can reflect the fluctuation and growth trend of electricity load in a week-long cycle.

[0043] Step S1043: Determine the rate of change of historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the second historical electricity load data sequence to obtain the second set of rate of change sequences.

[0044] This step is similar to step S1042. Therefore, the principle and specific implementation method of this step can be referred to step S1042, and will not be repeated here.

[0045] Step S1044: Determine the average value of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence, the second historical electricity load data sequence, the third historical electricity load data sequence, the fourth historical electricity load data sequence, the fifth historical electricity load data sequence, the sixth historical electricity load data sequence, and the seventh historical electricity load data sequence, and obtain the mean value sequence.

[0046] In this step, the mean sequence is determined in the above manner. The mean sequence can represent the overall average level of electricity load in the same sub-period within a consecutive week before the target date. It can effectively smooth the random fluctuations of daily data, thereby reflecting the stable benchmark characteristics of electricity load in that period.

[0047] Step S105: Determine the third sub-feature based on the mean, peak and valley values ​​of multiple sets of historical electricity load data sequences.

[0048] In this step, the third sub-feature can be determined based on the mean, peak, and valley values ​​of multiple sets of historical electricity load data sequences corresponding to the days most recent to the target date. The specific implementation of this step will be described below with reference to the accompanying diagram. Figure 3 The embodiments provided in this application are shown. Figure 1 A flowchart illustrating the sub-steps of step S105. (See attached flowchart.) Figure 3 As shown, the third sub-feature can be determined through the following steps S1051 to S1054.

[0049] Step S1051: Determine the first historical electricity load data sequence corresponding to the day before the target date, the second historical electricity load data sequence corresponding to the two days before the target date, and the third historical electricity load data sequence corresponding to the three days before the target date from multiple sets of historical electricity load data sequences.

[0050] Step S1052: Determine the average value of all historical electricity load data in the first historical electricity load data sequence, the second historical electricity load data sequence, and the third historical electricity load data sequence to obtain the recent average electricity load.

[0051] Taking a historical electricity load data sequence consisting of 96 historical electricity load data corresponding to 96 sub-time periods with a duration of 15 minutes as an example, in this step, the average value of all 288 historical electricity load data in the first historical electricity load data sequence, the second historical electricity load data sequence, and the third historical electricity load data sequence is determined, thereby obtaining the recent average electricity load value.

[0052] In this step, by calculating the average value of all historical electricity load data over the three days prior to the target date, the overall level of electricity load in the area can be comprehensively reflected by this average value, thus providing a reference value that can represent the short-term benchmark electricity load for subsequent forecasts.

[0053] Step S1053: Determine the peak and valley values ​​from all historical electricity load data in the first historical electricity load data sequence, the second historical electricity load data sequence, and the third historical electricity load data sequence to obtain the recent electricity load peak value and the recent electricity load valley value.

[0054] In this step, by filtering and comparing all historical electricity load data in the first, second, and third historical electricity load data sequences, the maximum and minimum values ​​are identified. The resulting recent electricity load peak and valley values ​​are then used to quantify the fluctuation range and extreme levels of the recent electricity load.

[0055] Taking a historical electricity load data sequence consisting of 96 historical electricity load data corresponding to 96 sub-time periods with a duration of 15 minutes as an example, in this step, the peak and valley values ​​of all 288 historical electricity load data in the first, second, and third historical electricity load data sequences are determined to obtain the recent electricity load peak value and the recent electricity load valley value.

[0056] Step S1054: Use the recent average electricity load, the recent peak electricity load, and the recent valley electricity load as the third sub-feature.

[0057] The third sub-feature has three dimensions. If the recent average electricity load is a, the recent peak electricity load is b, and the recent valley electricity load is c, then the third sub-feature is [a, b, c].

[0058] Step S106: Determine the first feature based on the first sub-feature, the second sub-feature, multiple sets of historical electricity load data sequences, multiple sets of change rate sequences, mean sequence, and the third sub-feature.

[0059] In this step, the first sub-feature, the second sub-feature, multiple sets of historical electricity load data sequences, multiple sets of change rate sequences, mean sequence, and the third sub-feature can be directly spliced ​​or combined to obtain the first feature containing all the above information.

[0060] Step S107: Input the first feature into the trained first model and the trained second model corresponding to the date attribute of the target day, respectively, to obtain the first electricity load prediction data sequence and the second electricity load prediction data sequence.

[0061] The first and second electricity load forecast data sequences are output by the trained first and second models, respectively. The first electricity load forecast data sequence includes multiple first electricity load forecast data corresponding to multiple sub-time periods of the target day, and the second electricity load forecast data sequence includes multiple second electricity load forecast data corresponding to multiple sub-time periods of the target day.

[0062] The first and second models can have the same or different structures. The first and second models can be one or more combinations of support vector machines (SVM), deep neural networks (DNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), or gated recurrent units (GRU).

