Short-term load forecasting method based on feature selection and LSSVR
Through the combination of feature selection and LSSVR, suitable feature combinations are screened out and prediction models are established, which solves the problem of insufficient load prediction accuracy in the prior art, and effectively fits and high-precision prediction of various influencing factors are achieved.
Patent Information
- Application Number
- CN202210220703.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-08
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-03-08
AI Technical Summary
The prior art is difficult to effectively consider the nonlinear complex relationship of multiple influencing factors in short-term load prediction, resulting in insufficient prediction accuracy. The feature selection method is mainly aimed at individual features rather than feature combinations, and cannot make full use of historical load information.
Using a method based on feature selection and LSSVR, the candidate feature set is determined by acquiring and preprocessing load data, and the sparse regular terms of the L1 paradigm are used to select features, and appropriate feature combinations are selected, and the LSSVR prediction model is established, and a combination of similar samples is used for feature selection. Load forecast.
The load prediction accuracy is improved, the model size is reduced, the overfitting phenomenon is avoided, and the nonlinear complex relationship between various influencing factors and load is achieved, with high prediction accuracy and generalization ability.
Smart Images

Figure CN114581141B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of short-term load forecasting for power systems, and in particular to a short-term load forecasting method based on feature selection and LSSVR. Background Art
[0002] The electric power industry is the lifeline of urban production and life. With the continuous and rapid development of smart grids, theoretical research on short-term load forecasting has become increasingly important. Improving the prediction accuracy is of great significance to the real-time dispatching of the power grid and the safe and economical operation of the power system.
[0003] The rapid development of artificial intelligence and machine learning in recent years has provided new ideas for load forecasting. Among them, support vector regression (SVR) is based on the principle of structural risk minimization, has the advantages of global optimality and strong generalization ability, and is widely used in small sample nonlinear regression problems. Least squares support vector machine (LSSVR) has fewer parameters and faster solution while retaining the excellent performance of SVR. The selection of input features is a key factor affecting the prediction effect of LSSV. Because of the nonlinearity and randomness of load, as well as multiple influencing factors, feature selection is more important. Subjective screening mainly considers the date type of the day to be predicted, meteorological information, daily average load in historical load, hourly load at the same time as the time to be predicted, and other obvious related information, while omitting important information such as the highest and lowest loads of the day in history, and hourly loads a few hours before and after the time to be predicted. Existing feature selection methods mainly adopt methods such as gray correlation analysis or feature scoring based on information theory. This method selects individual features and cannot consider feature combinations, so only some influencing factors can be selected, while the rest of the influencing factors are completely discarded, making it difficult to ensure prediction accuracy.
[0004] Therefore, it is of great significance to design a short-term load forecasting method based on feature selection and LSSVR, taking into account the prediction model that needs to fit the function mapping of the nonlinear and complex relationship between multiple influencing factors and load and the feature selection method with the ability to screen feature combinations. Summary of the invention
[0005] The present invention aims to solve at least one of the technical problems in the related art to a certain extent. To this end, one object of the present invention is to provide a short-term load forecasting method based on feature selection and LSSVR, so as to make full use of historical load information and load influencing factor information, improve load forecasting accuracy and have a larger scope of application.
[0006] To achieve the above object, the present invention is implemented through the following technical solutions:
[0007] A short-term load forecasting method based on feature selection and LSSVR, comprising:
[0008] Step S1: acquiring load data, meteorological data and calendar information of a region to be predicted, and performing data preprocessing on the load data, meteorological data and calendar information of the region to be predicted;
[0009] Step S2: Determine the candidate feature set as the input of the LSSVR prediction model based on the preprocessed data;
[0010] Step S3: using the LSSVR algorithm with L1 paradigm to perform feature selection on the candidate feature set, and obtaining the input feature vectors of the final historical data samples and the samples to be predicted at a preset future time according to the selected features;
[0011] Step S4: determining the similarity between the input feature vector of the sample to be predicted at a preset moment and the input feature vectors of all historical data samples at corresponding moments, and filtering out similar samples corresponding to the sample to be predicted from the historical data samples according to the similarity;
[0012] Step S5: The LSSVR prediction model is obtained by training similar samples corresponding to each sample to be predicted, and the load at a preset time in the future is predicted by the LSSVR prediction model.
