A training method and application method of electricity price prediction model

Abstract

The present invention discloses a training method and application method of an electricity price prediction model. First, historical data is used to calculate a typical curve of real-time price and system load rate and a segmentation point corresponding to the system load rate by using a 3sigma (3σ) criterion algorithm and a box plot algorithm. Then, historical data with strong correlation is selected in combination with the features of each segment as the input of a machine learning algorithm, thereby training a corresponding prediction model. Finally, the model is called and combined with a piecewise correction algorithm for the system load rate to accurately predict the real-time price of a certain window period in the future.

Description

Technical Field

[0001] The invention belongs to the field of electricity price prediction in the power market, and specifically relates to a training method and an application method of an electricity price prediction model. Background Art

[0002] In the process of developing green power energy transformation, wind power generation plays a pivotal role. However, due to the randomness and volatility of wind power output, this will inevitably bring certain uncertainties to the power balance of the power system, and also bring great challenges to the stable operation of the power system. In addition, real-time adjustment of electricity prices is one of the most effective ways for the power market to adjust users' electricity consumption habits and balance the output of thermal power and wind power. Realizing real-time price prediction can largely coordinate the contradiction between wind power grid connection and energy storage.

[0003] At present, the electricity market mainly focuses on the prediction of day-ahead prices, and the research on real-time prices, especially the rolling prediction of real-time prices, still has a certain lag. The more widely used prediction methods include machine learning algorithms, time series neural network algorithms, and time-based prediction algorithms. However, the unit electricity price will be affected by many factors such as the load of the entire network, the output of new energy, and the proportion of external power transmission. Therefore, even if the price of peak and valley sections is predicted by time period, it is not enough to deal with all situations. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of poor generalization ability and low parameter adjustment efficiency of time-segment prediction of electricity prices in the prior art, and to provide a training method and application method of an electricity price prediction model. First, historical data are used to calculate the typical curve of real-time price and system load rate and the segmentation point corresponding to the system load rate using the 3sigma (3σ) criterion algorithm and the box-and-whisker algorithm, and then the historical data with strong correlation is selected as the input of the machine learning algorithm in combination with the features of each segment, so as to train the corresponding prediction model. Finally, the model is called and combined with the segmented correction algorithm of the system load rate to accurately predict the real-time price of a certain window period in the future.

[0005] In order to solve the technical problem, the technical solution of the present invention is:

[0006] A training method for an electricity price prediction model, the method comprising:

[0007] Get real-time price-related data from the database;

[0008] Preprocessing the real-time price related data according to a preset 3sigma criterion algorithm and a preset least squares method to obtain a typical curve of the real-time price and a typical curve of the system load rate;

[0009] The typical curve of the real-time price and the typical curve of the system load rate are processed according to a preset box plot algorithm to obtain a segmentation standard of the system load rate;

[0010] According to the segmentation standard of the system load rate, the historical data with strong correlation is obtained after processing and used as a training data set for the machine learning algorithm;

[0011] Establishing a GBDT algorithm model, inputting the training data set to train the GBDT algorithm model, and obtaining an updated GBDT algorithm model;

[0012] The updated GBDT algorithm model is optimized by using a bionic algorithm to obtain a GBDT optimization model.

[0013] First, the segmentation standard of the system load rate is determined. This segmentation standard can make up for the shortcomings of the existing price forecasting technology that the time period forecast is too mechanical and cannot be fully applied to all forecast days; the GBDT forecasting model is established by combining the bionic parameter adjustment optimization algorithm. Compared with the traditional grid parameter adjustment algorithm, the bionic optimization algorithm combines the advantages of machine learning algorithm and intelligent optimization algorithm, thereby greatly improving the speed and efficiency of parameter adjustment; finally, the deviation correction algorithm is used to correct the deviation of the forecast result, and finally the corrected value of the real-time price is obtained. By correcting the forecast results, not only can the latest real-time price data features be utilized to the maximum extent, but also the abnormal values ​​and deviation values ​​can be corrected in real time, thereby further reducing the error between the forecast value and the true value.

[0014] Furthermore, the real-time price-related data include: date data, meteorological data, network-wide power transmission data, system load rate data, network-wide electricity price data, network-wide installed capacity data and network-wide load data.

[0015] Furthermore, the segmentation standard of the system load rate is to divide the system load rate into a low system load rate segment, a medium system load rate segment and a high system load rate segment.

[0016] Furthermore, the whole network historical data set is fitted respectively according to the preset 3sigma criterion algorithm and the preset least square method to obtain a typical curve of the real-time price and the system load rate.