[0063] It is worth noting that in this embodiment of the application, the data used to train the first model and the second model are different. By using different training data to train the first model and the second model, the first model and the second model can learn different patterns, thereby enabling the model to capture the changing patterns of electricity load under different characteristics from multiple dimensions. This combines the predictive advantages of both models to improve the accuracy and robustness of the final prediction results.

[0064] When training the first model, electricity load data for multiple days prior to and adjacent to a historical date are determined. These load data are used to define a first feature, which is then input into the first model to obtain the predicted electricity load for that historical date. The model's parameters are then adjusted based on the actual electricity load data for that historical date and the predicted load data, and the model is iteratively trained to obtain a well-trained first model. For example, if a historical date is selected as a training sample, electricity load data for multiple historical days prior to and adjacent to that date are determined. A first feature is constructed based on this data and input into the first model to obtain the predicted electricity load for that historical date. The error is calculated by comparing this feature with the actual electricity load data for that historical date, and the model parameters are updated using backpropagation based on the error. This process is repeated iteratively trained on a large number of historical samples until the model converges, resulting in a well-trained first model capable of capturing the overall load variation pattern.

[0065] In this embodiment, three second models are included: a first second model, a second second model, and a third second model, which correspond to weekdays, weekends, and public holidays, respectively. If the target day to be predicted is a weekday, the first feature is input into the first second model corresponding to a weekday; if the target day to be predicted is a weekend, the first feature is input into the second second model corresponding to a weekend; and if the target day to be predicted is a public holiday, the first feature is input into the third second model corresponding to a public holiday. The differences between these three second models will be described below.

[0066] When training the first second model (i.e., the second model corresponding to a workday), the electricity load data of the area for several days earlier than a certain historical workday and adjacent to that historical workday are first determined. This data is used to determine the first feature. Then, the first feature is input into the first second model to obtain the electricity load prediction data for that historical workday output by the model. Finally, the model parameters are adjusted based on the actual electricity load data for that historical workday and the predicted data output by the model. Through iterative training, the trained first second model is obtained. For example, if a historical Wednesday is selected as a training sample, the electricity load data for several historical days prior to and adjacent to that historical Wednesday are determined. Based on this data, the first feature is constructed, and the first feature is input into the first second model to obtain the electricity load prediction data for that historical Wednesday. The error is calculated by comparing it with the actual electricity load data for that historical Wednesday, and then the model parameters are updated based on backpropagation of the error. Through this process, a large number of historical workday samples are repeatedly iterated and trained until the model converges, thereby obtaining a trained first second model that can adapt to the workday load patterns.

[0067] When training the second model (i.e., the second model corresponding to the weekend), similar to training the first model, firstly, determine the electricity load data of the area for multiple days earlier than a certain historical weekend (such as Saturday or Sunday) and adjacent to that historical weekend, and use this data to determine the first feature; then, input the first feature into the second model to obtain the electricity load prediction data for that historical weekend output by the model; finally, adjust the model parameters based on the actual electricity load data for that historical weekend and the prediction data output by the model, and obtain the trained second model through iterative training.

[0068] When training the third second model (i.e., the second model corresponding to the statutory holiday), similar to training the first second model, firstly, the electricity load data of the electricity consumption area for multiple days earlier than a certain historical statutory holiday and adjacent to that historical statutory holiday is determined, and the first feature is determined using this data; then, the first feature is input into the third second model to obtain the electricity load prediction data for that historical statutory holiday output by the model; finally, the model parameters are adjusted based on the actual electricity load data of that historical statutory holiday and the prediction data output by the model, and the trained third second model is obtained through iterative training.

[0069] In this application, three secondary models are used to learn different patterns, specifically modeling the different electricity consumption characteristics and load change patterns for weekdays, weekends, and public holidays. This allows for more accurate capture of differences in electricity consumption characteristics under various date attributes, thereby improving the load forecasting accuracy for different types of dates. The features used to train the three secondary models have the same dimensions, and are the same as the features used to train the primary model. The differences between them are determined by the number of training samples. For example, the secondary model that only learns weekday samples learns the electricity load pattern of areas with "full production on weekdays"; the secondary model that only learns weekend samples learns the electricity load pattern of areas with "single-shift maintenance on weekends"; and the secondary model that only learns public holiday samples learns the electricity load pattern of areas with "full plant shutdown on holidays". Therefore, by inputting features with the same dimensions into different models for training, the different models learn different mapping relationships of electricity consumption patterns simply because the date attributes of the training subsets are different.

[0070] For example, when training a model, if the electricity load data of the area for 304 days between January 1, 2025 and October 31, 2025 is obtained for training the model, then the electricity load data corresponding to 290 days of these days is used to train the first model, the electricity load data corresponding to 196 weekdays is used to train the second model corresponding to weekdays, the electricity load data corresponding to 67 weekend days is used to train the second model corresponding to weekends, and the electricity load data corresponding to 27 statutory holidays is used to train the second model corresponding to statutory holidays.