[0013] Optionally, the data in the candidate feature set includes at least one of load-related meteorological data, week type data, daily load data and hourly load data.
[0014] Optionally, in step S1, the data preprocessing is to perform data normalization preprocessing on the load data of the area to be predicted, the meteorological data and the numerical data in the calendar information using the following formula:
[0015]
[0016] Among them, x is the eigenvalue corresponding to the numerical data, x min is the minimum value among the eigenvalues, x max is the maximum value among the eigenvalues, is the preprocessing value corresponding to the eigenvalue.
[0017] Optionally, step S2 includes:
[0018] Step S21: Calculate the daily average load, daily meteorological data, daily average load mean, daily maximum load, daily minimum load, hourly load data at the beginning and end of the day, and hourly load data at the predicted time and before and after the predicted time according to the preprocessed data, wherein the daily load data includes at least one of the daily average load mean, daily maximum load, daily minimum load, and hourly load data at the beginning and end of the day, and the hourly load data includes hourly load data at the predicted time and before and after the predicted time;
[0019] Step S22: Determine the load-related meteorological data by calculating the daily average load and the daily meteorological data; and determine the weekly type data by excluding the average daily average load of the same type of days including holidays.
[0020] Optionally, the daily meteorological data includes at least one of daily average temperature, daily maximum temperature, daily minimum temperature, daily average relative humidity and daily rainfall, and the step of determining the load-related meteorological data in step S22 includes:
[0021] Step S221: Determine the first Pearson correlation coefficient between the daily average load data and the daily meteorological data. The first Pearson correlation coefficient is expressed by the following formula:
[0022]
[0023] Among them, X i is the average load data of the ith day, Y i is the meteorological data of the ith day, is the average load value of each day, is the mean of daily meteorological data, m is the number of daily average load data and daily meteorological data, ρ XY is the first Pearson correlation coefficient;
[0024] Step S222: selecting daily meteorological data corresponding to a first Pearson correlation coefficient greater than a first preset threshold from each first Pearson correlation coefficient as the load-related meteorological data.
[0025] Optionally, step S3 includes:
[0026] Step S31: obtaining a number of historical data samples, and selecting a number of samples to be predicted at a corresponding future preset time from the candidate feature set, and establishing a first sample set based on the historical data samples and the samples to be predicted;
[0027] Step S32: establishing an LSSVR optimization problem with an L1 sparse regularization term according to the first sample set;
[0028] Step S33: Use the alternating direction multiplier method ADMM to solve the LSSVR optimization problem to obtain a set of feature solution vectors, and select feature values greater than a second preset threshold from the feature solution vectors as final features, so as to obtain the final historical data samples and the input feature vectors of the samples to be predicted at a preset time in the future through the final features.
[0029] Optionally, step S4 includes:
[0030] Step S41: determining a second Pearson correlation coefficient between the input feature vector of the sample to be predicted at a preset time and the input feature vectors of all historical data samples at corresponding times;
[0031] Step S42: selecting, from the determined second Pearson correlation coefficients, historical data samples corresponding to second Pearson correlation coefficients that are greater than a third preset threshold as similar samples corresponding to the sample to be predicted.
[0032] Optionally, step S5 includes:
[0033] Step S51: obtaining several similar sample hourly load data and several corresponding input feature quantities, and establishing a second sample set, and constructing a regression function according to the second sample set, wherein the regression function includes unknown regression parameters, and the regression function is the LSSVR prediction model;
[0034] Step S52: According to the principle of structural risk minimization, a constrained optimization problem is constructed, and the constrained optimization problem is solved to obtain the regression parameters to determine the LSSVR prediction model.
[0035] Compared with the prior art, the present invention has at least the following technical effects:
[0036] The present invention fully considers the various characteristic factors that affect the power load, and makes full use of historical load information, proposes a candidate feature set, and embeds feature selection into the LSSVR learning algorithm through the method of sparse regular terms, so that all features can be considered at the same time, and a suitable feature combination can be screened out, and then the features selected by the features are used to establish the LSSVR prediction model, which reduces the model scale and avoids overfitting. Therefore, the model obtained by subjective screening or scoring the features alone to achieve feature selection has better prediction performance. In addition, the present invention adopts the LSSVR method to realize data modeling, which has the characteristics of few parameters, fast solution speed, and strong generalization ability, and can better realize the fitting of the nonlinear complex relationship between various influencing factors and loads, and the method has a lower average absolute percentage error and a higher prediction accuracy.