[0017] Furthermore, the historical data with strong correlation is obtained after processing according to the segmentation standard of the system load rate as a training data set for the machine learning algorithm, including:

[0018] According to the segmentation standard of the system load rate, a segmentation result of the system load rate is obtained;

[0019] Extracting data features of each segment of the system load rate from the segmented results of the system load rate;

[0020] For the data characteristics of each segment of the system load rate, corresponding selection is performed in combination with pre-stored historical data to obtain historical data with strong correlation and use it as a training data set for the machine learning algorithm.

[0021] An application method of an electricity price prediction model, wherein the electricity price prediction model is trained using any one of the above-mentioned training methods for an electricity price prediction model; the application method comprises:

[0022] Get the dataset for the forecast day;

[0023] According to the data set, determining the system load rate segment to which the time to be predicted belongs;

[0024] For the system load rate segment, corresponding selection is performed in combination with pre-stored historical data to obtain historical data with strong correlation and use it as a test data set.

[0025] Calling the GBDT optimization model of the electricity price prediction model training method;

[0026] Input the test data set into the GBDT optimization model and perform rolling prediction to obtain a real-time price prediction value for a certain window period;

[0027] Performing deviation correction processing on the real-time price forecast value of the certain window period length to obtain a corrected real-time price;

[0028] Furthermore, the system load rate segment includes: a high system load rate segment, a medium system load rate segment and a low system load rate segment.

[0029] Further, the deviation correction process is performed on the predicted value of the real-time price, specifically including:

[0030] Using a deviation correction method to correct the predicted value of the real-time price for the high system load rate segment and the flat system load rate segment;

[0031] The predicted value of the real-time price is corrected by setting the low system load rate segment to zero.

[0032] Further, after performing deviation correction processing on the predicted value of the real-time price to obtain a corrected real-time price, the method further includes:

[0033] Check whether the real-time prices of all the opening period lengths have been predicted. If not, determine the system load rate segment to which the time to be predicted belongs based on the data set again.

[0034] Furthermore, the data set includes:

[0035] Meteorological data on the forecast day, power transmission data of the entire network on the forecast day, system load rate data on the forecast day, electricity price data of the entire network on the forecast day, installed capacity data of the entire network on the forecast day and load data of the entire network on the forecast day.

[0036] Compared with the prior art, the advantages of the present invention are:

[0037] A training method and application method of an electricity price prediction model are divided into three parts: the first part first determines the segmentation standard of the system load rate. This segmentation standard can make up for the shortcomings of the existing price prediction technology that the time period prediction is too mechanical and cannot be fully applied to all prediction days; the second part: the GBDT prediction model is established in combination with the bionic parameter adjustment optimization algorithm. Compared with the traditional grid parameter adjustment algorithm, the bionic optimization algorithm combines the advantages of the machine learning algorithm and the intelligent optimization algorithm, thereby greatly improving the speed and efficiency of the parameter adjustment; the third part: finally, the deviation correction algorithm is used to correct the deviation of the prediction result, and finally the corrected value of the real-time price is obtained. By correcting the prediction result, not only can the latest real-time price data characteristics be utilized to the maximum extent, but also the abnormal values ​​and deviation values ​​can be corrected in real time, thereby further reducing the error between the predicted value and the true value.

[0038] A training method and application method of an electricity price prediction model combines the advantages of multiple algorithms, which can make up for the shortcomings of poor generalization ability in existing price prediction technologies, improve the speed of training and parameter adjustment, and reduce the error between the predicted value and the true value. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 , a specific flow chart of a training method and application method of an electricity price prediction model of the present invention;

[0040] Figure 2 , a schematic diagram of a GBDT and bionic optimization algorithm model of a training method and application method of an electricity price prediction model of the present invention;

[0041] Figure 3 , Example 3: The prediction results of April 10 of a training method and application method of an electricity price prediction model;

[0042] Figure 4 , Example 3: The prediction results of April 11 of a training method and application method of an electricity price prediction model;

[0043] Figure 5 , Example 3: The prediction results of April 12 of a training method and application method of an electricity price prediction model;

[0044] Figure 6 , Example 3: The prediction results of April 13 of a training method and application method of an electricity price prediction model;

[0045] Figure 7 , Example 3: The prediction results on April 14 of a training method and application method of an electricity price prediction model. DETAILED DESCRIPTION

[0046] The specific implementation mode of the present invention is described below in conjunction with embodiments:

[0047] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0048] At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" cited in this specification are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments to their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0049] Embodiment 1:

[0050] like Figure 1 As shown, a training method for an electricity price prediction model comprises:

[0051] Get real-time price-related data from the database;