[0071] It should be noted that the model includes a LightGBM regressor. During model training, the LightGBM regressor iterates tree by tree across multiple training sets. The first tree fits the original residuals, the second tree fits the residuals of the first tree, and so on. After each tree is completed, the RMSE is evaluated using multiple validation sets. If the RMSE on the validation sets no longer decreases after 30 consecutive trees, the training stops early.

[0072] Step S108: Based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature, determine the second feature corresponding to each sub-period.

[0073] In this step, the historical electricity load data from multiple sets of historical electricity load data sequences, the rate of change from multiple sets of rate of change sequences, the mean from the mean value sequence, and the data from the third sub-feature can be directly spliced ​​or combined to obtain the second feature corresponding to each sub-period.

[0074] Step S109: Input the second feature corresponding to each sub-time period into the trained third model corresponding to each sub-time period to obtain the third electricity load prediction data sequence.

[0075] Taking a 15-minute duration for each sub-period as an example, this application includes 96 third models, each used to predict the electricity load data for the corresponding sub-period. Accordingly, the third electricity load prediction data sequence includes 96 third electricity load prediction data corresponding to the 96 sub-periods of the target day, with each third electricity load prediction data output by the trained third model corresponding to its sub-period.

[0076] Similar to the multiple second models, these 96 third models are also trained independently. Specifically, when training each third model, multiple historical electricity load data samples and corresponding features corresponding to the model's sub-periods are selected as the training set. The model parameters are adjusted by minimizing the error between the predicted and actual values, thus obtaining a model specifically designed to predict electricity load for a particular sub-period. This time-segmented modeling approach allows for fine-grained fitting of electricity load characteristics at different times of the day, further improving prediction accuracy.

[0077] Step S110: Perform weighted fusion processing on the first, second, and third electricity load forecast data corresponding to the same sub-period in the first, second, and third electricity load forecast data sequences to obtain the electricity load forecast data sequence.

[0078] The electricity load forecast data sequence includes electricity load forecast data corresponding to multiple sub-time periods of the target day. The electricity load forecast data U in the electricity load forecast data sequence corresponding to the i-th sub-time period... i =w1×U i1 +w2×U i2 +w3×U i3 Where w1, w2, and w3 are the first weight corresponding to the first electricity load forecast data, the second weight corresponding to the second electricity load forecast data, and the third weight corresponding to the third electricity load forecast data, respectively. i1 U i2 and U i3 These are the i-th first power load forecast data in the first power load forecast data sequence, the i-th second power load forecast data in the second power load forecast data sequence, and the i-th third power load forecast data in the third power load forecast data sequence, respectively. i is a positive integer.

[0079] w1, w2, and w3 can be predetermined as needed. The specific values ​​of w1, w2, and w3 can be determined through statistical analysis of the historical prediction performance of the first, second, and third models during the model validation phase. For example, the prediction accuracy or mean squared error of each model on the validation set can be calculated, and weights can be allocated based on the magnitude of the prediction error, assigning higher weights to models with smaller prediction errors. Alternatively, the weight coefficients corresponding to each sub-period can be dynamically adjusted based on the differences in prediction accuracy across different time periods, thereby achieving optimal combination and ensuring that the weighted and fused electricity load prediction data sequence approximates the actual electricity load curve to the greatest extent possible.

[0080] In this embodiment, through steps S101 to S110, the global trend capture capability of the first model, the specific date type adaptability of the second model, and the refined time period fitting capability of the third model can be fully utilized to achieve deep fusion and complementarity of multi-source feature information and multi-scale change patterns. Ultimately, the electricity load prediction data sequence obtained through weighted fusion can effectively overcome the limitations of a single model in the face of complex and ever-changing electricity consumption scenarios, thereby improving the prediction accuracy and stability of the target day's electricity load and providing reliable data support for power dispatching, demand-side management, and electricity cost control in the electricity consumption area.

[0081] In this application, the first model determines the general direction of electricity load. The cycle of electricity load variation is on an "annual" scale, and typically, weekday electricity load is greater than weekend electricity load, daytime electricity load is greater than nighttime electricity load, and summer electricity load is greater than spring and autumn electricity load. These patterns require a large amount of data to be stably learned; the first model is trained with the full sample size, providing the largest possible sample size and acting as a safety net. Relying solely on the second or third model could lead to extreme biases when the sample size is small (e.g., only a few dozen days of data). The first model ensures that even if the signals from the other two models are inaccurate, the overall trend will not deviate too far.

[0082] The second model determines the operating mode for each day. Within the same sub-period, production is 1000kW at full capacity on weekdays, 300kW at half capacity on weekends, and 100kW during statutory holidays when production is halted. If a single model were to learn the electricity load patterns for all three date attributes simultaneously, it would be forced to compromise to approximately 500kW, making it inaccurate for any one type. Therefore, each second model is trained using data from only one date attribute, ensuring the curve shape is not affected by data from other date attributes, and resulting in more accurate predictions for each date attribute. The core source of change in industrial and commercial electricity load is production scheduling switching. The first model alone cannot distinguish between the completely different scenarios of "today is Saturday, so production is reduced" and "today is Saturday, but a workday is adjusted, so full capacity is required."