[0037] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 A flow chart of a short-term load forecasting method based on feature selection and LSSVR provided by an embodiment of the present invention;
[0039] Figure 2A flowchart of a computational framework for solving a distributed convex optimization problem using an ADMM algorithm provided by an embodiment of the present invention;
[0040] Figure 3 A schematic diagram for comparing load forecasting results under different input characteristics provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0041] The present embodiment is described in detail below, and examples of the embodiment are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.
[0042] The short-term load forecasting method based on feature selection and LSSVR of this embodiment is described below with reference to the accompanying drawings.
[0043] Figure 1 Flow chart of a short-term load forecasting method based on feature selection and LSSVR provided by an embodiment of the present invention. Figure 1 As shown in Figure 2, the short-term load forecasting method based on feature selection and LSSVR includes:
[0044] Step S1: Obtain load data, meteorological data and calendar information of the area to be predicted, and perform data preprocessing on the load data, meteorological data and calendar information of the area to be predicted.
[0045] In step S1, data preprocessing is to perform data normalization preprocessing on the numerical data in the load data, meteorological data and calendar information of the forecasted area using the following formula:
[0046]
[0047] Among them, x is the eigenvalue corresponding to the numerical data, x min is the minimum value among the eigenvalues, x max is the maximum value among the eigenvalues, is the preprocessing value corresponding to the eigenvalue.
[0048] Step S2: Determine the candidate feature set as the input of the LSSVR prediction model based on the preprocessed data.
[0049] The data in the candidate feature set includes at least one of load-related meteorological data, week type data, daily load data and hourly load data.
[0050] The step S2 comprises:
[0051] Step S21: Calculate the daily average load, daily meteorological data, daily average load mean, daily maximum load, daily minimum load, hourly load data at the beginning and end of the day, and hourly load data before and after the predicted time based on the preprocessed data, wherein the daily load data includes at least one of the daily average load mean, daily maximum load, daily minimum load, and hourly load data at the beginning and end of the day, and the hourly load data includes hourly load data at the predicted time and before and after.
[0052] In this embodiment, the calculated daily average load mean, daily maximum load, daily minimum load and hourly load data at the beginning and end of the day are all daily load data; the hourly load data at the predicted time and before and after are hourly load data, both of which can be directly obtained through the calculation results.
[0053] Step S22: Determine load-related meteorological data by calculating the daily average load and daily meteorological data; and determine the weekly type data by excluding the average daily load value of the same type of days on holidays.
[0054] The daily meteorological data includes at least one of daily average temperature, daily maximum temperature, daily minimum temperature, daily average relative humidity and daily rainfall. The step of determining the load-related meteorological data in step S22 includes:
[0055] Step S221: Determine the first Pearson correlation coefficient between the daily average load data and the daily meteorological data. The first Pearson correlation coefficient is expressed by the following formula:
[0056]
[0057] Among them, X i is the average load data of the ith day, Y i is the meteorological data of the ith day, is the average load value of each day, is the mean of daily meteorological data, m is the number of daily average load data and daily meteorological data, ρ XY is the first Pearson correlation coefficient.
[0058] Step S222: selecting daily meteorological data corresponding to a first Pearson correlation coefficient greater than a first preset threshold from each first Pearson correlation coefficient as load-related meteorological data.
[0059] Specifically, in order to predict the hourly load of the next day, it is necessary to consider the factors affecting the daily load mean and the shape of the daily load curve. Therefore, the candidate feature set includes week type data, load-related meteorological data, daily load data and hourly load data. Among them, the week type data is represented by the average daily load of the same type of days excluding holidays and missing days, which affects the daily average load and the shape of the load curve; the load-related meteorological data mainly affects the daily average load, which can be determined by the first Pearson correlation coefficient between the daily meteorological data and the daily average load, that is, determined by the above formula (2).