[0052] Preprocessing the real-time price related data according to a preset 3sigma criterion algorithm and a preset least squares method to obtain a typical curve of the real-time price and a typical curve of the system load rate;

[0053] The typical curve of the real-time price and the typical curve of the system load rate are processed according to a preset box plot algorithm to obtain a segmentation standard of the system load rate;

[0054] According to the segmentation standard of the system load rate, the historical data with strong correlation is obtained after processing and used as a training data set for the machine learning algorithm;

[0055] Establishing a GBDT algorithm model, inputting the training data set to train the GBDT algorithm model, and obtaining an updated GBDT algorithm model;

[0056] The updated GBDT algorithm model is optimized by using a bionic algorithm to obtain a GBDT optimization model.

[0057] First, the segmentation standard of the system load rate is determined. This segmentation standard can make up for the shortcomings of the existing price forecasting technology that the time period forecast is too mechanical and cannot be fully applied to all forecast days; the GBDT forecasting model is established by combining the bionic parameter adjustment optimization algorithm. Compared with the traditional grid parameter adjustment algorithm, the bionic optimization algorithm combines the advantages of machine learning algorithm and intelligent optimization algorithm, thereby greatly improving the speed and efficiency of parameter adjustment; finally, the deviation correction algorithm is used to correct the deviation of the forecast result, and finally the corrected value of the real-time price is obtained. By correcting the forecast results, not only can the latest real-time price data features be utilized to the maximum extent, but also the abnormal values ​​and deviation values ​​can be corrected in real time, thereby further reducing the error between the forecast value and the true value.

[0058] The real-time price-related data include: date data, meteorological data, network-wide power transmission data, system load rate data, network-wide electricity price data, network-wide installed capacity data and network-wide load data.

[0059] The segmentation standard of the system load rate is to divide the system load rate into a low system load rate segment, a medium system load rate segment and a high system load rate segment.

[0060] The whole network historical data set is fitted respectively according to the preset 3sigma criterion algorithm and the preset least square method to obtain a typical curve of the real-time price and the system load rate.

[0061] The historical data with strong correlation is obtained after processing according to the segmentation standard of the system load rate as a training data set for the machine learning algorithm, including:

[0062] According to the segmentation standard of the system load rate, a segmentation result of the system load rate is obtained;

[0063] Extracting data features of each segment of the system load rate from the segmented results of the system load rate;

[0064] For the data characteristics of each segment of the system load rate, corresponding selection is performed in combination with pre-stored historical data to obtain historical data with strong correlation and use it as a training data set for the machine learning algorithm.

[0065] An application method of an electricity price prediction model, wherein the electricity price prediction model is trained using any one of the above-mentioned training methods for an electricity price prediction model; the application method comprises:

[0066] Get the dataset for the forecast day;

[0067] According to the data set, determining the system load rate segment to which the time to be predicted belongs;

[0068] For the system load rate segment, corresponding selection is performed in combination with pre-stored historical data to obtain historical data with strong correlation and use it as a test data set.

[0069] Calling the GBDT optimization model of the electricity price prediction model training method;

[0070] Input the test data set into the GBDT optimization model and perform rolling prediction to obtain a real-time price prediction value for a certain window period;

[0071] Performing deviation correction processing on the real-time price forecast value of the certain window period length to obtain a corrected real-time price;

[0072] The system load rate segment includes: a high system load rate segment, a medium system load rate segment and a low system load rate segment.

[0073] The deviation correction process for the predicted value of the real-time price specifically includes:

[0074] Using a deviation correction method to correct the predicted value of the real-time price for the high system load rate segment and the flat system load rate segment;

[0075] The predicted value of the real-time price is corrected by setting the low system load rate segment to zero.

[0076] After performing deviation correction processing on the predicted value of the real-time price to obtain a corrected real-time price, the method further includes:

[0077] Check whether the real-time prices of all the opening period lengths have been predicted. If not, determine the system load rate segment to which the time to be predicted belongs based on the data set again.

[0078] The dataset includes:

[0079] Meteorological data on the forecast day, power transmission data of the entire network on the forecast day, system load rate data on the forecast day, electricity price data of the entire network on the forecast day, installed capacity data of the entire network on the forecast day and load data of the entire network on the forecast day.

