Short-term load forecasting method based on PSO-LSSVM considering time interval division

Abstract

This invention discloses a PSO-LSSVM short-term load forecasting method considering time period division, comprising the following steps: Step 1, acquisition; Step 2, calculation; Step 3, establishment; Step 4, substitution; wherein in Step 1, power load data and meteorological statistics are acquired to construct a sample dataset D; this invention calculates the mean absolute error of the LSSVM forecast output under a specific time period, and divides the entire 24 hours into N time sets T1, T2, ..., T based on the changing trend of the mean absolute error. k Based on this, N PSO-LSSVM models are established, each predicting the output of the corresponding time set T using the LSSVM model. k The mean absolute error index is minimized as the fitness function, and the control parameters of the LSSVM model are iteratively optimized to obtain N sets of optimal LSSVM parameters. The N sets of optimal LSSVM parameters are then substituted into N LSSVM models for prediction, and the output results are concatenated to obtain the final short-term load forecast result. Finally, the effectiveness of the proposed method is verified by simulation, providing a new approach for optimizing the accuracy of short-term load forecasting.

Description

Technical Field

[0001] This invention relates to the field of short-term power load forecasting technology, specifically to the PSO-LSSVM short-term load forecasting method that considers time period division. Background Technology

[0002] With the accelerating pace of my country's power market liberalization, accurate short-term load forecasting is crucial for ensuring grid operation safety, improving power generation equipment utilization, assisting electricity retailers in managing deviation assessment risks, and facilitating demand-side response transactions within the power market environment. Common short-term load forecasting methods include backpropagation (BP) neural networks, long short-term memory (LSTM) neural networks, support vector machines (SVM), and least squares support vector machines (LSSVM). However, the accuracy of the forecast results largely depends on the quality of the control parameters. Therefore, intelligent algorithms such as particle swarm optimization (PSO), genetic algorithms, and sparrow algorithms have emerged and are combined with these forecasting algorithms. By minimizing the overall load forecasting error as the fitness function, some control parameters of the forecasting algorithm are optimized, thereby improving the overall load forecasting accuracy.

[0003] However, optimizing control parameters with the goal of minimizing overall load forecasting error makes it difficult to take into account time-of-day load characteristics. This inevitably leads to the discarding of some peak-valley characteristic data during the training process, resulting in reduced local load forecasting accuracy. Therefore, there is still significant room for improvement in overall forecasting accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a PSO-LSSVM short-term load forecasting method that takes into account time period division, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a PSO-LSSVM short-term load forecasting method considering time period division, comprising the following steps: Step 1, obtaining; Step 2, calculating; Step 3, establishing; Step 4, substituting.

[0006] In step one above, power load data and meteorological statistics are obtained to construct a sample dataset D;

[0007] In step two above, an LSSVM model is established to perform regression prediction on the dataset D constructed in step one, and the regression results are calculated over the time set T. k The LSSVM prediction outputs the mean absolute error. Based on the trend of the mean absolute error, the entire 24 hours of the day are finally divided into N time sets T1, T2, ..., T3 with no repeating elements. k ;

[0008] In step three above, N PSO-LSSVM models are established based on the N time sets obtained in step two. Each LSSVM model is trained on the dataset D constructed in step one. The PSO model predicts the output of the LSSVM model in the corresponding time set T. k The mean absolute error index is minimized as the fitness function, and the control parameters of the LSSVM model are iteratively optimized to obtain N sets of optimal LSSVM parameters.

[0009] In step four above, the N sets of optimal LSSVM parameters obtained in step three are substituted into N LSSVM models and used to predict the dataset D constructed in step one. The prediction outputs of the N LSSVM models are then concatenated based on the N time sets obtained in step two to obtain the final short-term load prediction result.

[0010] Preferably, in step one, the specific process of acquiring power load data and meteorological statistics to construct the sample dataset D is as follows:

[0011] 1) The power load data interval is 1 hour;

[0012] 2) Meteorological statistics include daily maximum temperature, minimum temperature, average temperature, and relative humidity data;

[0013] 3) The 24-hour load data of the day before the forecast date, meteorological statistics data, and meteorological statistics data of the forecast date are used as model inputs, and the 24-hour load data of the forecast date is used as model outputs. A pair of data samples are added to the dataset, and finally, after min-max normalization, the sample dataset D is obtained.