[0060] It can be understood that the larger the first Pearson correlation coefficient, the greater the correlation between the two. Generally, |ρ XY |>0.4 daily meteorological data, removing weakly correlated or irrelevant meteorological factors.
[0061] In this embodiment, the daily load data includes the historical daily average load that mainly affects the daily average load and the daily maximum load and daily minimum load that mainly affect the shape of the daily load curve, and the hourly load data includes the hourly load data before and after the forecast time.
[0062] As an example, the candidate feature set may be as shown in Table 1 below:
[0063] Table 1 Candidate feature set
[0064]
[0065]
[0066] Step S3: Use the LSSVR algorithm with L1 paradigm to perform feature selection on the candidate feature set, and obtain the input feature vectors of the final historical data samples and the samples to be predicted at a preset time in the future based on the selected features.
[0067] If the future preset time is the next 24 hours, since the power load is affected by multiple factors, in order to reduce the input dimension of the LSSVR prediction model and improve the calculation efficiency without affecting the prediction accuracy, the present invention adopts the L1 sparse regularization term instead of the L2 paradigm LSSVR model optimization to achieve feature selection of the original load candidate feature set.
[0068] The step S3 comprises:
[0069] Step S31: Acquire a number of historical data samples, select a number of samples to be predicted at a corresponding preset future time from the candidate feature set, and establish a first sample set based on the historical data samples and the samples to be predicted.
[0070] Step S32: Establishing an LSSVR optimization problem with an L1 sparse regularization term according to the first sample set.
[0071] Step S33: Use the alternating direction multiplier method ADMM to solve the LSSVR optimization problem to obtain a set of feature solution vectors, and select the feature values greater than the second preset threshold from the feature solution vectors as the final features, so as to obtain the final historical data samples and the input feature vectors of the samples to be predicted at the preset time in the future through the final features.
[0072] Specifically, we can select l historical data samples, such as l historical hourly load data, and record them as y i , and select the samples to be predicted from the candidate feature set, such as selecting n candidate input feature quantities corresponding to each hour, recorded as x i , the x i ∈R n , R is a real number set, and then establish the first sample set (x i ,y i ), where i=1,2,…l.
[0073] Furthermore, based on the first sample set (x i ,y i ) Establish an LSSVR optimization problem with an L1 sparse regularization term, and the LSSVR optimization problem is expressed by the following formula:
[0074]
[0075] sty i =w T x i +b+e i ,i=1,……,l (4)
[0076] Among them, J is the established LSSVR optimization problem, w and b are regression parameters, and e i is the fitting error, and C is the penalty coefficient, which is set to 100.
[0077] Furthermore, the above optimization problem is transformed into the following problem:
[0078]
[0079] stu-z=0 (6)
[0080] Among them, Y1=[y1,y2,……,y n ] T , z=[w;b],u is the relevant parameter, and then the alternating direction multiplier method ADMM is used to solve the above optimization problem. Among them, the ADMM algorithm is a computational framework for solving distributed convex optimization problems. The flow chart is as follows Figure 2 As shown, the algorithm principle is as follows:
[0081] 1) Set the algorithm parameters. In the present invention, ρ is set to 0.1.
[0082] 2) Initialize the algorithm, set k = 1, and randomly select u 0 ,d 0 ∈R n+1 .
[0083] 3) Update the coefficient according to the following formula:
[0084]
[0085] Where B = CG T Y1+ρI, I is the unit matrix, where v and d are related parameters.
[0086] 4) If the following stop conditions are not met:
[0087]
[0088]
[0089] Or, k<500, then return to step 3), otherwise the process ends and the optimal solution (w * ,b * )=z * , thus obtaining the eigensolution vector Then select from the feature solution vectors those values that are greater than a second preset threshold value, such as 1×10 -3 The characteristic value of is used as the final feature to reconstruct the input feature vector of the final historical data sample and the sample to be predicted, where ∈ abs and ∈ rel are the absolute error coefficient and the relative error coefficient, and they are 10 -3 and 10 -4 , the final number of features selected at hour h every day is n h .