[0080] Embodiment 2:

[0081] The second embodiment is applied to the first embodiment, and uses the historical data information of the entire network where the unit is located, including the electricity price data of the entire network, the load data of the entire network, the power transmission data of the entire network, the installed capacity of the entire network, and date-related data (whether it is a holiday, week type, quarter type) and weather data, to establish a set of prediction models of system load rate segmentation and bionic optimization algorithms with stronger generalization ability and more efficient parameter adjustment. The modeling process is as follows: Figure 1As shown. First, using the historical data of the most recent week, the typical curves of real-time price and system load rate are fitted according to the 3sigma (3σ) criterion algorithm and the least squares method, and then the box plot algorithm is used to calculate the two segmentation points of the typical curve of real-time price and the two corresponding segmentation points of the typical curve of system load rate corresponding to the segmentation points, thereby dividing the 96 moments of the whole day into three sections: low system load rate, average system load rate, and high system load rate. Then, combining the data characteristics of each section, the characteristic data of the same distribution is obtained from the historical data, and then the corresponding prediction model is trained through the machine learning algorithm, and the model is optimized using the bionic optimization algorithm. Finally, the model and the segmented correction algorithm of the system load rate are applied to accurately predict the real-time price of a certain window period in the future.

[0082] 3.1 Determination of system load rate by segment

[0083] Use historical data to obtain the real-time price data corresponding to 96 times every day in the past week. ij (i=1,2,...,7,j=1,2,...,96) and system load factor data R ij (i=1,2,...,7,j=1,2,...,96), and then calculate the mean of the real-time price and system load rate at each moment respectively Combined with the least squares formula (3) and formula (4), we can get the typical curve of real-time price and the typical curve of system load rate respectively. Finally, we use the box plot algorithm to obtain the system load rate values ​​corresponding to the lower quartile Q1 and upper quartile Q3 of the real-time electricity price in the typical curve of system load rate, which are used as the upper limit of the low system load rate segment and the upper limit of the flat system load rate segment, respectively, so as to divide the system load rate into three segments: low, flat and high.

[0084]

[0085]

[0086]

[0087]

[0088] in:

[0089] minf P (x) represents the typical curve function of real-time price after fitting;

[0090] minf R (x) represents the typical curve function of the system load rate after fitting;

[0091] 3.2 Establishing GBDT algorithm model and adjusting bionic optimization algorithm

[0092] like Figure 2 As shown in the figure, based on the calculated system load rate segmentation point, when using the rolling forecast (forecast every 15 minutes) method to predict the real-time price of a certain window period in the future, first determine the segment to which the moment belongs according to the system load rate corresponding to each moment in the future window period, and then select the training data set corresponding to the segment trend from the historical data according to the data characteristics of the segment, and then use all the data to establish N decision trees with a depth of D, and generate the corresponding weak learner And the weights θ learned by each weak learner i (i=1,2,...,N), and finally all weak learners are combined to form a strong learner algorithm model. In the GBDT algorithm model, the bionic optimization algorithm is used to optimize the number of decision trees, the minimum number of samples of the decision tree cotyledons, the maximum depth of the decision tree, and the maximum feature number ratio of the data, so as to train a real-time price prediction model with stronger generalization ability.

[0093] 3.3 Algorithm correction of prediction data deviation

[0094] The prediction model formed by adjusting the parameters of the GBDT algorithm model and the bionic optimization algorithm can predict the real-time price of a certain window period in the future. However, the predicted results may still have certain errors, so according to the characteristics of the real-time prices in different system load rate segments and the relationship between the real-time prices and the day-ahead prices, the deviation correction method can be used to correct the predicted data in the high system load rate segment and the flat system load rate segment, and the zeroing method can be used to correct the predicted data in the low system load rate segment.

[0095] Correction algorithms for high system load factor segments and flat system load factor segments:

[0096]

[0097]

[0098] Correction algorithms for low system load rate segments and flat system load rate segments:

[0099]

[0100] in:

[0101] M represents the length of the window period;

[0102] P i(日前) It represents the day-ahead price at the i-th moment before the forecast time;

[0103] P i(实时) Represents the real-time price at the i-th moment before the prediction moment;

[0104] P represents the mean difference between the day-ahead price and the real-time price in the M-window period forward from the prediction time;

[0105] P i(预测) It represents the predicted value of the real-time price at the i-th moment after the prediction moment;

[0106] P i(修正) It represents the correction value of the real-time price at the i-th moment after the prediction moment;

[0107] Embodiment three:

[0108] like Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown, based on the real-time electricity price data of a wind farm in Shanxi in 2021, the load data of the entire network, the new energy load data, the proportion of power transmitted, as well as the historical data such as meteorology and holidays, the rolling forecasting method (forecasting every 15 minutes) is used to predict the real-time price data of the wind farm for the next four window periods corresponding to each moment from April 10, 2021 to April 14, 2021.

