Power load forecasting method and system

Abstract

The present invention provides a method and system for power load forecasting, which relates to the fields of meteorology and power research, and includes the following steps: S1, obtaining load data and meteorological data within the past two years, and marking the data with date types; S2, constructing training data sets for different dates in a short period of time to carry out training, and obtaining prediction models for corresponding dates; S3, obtaining input data for short-term load forecasting; S4, inputting the data obtained in S3 into the corresponding prediction models to obtain short-term load forecasting results. The present invention uses power load data from the past two years and meteorological data from the corresponding time period to redefine the daily time range, distinguish between working days and non-working days, and adopts the LightGBM machine learning method to learn and train the changes in short-term power load, and obtain a 15-minute power load forecast for the next three days. The output power load forecast is more accurate and precise, providing strong support for the business decision-making of the power sector.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and in particular to a power load forecasting method and system. Background Art

[0002] With the development of the economy and the acceleration of urbanization, the demand for electricity continues to grow. The supply and demand of electric energy is closely related to weather changes. High temperatures, low temperatures, dryness and wetness often affect the grid dispatching and control strategy. Wind and light conditions determine the output capacity of new energy sources. Accurate grid load forecasting is a key factor in ensuring energy supply. In recent years, with the increase in photovoltaic access capacity, the impact of meteorological and other factors on load has become increasingly significant. The efficient operation of new power systems will increasingly rely on the assessment and prediction capabilities and integration level of meteorological factors. The existing algorithm for predicting power load is an artificial intelligence-based power load forecasting method. This method can reduce the impact of noise data on the forecast results and improve the accuracy of power load forecasting.

[0003] In the related art, an artificial intelligence-based power load forecasting method is disclosed in an application document with publication number CN118520428A. The method includes: fitting the load data in the acquired historical power load data segment, determining the extreme points of the fitting curve, and then calculating the initial noise level of the historical power load data segment, and the initial noise level of the corresponding historical power grid frequency data segment, as well as the correlation between the historical power load data segment and the corresponding historical power grid frequency data segment to correct the initial noise level, and obtain the final noise level of the historical power load data segment, so as to use the final noise level of each historical power load data segment to weight the predicted value determined by the corresponding historical power load data segment to obtain the final predicted value.

[0004] However, power load is closely related to meteorological factors. The above scheme only uses historical power load data for fitting training. The prediction system only considers the inherent laws of the load, and does not consider the impact of changes in meteorological factors such as temperature, precipitation, and humidity on power load changes. This, to a certain extent, limits the improvement of the accuracy of load forecasting. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting power load.

[0006] According to a first aspect of an embodiment of the present invention, there is provided a method for predicting power load, comprising the following steps:

[0007] S1. Obtain load data and meteorological data from the past two years and mark the data with date type;

[0008] S2. Construct training data sets for different dates in a short period of time and conduct training to obtain prediction models for the corresponding dates;

[0009] S3. Obtain input data for short-term load forecasting;

[0010] S4. Input the data obtained in S3 into the corresponding prediction models to obtain short-term load forecast results.

[0011] According to a first aspect of an embodiment of the present invention, there is provided a power load forecasting system, comprising:

[0012] The historical data acquisition module is used to obtain load data and meteorological data in the past two years and mark the data with date type;

[0013] The model acquisition module is used to construct training data sets for different dates in a short period of time and conduct training to obtain the prediction model for the corresponding date;

[0014] Forecast input data acquisition module, used to obtain input data for short-term load forecasting;

[0015] The load forecast result acquisition module is used to input the forecast input data into the corresponding forecast model to obtain the short-term load forecast result.

[0016] The beneficial effects of the present invention are:

[0017] The power load forecasting method and system of the present invention utilizes the relationship between load and weather factors, applies machine learning algorithms, and explores the impact of weather on load. Based on numerical weather forecasts, they accurately predict short-term power load. Based on these accurate predictions, power companies can formulate load response plans for different scenarios based on load forecasts, increase the size of the adjustable load resource pool, participate in the regulation of distribution network operations, effectively alleviate the pressure of peak shaving and valley filling on the power grid, maintain a balanced power supply, and ensure normal electricity consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 is a flow chart of the power load forecasting method of the present invention;

[0020] Figure 2 is a short-term load forecast result in an embodiment of the present invention;

[0021] Figure 3 2 is a schematic diagram comparing the prediction result and the actual load in an embodiment of the present invention;

[0022] Figure 4 3 is a schematic diagram showing a comparison between the load forecast and the actual load from August 16 to August 31, 2024, in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific implementation cases. It should be noted that the embodiments described here are only for illustration and are not intended to limit the present invention.