[0090] Step S4: Determine the similarity between the input feature vector of the sample to be predicted at a preset moment and the input feature vectors of all historical data samples at corresponding moments, and select similar samples corresponding to the sample to be predicted from the historical data samples according to the similarity.
[0091] The step S4 comprises:
[0092] Step S41: Determine a second Pearson correlation coefficient between the input feature vector of the sample to be predicted at a preset time and the input feature vectors of all historical data samples at corresponding times.
[0093] Step S42: selecting, from the determined second Pearson correlation coefficients, historical data samples corresponding to second Pearson correlation coefficients that are greater than a third preset threshold as similar samples corresponding to the sample to be predicted.
[0094] Among them, the preset moment is the moment to be predicted. In this embodiment, for the samples to be predicted every hour, similar samples are selected according to the similarity between the input feature vector of the sample to be predicted and the input feature vector of the historical data samples of all corresponding moments, such as calculating the second Pearson correlation coefficient between the input feature vector of the sample to be predicted and the input feature vector of the historical data samples of all corresponding moments, and then taking the historical data samples whose second Pearson correlation coefficient is greater than the third preset threshold (such as 0.7-0.8) as similar samples. If there are fewer similar samples, the third preset threshold can be appropriately reduced, but it should not be lower than 0.6. In this embodiment, l1 similar samples can be obtained.
[0095] Step S5: A LSSVR prediction model is obtained by training similar samples corresponding to each sample to be predicted, and the load at a preset time in the future is predicted by the LSSVR prediction model.
[0096] The step S5 comprises:
[0097] Step S51: Acquire several similar sample hourly load data and several corresponding input feature quantities, establish a second sample set, and construct a regression function according to the second sample set, wherein the regression function includes unknown regression parameters, and the regression function is a LSSVR prediction model.
[0098] Step S52: According to the principle of structural risk minimization, a constrained optimization problem is constructed, and the constrained optimization problem is solved to obtain regression parameters to determine the LSSVR prediction model.
[0099] Specifically, the selected similar samples can be used as training samples, such as taking the hourly load data of l1 similar samples as p i and corresponding n h The input feature quantity is q i , Then establish the second sample set (p i ,q i ), i = 1, 2, ..., l1, and construct the following regression function based on the second sample set:
[0100]
[0101] Among them, Φ(q i ) is q i Correlation function, f(q i) is the constructed regression function. Since formula (10) includes unknown regression parameters w1 and b1, the following constrained optimization problem can be constructed based on the principle of structural risk minimization, and the regression parameters w1 and b1 can be obtained by solving the following constrained optimization problem:
[0102]
[0103] stp i =w1 T Φ(q i )+b1+e i ,i=1,2……,l1 (12)
[0104] Among them, J1 is the constrained optimization problem constructed, and C1 is the penalty coefficient.
[0105] Specifically, the kernel function uses the RBF (Radial Basis Function) kernel function:
[0106]
[0107] Among them, K(x i ,x j ) is the RBF kernel function, δ is the kernel width, and then the Lagrange multiplier method is used. According to the KKT condition, we can get:
[0108]
[0109] N is the number of samples to be predicted.
[0110] Formulas (14)-(15) can be used to solve α and b1, so the regression parameters w1 and b1 can be obtained. Substituting them into formula (10) and combining them with formula (13), the following LSSVR prediction model is obtained for load forecasting:
[0111]
[0112] For the penalty coefficient C1 and kernel width δ in the LSSVR prediction model, their specific values can be determined by grid search, that is, the possible values of each parameter are arranged and combined, all possible combination results are listed to generate a "grid", and then each combination is used for LSSVR training, and the performance is evaluated by cross-validation to select the best parameter combination. C1 is in (1,10,10 2 ,10 3 ,10 4 ) is selected, δ 2 In (10 -2 ,10 -1 ,1,10,102 )
[0113] In order to verify the accuracy of the load forecasting method of the present invention, the hourly load forecast of a certain region in China from July 31 to August 3, 2014 was carried out, and the data of the day was included in the training set each time the next day was predicted. The training set data is the historical load data (with an interval of 1 hour) of the region from June to September 2012, 2013 and from June 1 to July 30, 2014, daily weather data and holiday information.