[0109] As shown in Tables 1 and 2, in order to verify the advantages and generalization ability of the system load rate segmentation and bionic optimization algorithm, three different schemes are used to compare and analyze the prediction results. Scheme 1: Only use the system load rate segmentation and GBDT algorithm, and set the number of decision trees to 100, the maximum depth of the decision tree to 3, the minimum number of samples of the cotyledon to 2, and the maximum feature ratio to 1; Scheme 2: Add bionic optimization algorithm adjustment on the basis of Scheme 1, and the number of decision trees, the maximum depth of the decision tree, the minimum number of samples of the cotyledon, and the maximum feature ratio are all optimized and adjusted through this optimization algorithm; Scheme 3: Use the deviation algorithm to correct on the basis of Scheme 2.

[0110] In order to verify the performance of the model, since it is common for real-time electricity prices to be 0, the traditional MAPE algorithm cannot evaluate the prediction results of real-time electricity prices, so the improved MAPE algorithm is used. The algorithm formula is as follows:

[0111]

[0112]

[0113] in:

[0114] P i(真实) Indicates the true value of the real-time price corresponding to the i-th moment in the 96 moments of a day;

[0115] P′ i(修正)Indicates the revised value of the real-time price corresponding to the ith moment in the 96 moments of a day;

[0116] P represents the average of the real-time price correction values ​​corresponding to 96 moments in a day;

[0117] In order to verify the advantages of the real-time electricity price forecasting method based on system load rate segmentation and bionic optimization algorithm, the rolling forecasting (forecast every 15 minutes) method is used to predict the real-time price data of the wind farm for the next four window periods corresponding to each moment from April 10, 2021 to April 14, 2021, and the predicted value at the last moment of the window period is used as a comparison object.

[0118] Table 1 Test plan

[0119]

[0120] Table 2 Real-time price prediction results for some dates

[0121]

[0122] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

[0123] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. An application method of an electricity price prediction model, characterized in that: The application method comprises: Get the dataset for the forecast day; According to the data set, determining the system load rate segment to which the time to be predicted belongs; For the system load rate segment, corresponding selection is performed in combination with pre-stored historical data to determine a test data set; Call GBDT optimization model; Input the test data set into the GBDT optimization model and perform rolling prediction to obtain a real-time price prediction value for a certain window period; Performing deviation correction processing on the real-time price forecast value of the certain window period length to obtain a corrected real-time price; The training process of the electricity price prediction model is as follows: Get real-time price-related data from the database; Preprocessing the real-time price related data according to a preset 3sigma criterion algorithm and a preset least squares method to obtain a typical curve of the real-time price and a typical curve of the system load rate; The typical curve of the real-time price and the typical curve of the system load rate are processed according to a preset box plot algorithm to obtain a segmentation standard of the system load rate; According to the segmentation standard of the system load rate, characteristic data of the same distribution is obtained from the historical data in combination with the data characteristics of each segment; Establishing a GBDT algorithm model, inputting the training data set to train the GBDT algorithm model, and obtaining an updated GBDT algorithm model; Optimizing the parameters of the updated GBDT algorithm model using a bionic algorithm to obtain a GBDT optimization model; The segmentation standard of the system load rate is to divide the system load rate into a low system load rate segment, a medium system load rate segment and a high system load rate segment.

2. The application method of the electricity price prediction model according to claim 1, characterized in that: The real-time price-related data include: date data, meteorological data, network-wide power transmission data, system load rate data, network-wide electricity price data, network-wide installed capacity data and network-wide load data.

3. The application method of the electricity price prediction model according to claim 1, characterized in that: The historical data set of the entire network is fitted according to the preset 3sigma criterion algorithm and the preset least squares method to obtain the typical curves of real-time price and system load rate.

4. The application method of an electricity price prediction model according to claim 1, characterized in that: The deviation correction process for the predicted value of the real-time price specifically includes: For the high system load rate segment and the flat system load rate segment, a deviation correction method is used to correct the predicted value of the real-time price; For the low system load rate segment, the predicted value of the real-time price is corrected by setting it to zero.

5. The application method of an electricity price prediction model according to claim 1, characterized in that: After performing deviation correction processing on the predicted value of the real-time price to obtain a corrected real-time price, the method further includes: Check whether the real-time prices of all window period lengths have been predicted. If not, determine the system load rate segment to which the time to be predicted belongs based on the data set again.

6. The application method of an electricity price prediction model according to claim 1, characterized in that: The dataset includes: Meteorological data on the forecast day, power transmission data of the entire network on the forecast day, system load rate data on the forecast day, electricity price data of the entire network on the forecast day, installed capacity data of the entire network on the forecast day and load data of the entire network on the forecast day.

Images