[0083] The third model depicts patterns that occur almost daily within a fixed sub-period—for example, someone always starts work at 8:00 AM, someone always eats lunch at 12:00 PM, and someone always changes shifts at 6:00 PM. These patterns change daily, regardless of the day of the week or the season. If the first or second model simultaneously processes electricity load patterns across different sub-periods, these operational patterns can easily be overwhelmed by calendar and seasonal signals—the model might mistakenly assume that high load at 8:00 AM is due to high summer temperatures, rather than because the factory starts work at 8:00 AM every day. The third model deliberately removes seasonal / absolute time features such as months, days of the week, and temperatures, allowing the model to focus on the electricity consumption patterns of specific sub-period areas, thus making the localized predictions of electricity load for each sub-period more accurate.

[0084] It is understandable that the higher the dimensionality of the constructed features, the richer the amount of information the model can receive and utilize when inputting the features, thus enabling a more accurate prediction of the electricity load on the target day based on the input features. Based on this, in some embodiments, before executing step S106, one or more steps S11, S12, S13, and S14 are executed to obtain relevant data. The relevant data includes one or more sets of ratio sequences, the third feature, the fourth feature, and the fifth feature. The specific methods for determining the multiple sets of ratio sequences, the third feature, the fourth feature, and the fifth feature will be described below.

[0085] Step S11 involves determining multiple ratio sequences based on multiple sets of historical electricity load data sequences. Each ratio sequence includes the ratio of historical electricity load data corresponding to the same sub-period in two sets of historical electricity load data sequences.

[0086] The multiple ratio sequences include a first set of ratio sequences and a second set of ratio sequences. Specifically, the first set of ratio sequences is obtained by determining the ratio between the first historical electricity load data sequence corresponding to the day before the target date and the second historical electricity load data sequence corresponding to the two days before the target date. The first set of ratio sequences includes ratios corresponding to multiple sub-time periods, where each ratio represents the ratio relationship between the historical electricity load data of that sub-time period in the first historical electricity load data sequence and the historical electricity load data of the same sub-time period in the second historical electricity load data sequence.

[0087] The ratio of the seventh historical electricity load data sequence corresponding to seven days prior to the target date to the fourteenth historical electricity load data sequence corresponding to fourteen days prior to the target date is determined to obtain the second set of ratio sequences. The second set of ratio sequences is similar to the first set of ratio sequences, and will not be described in detail here.

[0088] Step S12 is: to determine the target historical electricity load data sequence from multiple sets of historical electricity load data sequences that has the same date attribute as the target date and is closest to the target date, and to determine the third feature based on the target historical electricity load data sequence.

[0089] For example, if the target date is November 6, 2025, which is a working day, the day with the same date attribute as the target date and closest to the target date is November 5, 2025. The historical electricity load data sequence corresponding to November 5, 2025 is the target historical electricity load data sequence.

[0090] Taking a 15-minute duration for each sub-period as an example, the third feature includes a total of 192 data points. Specifically, when determining the first 96 data points of the third feature, the ratio of the historical electricity load data corresponding to the j-th sub-period to the historical electricity load data corresponding to the (j+1)-th sub-period in the target historical electricity load data sequence is determined, thus obtaining the (j+1)-th data point from the first 96 data points of the third feature. j is a positive integer, and the first data point in the third feature can be preset as needed.

[0091] When determining the 97th to 192nd data points of the third feature, the ratio of the historical electricity load data corresponding to the j-th sub-period to the historical electricity load data corresponding to the (j+1)-th sub-period in the target historical electricity load data sequence is determined to obtain the (j+2)-th data point among the 97th to 192nd data points of the third feature. Furthermore, the 97th and 98th data points of the third feature can be preset as needed.

[0092] Step S13 is: Determine the fourth feature based on the date attributes of the n days before and m days after the target date, where n and m are both positive integers.

[0093] In this step, preferably, n includes 1 and 7, and m includes 1 and 7. Specifically, the fourth feature has four dimensions, which are: whether the day before the target day is a working day, whether the day after the target day is a public holiday, whether the seven days before the target day are working days, and whether the seven days after the target day are public holidays. When constructing the fourth feature, the date attributes corresponding to the day before, the day after, the seven days before, and the seven days after the target day are judged in sequence. If the corresponding attribute conditions are met, the value of the corresponding dimension is set to 1; otherwise, it is set to 0. By determining the fourth feature based on the date attributes of the days before and after the target day, the electricity load correlation characteristics of the target day and its adjacent time periods can be effectively captured, and the potential impact of special periods such as before and after holidays on electricity consumption behavior can be identified, thereby further improving the accuracy and robustness of the prediction results.

[0094] Step S14 is: Determine the fifth feature based on the meteorological forecast data of the target day and the month to which the target day belongs.