[0114] The mean absolute percentage error (MAPE) is selected to evaluate the short-term load forecasting effect of each method. The smaller the MAPE, the higher the prediction accuracy. MAPE is calculated as follows:
[0115]
[0116] Among them, p i is the load forecast value, r i is the true value of the load.
[0117] 1) candidate feature set, 2) features obtained by feature selection of candidate feature set as shown in this patent, 3) empirical feature set (type of day and week to be predicted, daily average temperature, daily average relative humidity, daily average load of the same type of day three days ago and one week ago, load of the same type of day to be predicted two days ago and one week ago), 4) features obtained by feature selection of candidate feature set based on maximum correlation and minimum redundancy criterion of mutual information as sample input features are used to realize load forecasting according to the similar sample selection method and LSSVR modeling method in the present invention. The forecast error is shown in Table 2. The load forecast results for 7.31 are as follows. Figure 3 shown.
[0118] Table 2 Load forecast results
[0119] feature 7.31 8.1 8.2 8.3 Mean 1 1.77% 1.35% 1.51% 2.40% 0.0176 2 1.60% 1.72% 1.25% 2.28% 0.0171 3 1.72% 1.43% 1.40% 3.03% 0.0190 4 2.43% 1.33% 3.51% 4.05% 0.0283
[0120] By comparing the prediction results obtained with different input features, it can be seen that the prediction accuracy of the candidate feature set selected according to the present invention is significantly improved compared with the empirical feature set, and the feature selection method according to the present invention can further improve the prediction accuracy on the basis of the candidate feature set more stably than the feature selection method based on mutual information. Figure 3 It can be seen that the advantage of using the candidate feature set and the features selected based on the candidate feature set as input for load forecasting is mainly at the beginning and end of the day, while feature selection according to the method shown in the present invention further improves the model prediction accuracy for most forecast moments.
[0121] In summary, the present invention obtains the load and related data of the area to be predicted and performs data preprocessing, then determines the candidate feature set based on the preprocessed data, and trains the LSSVR of the L1 paradigm on the global data to achieve feature selection, and then selects the daily similar data as similar samples according to the similarity of the selected features, and uses the similar samples as training samples to train the LSSVR model to achieve load prediction, so that the present invention fully considers the weather, time and other influencing factors and historical load data information. Under the premise, a power grid load feature selection and prediction method that can integrate multiple data characteristics is established, and the method can reduce the model scale and avoid overfitting at the same time, so that the model obtained by the method has better prediction performance than the model obtained by subjective screening or scoring the features alone to achieve feature selection. In addition, the present invention adopts the LSSVR method to achieve data modeling, which has the characteristics of few parameters, fast solution speed, and strong generalization ability, and can better achieve the fitting of the nonlinear complex relationship between multiple influencing factors and loads, and the method has a lower average absolute percentage error and a higher prediction accuracy.
[0122] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.
[0123] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.
Claims
1. A short-term load forecasting method based on feature selection and LSSVR, characterized in that: include: Step S1: acquiring load data, meteorological data and calendar information of a region to be predicted, and performing data preprocessing on the load data, meteorological data and calendar information of the region to be predicted; Step S2: Determine the candidate feature set as the input of the LSSVR prediction model based on the preprocessed data; Step S3: using the LSSVR algorithm with L1 paradigm to perform feature selection on the candidate feature set, and obtaining the input feature vectors of the final historical data samples and the samples to be predicted at a preset future time according to the selected features; Step S4: determining the similarity between the input feature vector of the sample to be predicted at a preset moment and the input feature vectors of all historical data samples at corresponding moments, and filtering out similar samples corresponding to the sample to be predicted from the historical data samples according to the similarity; Step S5: obtaining the LSSVR prediction model by training similar samples corresponding to each sample to be predicted, and predicting the load at a preset time in the future by using the LSSVR prediction model; The step S3 comprises: Step S31: obtaining a number of historical data samples, and selecting a number of samples to be predicted at a corresponding future preset time from the candidate feature set, and establishing a first sample set based on the historical data samples and the samples to be predicted; Step S32: Establish an LSSVR optimization problem with an L1 sparse regularization term according to the first sample set. The LSSVR optimization problem is expressed by the following formula: s.t.y i =w T x i +b+e i ,i=1,……,l Among them, J is the established LSSVR optimization problem, w and b are regression parameters, and e i is the fitting error, C is the penalty coefficient, and l historical data samples are recorded as y i , the selected sample to be predicted is recorded as x i ; Step S33: Use the alternating direction multiplier method ADMM to solve the LSSVR optimization problem to obtain a set of feature solution vectors, and select feature values greater than a second preset threshold from the feature solution vectors as final features, so as to obtain the final historical data samples and the input feature vectors of the samples to be predicted at a preset time in the future through the final features.