[0024] Reference Figure 1 In one embodiment of the present invention, a method and system for predicting power load includes the following steps:

[0025] S1: Obtain load data and meteorological data from the past two years and mark the data with date type.

[0026] In this embodiment, the past two years refers to two years counting backward from the current day.

[0027] Specifically, the load data is the real-time historical data of the power dispatching load, and the time resolution of the data is 5 minutes; the meteorological data is the real-time historical data of the meteorological conditions in the same period as the load data, including: temperature TEM, precipitation RAIN, and humidity RH.

[0028] Furthermore, considering the significant difference in power load between weekdays and holidays, all data need to be marked with date type.

[0029] Specifically, the method of marking the date type of the data is: mark working days (such as ordinary Monday to Friday) as 0; mark holidays that are not ordinary Saturdays as 1 (such as Sundays and statutory holidays such as the Spring Festival, Qingming Festival, Dragon Boat Festival, May Day, Mid-Autumn Festival, National Day, and New Year's Day); ordinary Saturdays are marked as 6, and if Saturday happens to be the above-mentioned Spring Festival, Qingming Festival, etc., it is marked as 1.

[0030] S2: Build training data sets for different dates in a short period of time and conduct training to obtain the prediction model for the corresponding date.

[0031] Because the power company's marketing decision-making department focuses on short-term load changes, in this example, the short-term period is the next three days. Therefore, a 15-minute load change forecast is required for the next three days. In this example, the time range for each day is defined as "08:00 on the first day to 08:00 on the second day."

[0032] Taking into account the different patterns of load changes in different time periods, in this embodiment, corresponding power load forecasting models are established for dates in the short term, that is, the models for predicting the load of "today", "tomorrow" and "the day after tomorrow" are Model 1, Model 2 and Model 3 respectively, and the training data sets input into different models are different.

[0033] Taking Model 1 as an example, the method of constructing the training dataset includes:

[0034] For the load situation of the day, the actual business operation is "today" and the date is set to t. Collect the actual data of the same period last year and the recent period respectively:

[0035] 1. Go back to the same period last year, i.e., t-1y (y is year); and go back one month, i.e., t-1y-1m (m is month); starting from t-1y-1m, obtain load data for two months in the future, with the time range being: t-1y-1m to t-1y+1m. This is equivalent to obtaining load data data1 for the past two months before and after the same period last year;

[0036] 2. Using the previous day (t-1) as the benchmark, look back two months, i.e., t-1-2m; and using this as the starting time, collect load data for two consecutive months, with the time range being: t-1-2m to t-1. This is equivalent to obtaining the load data for the most recent two months, data2;

[0037] 3. Merge data1 and data2 to form a load data set with similar climate background conditions to "today". This data set includes the load conditions of the same period last year and the recent load conditions.

[0038] Use the collected load data set as the target data and search for other data sets that are directly related to the target data:

[0039] 1. Considering the obvious daily variation characteristics of load changes, obtain the number of minutes corresponding to the time span of the target data. For example, the number of minutes corresponding to the target load value at 14:00 on a certain day is 840, that is, 14×60.

[0040] 2. Obtain meteorological data at the same time period and time point as the target data, including TEM, RAIN, and RH.

[0041] 3. Considering that weekday loads better represent actual electricity consumption for production and daily life, and that actual loads fluctuate continuously, the load on the day (d) is closely related not only to the weather conditions of the day but also to the load and weather changes of the previous few days. Therefore, we obtain the load and weather conditions for the five most recent weekdays before the current day.

[0042] For example, if the target load value P at 2:00 PM on a particular day is a consecutive five-day historical workday, the associated data includes not only the TEM, RAIN, and RH at that time on day d, but also the P, TEM, RAIN, and RH at 2:00 PM on days d-5, d-4, d-3, d-2, and d-1. If one or more days between d-5 and d-1 are non-workdays, for example, d-5 and d-4 are Saturday and Sunday, respectively, the five historical workdays are d-7, d-6, d-3, d-2, and d-1.

[0043] 4. Read the date type data corresponding to the current date of the target data, that is, 0, 1, or 6, as the last associated data.

[0044] 5. The target load data and all associated data are collated to form the training dataset for Model 1. In this embodiment, the dataset is optionally arranged in the following order: minutes, date type data, meteorological data (TEM, RAIN, RH) for the same period and time as the load data, and historical data for each of the five working days mentioned above, where the historical data for each working day is P, TEM, RAIN, and RH, respectively. The temporal resolution of this dataset is 5 minutes.