[0095] In this step, specifically, the fifth feature has three dimensions: temperature, relative humidity, and whether it belongs to a summer month. The meteorological forecast data for the target day includes predicted temperature and predicted humidity. These predicted temperature and humidity are used as the data for the first two dimensions of the fifth feature. If the target day belongs to a summer month, the value of the third dimension of the fifth feature is set to 1; otherwise, it is set to 0.

[0096] Considering that temperature and relative humidity affect electricity load, and that turning on air conditioning in summer leads to a significant increase in electricity load, the fifth feature is constructed by introducing meteorological data and monthly attributes. This enables the model to accurately capture the impact of environmental factors and seasonal electricity consumption patterns on electricity load, thereby effectively improving the accuracy of the prediction results.

[0097] Based on this, step S106 includes: determining the first feature based on the first sub-feature, the second sub-feature, multiple sets of historical electricity load data sequences, multiple sets of rate of change sequences, the mean sequence, the third sub-feature, and related data. Specifically, the first sub-feature, the second sub-feature, multiple sets of historical electricity load data sequences, multiple sets of rate of change sequences, the mean sequence, the third sub-feature, and related data are concatenated to integrate multi-dimensional feature information and obtain the first feature.

[0098] Step S108 includes: if the relevant data includes multiple sets of ratio sequences, then based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, the third sub-feature, and the multiple sets of ratio sequences, determine the second feature corresponding to each sub-period; if the relevant data does not include multiple sets of ratio sequences, then based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature, determine the second feature corresponding to each sub-period.

[0099] If the relevant data includes multiple ratio sequences, then the historical electricity load data from multiple historical electricity load data sequences corresponding to each sub-period, the rate of change from multiple rate of change sequences, the mean from the mean sequence, the third sub-feature, and the multiple ratio sequences are spliced ​​or combined to construct a second feature containing multi-dimensional information. This allows full utilization of the relative load change pattern provided by the ratio sequences and enhances the feature's representational ability. If the relevant data does not include multiple ratio sequences, the specific implementation of step S108 can be referred to the aforementioned method. Figure 1 The provided embodiments are not described in detail here.

[0100] In this embodiment, by determining relevant information and, based on that information, determining the first and second features, comprehensive feature extraction and fusion of electricity load data can be performed from both daily and sub-period levels. This effectively uncovers key information from historical load data, trends, date attributes, meteorological factors, and relative variation patterns. This multi-dimensional feature construction method enriches the data dimensions of the input model, enabling it to more accurately capture the complex changing characteristics of electricity load and thus improve the prediction accuracy of the target day's electricity load.

[0101] The following section will explain, with reference to the accompanying diagram, how the first weight w1, the second weight w2, and the third weight w3 are determined. Figure 4 A flowchart illustrating the method for determining the first weight, second weight, and third weight provided in an embodiment of this application is shown. Figure 4 As shown, the first weight, the second weight, and the third weight can be determined through the following steps S201 to S204.

[0102] Step 201: Obtain multiple sets of electricity load training data sequences corresponding to historical multiple days of the electricity consumption area.

[0103] For example, the multiple sets of electricity load training data sequences obtained in this step include 304 sets of electricity load training data sequences for electricity consumption areas corresponding to January 1, 2025 to October 31, 2025. Each set of electricity load training data sequences includes electricity load training data corresponding to multiple sub-time periods (i.e., the actual electricity load data of the electricity consumption area corresponding to that sub-time period). Step 202: Divide the multiple sets of electricity load training data sequences into multiple datasets, each dataset including multiple sets of electricity load training data sequences.

[0104] The 304 sets of electricity load training data sequences were divided into five datasets in chronological order. The first dataset includes sets 1 to 61 of the 304 sets of electricity load training data sequences; the second dataset includes sets 62 to 122 of the 304 sets of electricity load training data sequences; the third dataset includes sets 123 to 183 of the 304 sets of electricity load training data sequences; the fourth dataset includes sets 184 to 244 of the 304 sets of electricity load training data sequences; and the fifth dataset includes sets 245 to 304 of the 304 sets of electricity load training data sequences.

[0105] Step 203: Based on the multiple sets of datasets, determine k sets of first training features, k sets of second training features, and k sets of third training features, and use the k sets of first training features, k sets of second training features, and k sets of third training features to input k first models, k second models, and k third models respectively to obtain k sets of first prediction data, k sets of second prediction data, and k sets of third prediction data, where k is a positive integer and k > 1.

[0106] Specifically, this step can be achieved through the following steps S21 to S22.

[0107] Step S21: Based on the first r sets of datasets in the multiple datasets, determine the rth first training feature in the k sets of first training features, the rth second training feature in the k sets of second training features, and the rth third training feature in the k sets of third training features. Input the rth first training feature into the rth first model in the k sets of first models, input the rth second training feature into the rth second model in the k sets of second models, and input the rth third training feature into the rth third model in the k sets of third models to obtain the rth first prediction data in the k sets of first prediction data, the rth second prediction data in the k sets of second prediction data, and the rth third prediction data in the k sets of third prediction data, where r is a positive integer and the initial value of r is 1.