2. The short-term load forecasting method based on feature selection and LSSVR according to claim 1 is characterized in that: The data in the candidate feature set includes at least one of load-related meteorological data, week type data, daily load data and hourly load data.
3. The short-term load forecasting method based on feature selection and LSSVR according to claim 1 is characterized in that: In step S1, the data preprocessing is to perform data normalization preprocessing on the load data of the area to be predicted, the meteorological data and the numerical data in the calendar information using the following formula: Among them, x is the eigenvalue corresponding to the numerical data, x min is the minimum value among the eigenvalues, x max is the maximum value among the eigenvalues, is the preprocessing value corresponding to the eigenvalue.
4. The short-term load forecasting method based on feature selection and LSSVR according to claim 2 is characterized in that: The step S2 comprises: Step S21: Calculate the daily average load, daily meteorological data, daily average load mean, daily maximum load, daily minimum load, hourly load data at the beginning and end of the day, and hourly load data at the predicted time and before and after the predicted time according to the preprocessed data, wherein the daily load data includes at least one of the daily average load mean, daily maximum load, daily minimum load, and hourly load data at the beginning and end of the day, and the hourly load data includes hourly load data at the predicted time and before and after the predicted time; Step S22: Determine the load-related meteorological data by calculating the daily average load and the daily meteorological data; and determine the weekly type data by excluding the average daily average load of the same type of days including holidays.
5. The short-term load forecasting method based on feature selection and LSSVR according to claim 4 is characterized in that: The daily meteorological data includes at least one of daily average temperature, daily maximum temperature, daily minimum temperature, daily average relative humidity and daily rainfall. The step of determining the load-related meteorological data in step S22 includes: Step S221: Determine the first Pearson correlation coefficient between the daily average load data and the daily meteorological data. The first Pearson correlation coefficient is expressed by the following formula: Among them, X i is the average load data of the ith day, Y i is the meteorological data of the ith day, is the average load value of each day, is the mean of daily meteorological data, m is the number of daily average load data and daily meteorological data, ρ XY is the first Pearson correlation coefficient; Step S222: selecting daily meteorological data corresponding to a first Pearson correlation coefficient greater than a first preset threshold from each first Pearson correlation coefficient as the load-related meteorological data.
6. The short-term load forecasting method based on feature selection and LSSVR according to claim 1, characterized in that: The step S4 comprises: Step S41: determining a second Pearson correlation coefficient between the input feature vector of the sample to be predicted at a preset time and the input feature vectors of all historical data samples at corresponding times; Step S42: selecting, from the determined second Pearson correlation coefficients, historical data samples corresponding to second Pearson correlation coefficients that are greater than a third preset threshold as similar samples corresponding to the sample to be predicted.
7. The short-term load forecasting method based on feature selection and LSSVR according to claim 6, characterized in that: The step S5 comprises: Step S51: obtaining several similar sample hourly load data and several corresponding input feature quantities, and establishing a second sample set, and constructing a regression function according to the second sample set, wherein the regression function includes unknown regression parameters, and the regression function is the LSSVR prediction model; Step S52: According to the principle of structural risk minimization, a constrained optimization problem is constructed, and the constrained optimization problem is solved to obtain the regression parameters to determine the LSSVR prediction model.
Citation Information
Patent Citations
Short-term load prediction method based on similar day segmentation and LM-BP network
CN108229754A
Multi-model integrated load prediction method based on wavelet transform
CN110443417A
Short-term power load prediction method based on data mining
CN112132347A
Power load prediction method during typhoon
CN113743673A
Short-term load prediction method based on feature selection and LSSVR
CN114581141A