[0045] In this embodiment, the basic idea of ​​constructing the training data set for Model 2 and Model 3 is the same as that for Model 1, except that in the "obtaining the load and weather conditions for the five most recent historical working days before the current day", Models 2 and 3 respectively obtain the P, TEM, RAIN, and RH data for the working days before d-1 and d-2. The reason for this is that in Model 2, if the load for "tomorrow" (d) is to be predicted, the load for "today" (d-1) is still unknown, so the load can only be based on the historical data for the five historical working days before "yesterday" (d-1). Similarly, for Model 3, if the load for "the day after tomorrow" (d) is to be predicted, the load for "today" (d-2) and tomorrow (d-1) is unknown, so the data for the five historical working days before "yesterday" (d-2) can only be used to correspond to the target load data.

[0046] In this embodiment, the method for training the model includes:

[0047] Use Python to call the LightGBM package as the machine learning model framework. The model parameters can be set as follows: boosting type is "bdbt", objective function is "regression", evaluation function is "rmse", number of leaf nodes is set to 100, learning rate is set to 0.1, feature selection ratio for tree building is set to 0.9, sample sampling ratio for tree building is set to 0.8, number of trees in the model is set to 2000, maximum tree depth is set to 10, frequency of ensemble learning during training is set to 5, and all other settings use the default values. The training datasets for each model are input into the framework for training, resulting in load forecasting models based on the corresponding relationship between power load and meteorological factors and the changes in the previous five consecutive working days. The trained models 1, 2, and 3 are obtained.

[0048] S3: Obtain input data for short-term load forecasting.

[0049] After establishing the power load forecast model, read the latest numerical weather forecast product and input the short-term forecast data into the model to obtain the power load forecast results within three days. The specific steps are as follows:

[0050] 1. Read the latest numerical weather forecast

[0051] Site forecast data is retrieved daily from the meteorological big data cloud platform. For example, site forecast data with the code NAFP_FOR_FTM_LOW_RIOF_JS is obtained from the Jiangsu Provincial Meteorological Bureau data platform. The forecast period is 0 to 72 hours, with a time resolution of 3 hours. This data product is updated twice daily at 08:00 and 20:00, respectively, and uses the following forecast elements: TEM, RAIN, and RH. Because 15-minute load forecasts are required, linear interpolation is performed on the forecast to generate 15-minute TEM, RAIN, and RH forecast data for the next 72 hours.

[0052] 2. Construct input data sets for different models

[0053] For Model 1, the steps to construct the data are as follows:

[0054] 1. Calculate the number of minutes from 08:00 today to 08:00 tomorrow, for every 15 minutes. For example, the number of minutes at 14:00 is 840.

[0055] 2. Extract the forecast data from step 1 that falls within the time range of "today", specifically from 08:00 "today" to 08:00 "tomorrow";

[0056] 3. Obtain the 15-minute daily real-time data of P, TEM, RAIN, and RH for the five most recent historical working days before "today" (t). The start and end time of each day's data are 08:00 on the current day to 08:00 the next day;

[0057] 4. According to the date attribute of "today", obtain the day type data of "today", that is, 0, 1, or 6;

[0058] 5. Summarize all the data in steps 1 to 4 and arrange them in a certain time and order. The arrangement style is consistent with the arrangement order of the training data set in S2.

[0059] For Model 2 and Model 3, the input data sets are similar to the above steps. The main differences are: Step 2 extracts the forecast data for "tomorrow" and "the day after tomorrow" from the forecast data in Step 1; Step 3 obtains the P, TEM, RAIN, and RH data for each of the five historical working days before t-1 (in Model 2, "tomorrow" is t) and before t-2 (in Model 3, "the day after tomorrow" is t); Step 4 obtains the daily type data for "tomorrow" and "the day after tomorrow".

[0060] S4: Input the data obtained in S3 into the corresponding prediction models to obtain short-term load forecast results.

[0061] Input the data set obtained from S3 into the corresponding model to automatically obtain the 15-minute load forecast results for "today", "tomorrow" and "the day after tomorrow". By splicing these three parts of forecast data in sequence, the power load forecast for the next three days is obtained. The forecast results can be automatically generated into Excel files or txt files, or can be automatically used as Figure 2 The picture shown is displayed in real time, making the prediction results more intuitive and easier to transmit.

[0062] While implementing the power load forecast for the next three days based on the LightGBM algorithm, the present invention also provides an automatic verification function for the forecast results, and the verification results can be generated synchronously with the forecast results. Assuming that the power department is concerned about the forecast for the next day every day, the focus is on verifying the forecast results of "tomorrow" in the output of the forecast model. Assuming that today is t, the specific steps are:

[0063] 1. Retrieve the actual load data of yesterday (t-1) for every 5 minutes;

[0064] 2. Retrieve the 15-minute load forecast data for the next day (t-1) reported from the day before yesterday (t-2), which should coincide with the actual load data date in step 1;

[0065] 3. Get the actual load data corresponding to the 15-minute load forecast data time, and use Figure 3The graphical form shown allows for an intuitive comparison of the actual and predicted curves. The root mean square error (RMSE) and mean absolute percentage error (MAPE) between the two can also be calculated using the following formulas:

[0066]

[0067] See also Figure 4 ,As can be seen from the figure, the load forecast during the high temperature period from August 16th to 31st 2024 is more accurate, with an average accuracy of 97.4%, and the average prediction accuracy of the daily maximum ,load value is 97.6%.