[0108] In this step, the method for determining the first training feature is similar to the method for determining the first feature in the foregoing embodiments. Therefore, the first training feature can be determined by referring to the relevant descriptions in the foregoing embodiments.

[0109] It's worth noting that in this step, the second model is categorized based on date attributes: second models for weekdays, second models for weekends, and second models for public holidays. If subsequent predictions only need to forecast the electricity load of weekday areas, then the k second models correspond to the k weekdays; if subsequent predictions only need to forecast the electricity load of weekend areas, then the k second models correspond to the k weekends; and if subsequent predictions only need to forecast the electricity load of public holiday areas, then the k second models correspond to the k public holidays. If subsequent predictions need to forecast the electricity load of weekday, weekend, and public holiday areas, then there are corresponding second models for the k weekdays, the k weekends, and the k public holidays.

[0110] The method for determining the second training feature is similar to that for determining the first training feature. The only difference is that the date attributes of the dataset used to determine the second training feature are different, namely, they correspond to weekdays, weekends, or public holidays, to adapt to the differences in electricity load patterns under different date types.

[0111] The method for determining the third training feature is similar to the method for determining the second feature in the foregoing embodiments. Therefore, the third training feature can be determined by referring to the relevant descriptions in the foregoing embodiments, and will not be repeated here.

[0112] Step S22: If r < k, then increment r and go to step S21.

[0113] If r is less than k, then after each increment of r by 1, proceed to step S21.

[0114] Step 204: Based on k sets of first prediction data, k sets of second prediction data, k sets of third prediction data, and multiple sets of electricity load training data sequences, determine the first weight, the second weight, and the third weight.

[0115] Specifically, this step can be achieved through the following steps a1 to a4.

[0116] Step a1: Construct matrix H.

[0117] In matrix H, the first column represents k sets of first-predicted data, the second column represents k sets of second-predicted data, and the third column represents k sets of third-predicted data. Each row of data in matrix H corresponds to the same sub-time period.

[0118] Step a2: Construct vector Y based on multiple sets of electricity load training data sequences.

[0119] The vector Y includes a column of data, and each row of data in the vector Y is the electricity load training data for the sub-period corresponding to each row of predicted data in the matrix H.

[0120] Step a3: Determine if the formula is satisfied The matrix w at time, where, .

[0121] In the formulas used in this step, the subscript 2 represents the norm, and the superscript 2 represents the square.

[0122] Step a4: , and They were respectively designated as the first weight, the second weight, and the third weight.

[0123] It should be noted that the first weight Second weight and third weight The sum is 1. If determined through step a3... , and If the sum is greater than 1, then for , and Perform normalization processing, and then normalize the result. , and They were respectively designated as the first weight, the second weight, and the third weight.

[0124] In this embodiment of the application, by using the above steps S201 to S204, the prediction results of multiple models and historical actual electricity load data can be comprehensively utilized to construct a matrix and target vector containing different prediction data. The optimal weight coefficients are determined based on the non-negative least squares (NNLS) method, which can better quantify the contribution of different prediction models in different sub-periods, ensure the reasonable allocation of the weights of each model, and thus effectively improve the accuracy of the determined weights.

[0125] Figure 5 The diagram shows a structural schematic of the power load forecasting device provided in the embodiment of this application. The specific embodiments of this application do not limit the specific implementation of the power load forecasting device.

[0126] like Figure 5 As shown, the power load prediction device 300 may include a processor 302 and a memory 304.

[0127] The memory 304 is used to store the computer program 306. The memory 304 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device. The computer program 306 may include computer-executable instructions.

[0128] The processor 302 is used to execute the computer program 306 to implement the above-described embodiment of the power load forecasting method.

[0129] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The power load prediction device 300 may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0130] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described electricity load prediction method embodiment.

[0131] This application provides a computer program that can be executed by a processor to implement the above-described electricity load forecasting method embodiment.

[0132] This application provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described embodiment of the power load prediction method.

[0133] In the several embodiments provided in this application, any function, if implemented as a software functional module / unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the technical solution of this application 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 other electronic device) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing computer program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of this application are not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0135] It should be noted that the above embodiments are illustrative of this application and not restrictive, and those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In claims enumerating several means, several units or modules of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