[0068] Based on the above embodiment, an embodiment of the present invention further provides a power load forecasting system, including:

[0069] The historical data acquisition module is used to obtain load data and meteorological data in the past two years and mark the data with date type;

[0070] The model acquisition module is used to construct training data sets for different dates in a short period of time and conduct training to obtain the prediction model for the corresponding date;

[0071] Forecast input data acquisition module, used to obtain input data for short-term load forecasting;

[0072] The load forecast result acquisition module is used to input the forecast input data into the corresponding forecast model to obtain the short-term load forecast result.

[0073] In this embodiment, the working process of each module is as described above and will not be repeated here.

[0074] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A method for predicting power load, characterized in that: The following steps are involved: S1. Obtain load data and meteorological data from the past two years and mark the data with date type; wherein the load data is real-time historical data of power dispatching load, and the time resolution of the data is 5 minutes; the meteorological data is real-time historical data of meteorological data from the same period as the load data, including: temperature TEM, precipitation RAIN, and humidity RH; and the method of marking the data with date type includes: marking weekdays as 0, marking holidays that are not ordinary Saturdays as 1, and marking ordinary Saturdays as 6; S2. Construct training data sets for different dates in a short period of time and conduct training to obtain the prediction model for the corresponding date; The method of constructing a training data set includes: obtaining target data and associated data; The method for obtaining target data includes: Get the load data data1 for the past two months before and after the same period last year; Get the load data data2 for the last two months; Merge data1 and data2 as the target data; and The method for obtaining associated data includes: Get the number of minutes corresponding to the time span of the target data; Acquire meteorological data at the same time period and time point as the target data, including TEM, RAIN, and RH; Get the load and weather conditions for the last five historical working days before the current day; Read the date type data corresponding to the current date of the target data, that is, 0, 1, or 6, as the last associated data; Arrange the target load data and all related data to form a training data set; S3. Obtain input data for short-term load forecasting; S4, input the data obtained in S3 into the corresponding prediction model to obtain the short-term load forecast result; In step S3, the method for obtaining input data for carrying out short-term load forecasting includes: reading the latest numerical weather forecast and constructing input data sets for different models respectively; The method for constructing input data sets for different models includes: Calculate the number of minutes from 08:00 on the current day to 08:00 on the next day for every 15 minutes; Extract the part of the forecast data that falls within the "today" time range; Obtain the 15-minute real-time data of P, TEM, RAIN, and RH for the five most recent historical working days before "the current day". The start and end time of each day's data are both 08:00 on the current day and 08:00 the next day. P is the target load value. According to the date attribute of "this day", obtain the day type data of "this day", that is, 0, 1, or 6; The data of each forecasted day in the short term are summarized separately and arranged in a certain time and sequence to obtain the input data sets of different models.

2. The power load forecasting method according to claim 1, characterized in that: In step S2, the method for training the model includes: Use Python to call the LightGBM package as the model framework for machine learning and set model parameters; The training data sets of each model are input into the framework for learning and training, so as to obtain the prediction model of the corresponding date after training.

3. The power load forecasting method according to claim 1, characterized in that: In step S4, the data collected in S3 is organized into the same format and order as the associated data in the training model, and input into the corresponding models according to different time periods to obtain the short-term 15-minute load forecast results; the forecast data in different time periods are arranged in order, and finally the short-term power load forecast results for the next three days can be obtained.

4. The power load forecasting method according to claim 3, characterized in that: Step S4 also includes: checking the forecast situation of the previous day; The method for testing the forecast situation of the previous day includes: Retrieve yesterday's 5-minute load actual data; Retrieve the 15-minute load forecast data for the next day reported from the day before yesterday; Take the actual load data corresponding to the 15-minute load forecast data time and verify it through one or more methods of curve comparison, calculation of error root mean square error or calculation of mean absolute percentage error.

5. A power load forecasting system using the power load forecasting method according to any one of claims 1 to 4, characterized in that: include: The historical data acquisition module is used to obtain load data and meteorological data in the past two years and mark the data with date type; The model acquisition module is used to construct training data sets for different dates in a short period of time and conduct training to obtain the prediction model for the corresponding date; The forecast input data acquisition module is used to obtain input data for short-term load forecasting; the load forecast result acquisition module is used to input the forecast input data into the corresponding forecast model to obtain the short-term load forecast result.

Images