[0136] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting electricity load on a target day in an electricity consumption area, characterized in that, The method includes: The date attribute and weekday attribute of the target day are determined, and a first sub-feature is determined based on the date attribute and weekday attribute of the target day, wherein the date attribute includes weekdays, weekends and public holidays, and the weekday attribute includes Monday, Tuesday, Wednesday, Thursday, Friday, Saturday and Sunday; Determine the start date and end date of a statutory holiday that is earlier than the target date and closest to the target date, and determine a second sub-feature based on a first interval of days between the start date and the target date and a second interval of days between the end date and the target date; Obtain multiple sets of historical electricity load data sequences corresponding to multiple historical days in the electricity consumption area, wherein each set of historical electricity load data sequences includes historical electricity load data corresponding to multiple sub-time periods; Based on multiple sets of historical electricity load data sequences, multiple sets of change rate sequences and one set of mean sequence are determined. Each set of change rate sequences includes the change rate between the historical electricity load data corresponding to the same sub-period of the two sets of historical electricity load data sequences. The mean sequence includes the mean of the historical electricity load data corresponding to the same sub-period of the multiple sets of historical electricity load data sequences. The third sub-feature is determined based on the mean, peak, and valley values ​​of multiple sets of historical electricity load data sequences; The first feature is determined based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of change rate sequences, the mean sequence, and the third sub-feature; The first feature is input into a trained first model and a trained second model corresponding to the date attribute of the target day, respectively, to obtain a first electricity load prediction data sequence output by the trained first model and a second electricity load prediction data sequence output by the trained second model. The first electricity load prediction data sequence includes multiple first electricity load prediction data corresponding to multiple sub-time periods of the target day, and the second electricity load prediction data sequence includes multiple second electricity load prediction data corresponding to multiple sub-time periods of the target day. The training data used to train the trained second model is determined based on the electricity load data of the electricity consumption area for multiple days earlier than the day with the same date attribute and adjacent to the day with the same date attribute. The day with the same date attribute is a day with the same date attribute as the target day. Based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature, the second feature corresponding to each sub-period is determined; The second feature corresponding to each sub-time period is input into the trained third model corresponding to each sub-time period to obtain a third electricity load prediction data sequence. The third electricity load prediction data sequence includes multiple third electricity load prediction data corresponding to multiple sub-time periods of the target day. Each third electricity load prediction data is output by the trained third model corresponding to its corresponding sub-time period. The first, second, and third electricity load forecast data corresponding to the same sub-period in the first, second, and third electricity load forecast data sequences are weighted and fused to obtain an electricity load forecast data sequence, wherein the electricity load forecast data sequence includes electricity load forecast data corresponding to multiple sub-periods of the target day.

2. The method according to claim 1, characterized in that, Before determining the first feature based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, and the third sub-feature, the method further includes: Perform one or more of steps S11, S12, S13, and S14 to obtain relevant data, wherein step S11 is: determining multiple ratio sequences based on multiple sets of historical electricity load data sequences, each set of ratio sequences including the ratio of historical electricity load data corresponding to the same sub-period of two sets of historical electricity load data sequences; step S12 is: determining a target historical electricity load data sequence from the multiple sets of historical electricity load data sequences that has the same date attribute as the target day and is closest to the target day, and determining a third feature based on the target historical electricity load data sequence; step S13 is: determining a fourth feature based on the date attributes of the n days before and m days after the target day, where n and m are both positive integers; step S14 is: determining a fifth feature based on the meteorological forecast data of the target day and the month to which the target day belongs, wherein the relevant data includes one or more of the multiple ratio sequences, the third feature, the fourth feature, and the fifth feature; The determination of the first feature based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of rate of change sequences, the mean sequence, and the third sub-feature includes: The first feature is determined based on the first sub-feature, the second sub-feature, the multiple sets of historical electricity load data sequences, the multiple sets of change rate sequences, the mean sequence, the third sub-feature, and the relevant data; The step of determining the second feature corresponding to each sub-period based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature includes: If the relevant data includes the multiple sets of ratio sequences, then based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, the third sub-feature, and the multiple sets of ratio sequences, the second feature corresponding to each sub-period is determined; If the relevant data does not include the multiple sets of ratio sequences, then the second feature corresponding to each sub-period is determined based on the historical electricity load data in the multiple sets of historical electricity load data sequences corresponding to each sub-period, the rate of change in the multiple sets of rate of change sequences, the mean in the mean sequence, and the third sub-feature.

3. The method according to claim 2, characterized in that, The multiple sets of ratio sequences include a first set of ratio sequences and a second set of ratio sequences. Step S11 includes: From the multiple sets of historical electricity load data sequences, determine the first historical electricity load data sequence corresponding to the day before the target day, the second historical electricity load data sequence corresponding to the two days before the target day, the seventh historical electricity load data sequence corresponding to the seven days before the target day, and the fourteenth historical electricity load data sequence corresponding to the 14 days before the target day. The ratio of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the second historical electricity load data sequence is determined to obtain the first set of ratio sequences; The ratio of the historical electricity load data corresponding to the same sub-period of the seventh historical electricity load data sequence to the historical electricity load data of the fourteenth historical electricity load data sequence is determined to obtain the second set of ratio sequences.

4. The method according to claim 1, characterized in that, The multiple sets of rate of change sequences include a first set of rate of change sequences and a second set of rate of change sequences; The determination of multiple sets of rate of change sequences and one set of mean sequences based on multiple sets of historical electricity load data sequences includes: From the multiple sets of historical electricity load data sequences, determine the first historical electricity load data sequence corresponding to the day before the target day, the second historical electricity load data sequence corresponding to the two days before the target day, the third historical electricity load data sequence corresponding to the three days before the target day, the fourth historical electricity load data sequence corresponding to the four days before the target day, the fifth historical electricity load data sequence corresponding to the five days before the target day, the sixth historical electricity load data sequence corresponding to the six days before the target day, and the seventh historical electricity load data sequence corresponding to the seven days before the target day. Determine the rate of change of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the seventh historical electricity load data sequence to obtain the first set of rate of change sequences; Determine the rate of change of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence and the second historical electricity load data sequence to obtain the second set of rate of change sequences; The average value sequence is obtained by determining the average value of the historical electricity load data corresponding to the same sub-period of the first historical electricity load data sequence, the second historical electricity load data sequence, the third historical electricity load data sequence, the fourth historical electricity load data sequence, the fifth historical electricity load data sequence, the sixth historical electricity load data sequence, and the seventh historical electricity load data sequence.

5. The method according to claim 1, characterized in that, The determination of the third sub-feature based on the mean, peak, and valley values ​​of multiple sets of historical electricity load data sequences includes: From the multiple sets of historical electricity load data sequences, determine the first historical electricity load data sequence corresponding to the day before the target day, the second historical electricity load data sequence corresponding to the two days before the target day, and the third historical electricity load data sequence corresponding to the three days before the target day; The average value of all historical electricity load data in the first historical electricity load data sequence, the second historical electricity load data sequence, and the third historical electricity load data sequence is determined to obtain the recent average electricity load. Peak and valley values ​​are determined from all historical electricity load data in the first historical electricity load data sequence, the second historical electricity load data sequence, and the third historical electricity load data sequence to obtain recent electricity load peak values ​​and recent electricity load valley values; The recent average electricity load, the recent peak electricity load, and the recent valley electricity load are used as the third sub-feature.

6. The method according to claim 1, characterized in that, The first weight corresponding to the first electricity load forecast data, the second weight corresponding to the second electricity load forecast data, and the third weight corresponding to the third electricity load forecast data are determined in advance through the following steps: Obtain multiple sets of electricity load training data sequences corresponding to historical multiple days of the electricity consumption area, wherein each set of electricity load training data sequences includes electricity load training data corresponding to multiple sub-time periods; The multiple sets of electricity load training data sequences are divided into multiple datasets, and each dataset includes multiple sets of electricity load training data sequences. Based on the multiple datasets, k sets of first training features and k sets of second training features are determined. The k sets of first training features are then input into k first models and k second models, respectively. The k sets of second training features are then input into k third models, respectively, to obtain k sets of first prediction data, k sets of second prediction data, and k sets of third prediction data, where k is a positive integer and k > 1. Based on the k sets of first prediction data, the k sets of second prediction data, the k sets of third prediction data, and the multiple sets of electricity load training data sequences, the first weight, the second weight, and the third weight are determined.

7. The method according to claim 6, characterized in that, The process involves determining k sets of first training features and k sets of second training features based on multiple datasets, inputting the k sets of first training features into k first models and k second models respectively, and inputting the k sets of second training features into k third models respectively, to obtain k sets of first prediction data, k sets of second prediction data, and k sets of third prediction data, including: Step S21: Based on the first r sets of datasets in the multiple datasets, determine the rth first training feature in the k sets of first training features and the rth second training feature in the k sets of second training features. Input the rth first training feature into the rth first model in the k sets of first models and the rth second model in the k sets of second models, and input the rth second training feature into the rth third model in the k sets of third models to obtain the rth first prediction data in the k sets of first prediction data, the rth second prediction data in the k sets of second prediction data, and the rth third prediction data in the k sets of third prediction data, where r is a positive integer and the initial value of r is 1. Step S22: If r < k, then increment r and go to step S21.

8. The method according to claim 6, characterized in that, The determination of the first weight, the second weight, and the third weight based on the k sets of first prediction data, the k sets of second prediction data, the k sets of third prediction data, and the multiple sets of electricity load training data sequences includes: Construct a matrix H, wherein the first column of the matrix H contains the first predicted data of the k groups, the second column of the matrix H contains the second predicted data of the k groups, and the third column of the matrix H contains the third predicted data of the k groups. A vector Y is constructed based on the multiple sets of electricity load training data sequences, wherein the vector Y includes a column of data, and each row of data in the vector Y is the electricity load training data for the sub-period corresponding to each row of predicted data in the matrix H; Determine if the formula is satisfied The matrix w at time, where, ; Will , and The weights are respectively determined as the first weight, the second weight, and the third weight.

9. An electricity load forecasting device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the power load forecasting method according to any one of claims 1 to 8.

10. A 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 power load prediction method according to any one of claims 1 to 8.

Citation Information

Patent Citations

  • Power system load prediction method based on multivariate data fusion

    CN119377809A

  • Load prediction method

    CN122136790A