[0014] 4) In the sample dataset D, the last pair of samples is the test set, and the remaining samples are the training set.

[0015] Preferably, in step two, an LSSVM model is established to perform regression prediction on the dataset D constructed in step one, and the regression prediction is calculated over the time set T. k The LSSVM prediction outputs the mean absolute error. Based on the trend of the mean absolute error, the entire 24 hours of the day are finally divided into N time sets T1, T2, ..., T3 with no repeating elements. k The specific process is as follows:

[0016] 1) LSSVM prediction output in time set T k The formula for calculating the mean absolute error index is:

[0017]

[0018] Among them, T k Let n be the k-th time set, and let n be the set T. kLength, y tp For the actual value, y tm This is a predicted value;

[0019] 2) Initialization parameters k=1 and time set T for time period division k ={t start ,t start +1,t start +2,t end}, T k The elements in the set represent specific times within a 24-hour period, with the first element being t. start =1, representing 1:00:00, the last element t end =t start +3 represents 4:00:00;

[0020] 3) The LSSVM model uses a radial basis function kernel and is trained based on the dataset D constructed in step one;

[0021] 4) Based on step 2, 1), calculate the LSSVM prediction output in time set T. k The mean absolute error index MAE0;

[0022] 5) Try moving to set T k Add t end +1 element;

[0023] 6) If t end +1>24, then T k Once the construction is complete, exit the loop and output the time period segmentation results T1, T2, ..., T k Otherwise, calculate MAE1 according to step 2(1).

[0024] 7) If MAE1≤MAE0, then T k =T k ∪{T end +1},T end =T end +1;

[0025] 8) If MAE1>MAE0, then T k Construction complete, k = k + 1, T start =T end +1;

[0026] 9) Initialize the new time set T k ={t start ,t start +1,t start +2,t end}, T start =T start +3;

[0027] 10) If the new time set T k If any element is greater than 24, exit the loop and output the time period division results T1, T2, ..., T k Otherwise, repeat steps 2.4) to 2.9).

[0028] Preferably, in step three, N PSO-LSSVM models are established based on the N time sets obtained in step two. Each LSSVM model is trained on the dataset D constructed in step one, and the PSO model predicts the output of the LSSVM model in the corresponding time set T. n The process of iteratively optimizing the control parameters of the LSSVM model by minimizing the mean absolute error index as the fitness function to obtain N sets of optimal LSSVM parameters is as follows:

[0029] 1) The LSSVM control parameters that need to be optimized for each PSO model are the penalty factor C and the kernel parameter σ;

[0030] 2) The PSO models of each group predict the output of the LSSVM model at the corresponding time set T. n The mean absolute error index is minimized as the fitness function, as shown below:

[0031]

[0032] Among them, T k The time set responsible for the k-th PSO-LSSVM model, where n is the set T. k Length, y tp For the actual value, y tm These are predicted values.

[0033] Preferably, in step four, the process of substituting the N sets of optimal LSSVM parameters obtained in step three into N LSSVM models and predicting the dataset D constructed in step one, and then concatenating the prediction outputs of the N LSSVM models based on the N time sets obtained in step two to obtain the final short-term load prediction result is as follows:

[0034] 1) The output of each model group is the 24-hour load data for the prediction day of the test set;

[0035] 2) According to the time set T k The elements represent specific hourly times within a 24-hour period. The load data corresponding to the k-th PSO-LSSVM model at that time is extracted into set L. k ;

[0036] 3) The final model prediction output is L = L1∪L2···L k .

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows: The PSO-LSSVM short-term load forecasting method that considers time period division, based on the historical load time characteristics, divides 24 hours into multiple time sets, establishes multiple PSO-LSSVM models to capture the features between input and output data in each time set area to improve the accuracy of local load forecasting, thereby obtaining more accurate load forecasting results and providing decision support for regional power grid dispatching and stable operation of the power market. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 This is a comparison chart of the prediction results, actual values, and PSO-LSSVM method in the embodiments of the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the working principle of the invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1-3 The present invention provides an embodiment of a PSO-LSSVM short-term load forecasting method considering time period division, comprising the following steps: Step 1, obtaining; Step 2, calculating; Step 3, establishing; Step 4, substituting.

[0043] The specific process of acquiring power load data and meteorological statistics to construct the sample dataset D in step one above is as follows:

[0044] 1) The power load data interval is 1 hour;

[0045] 2) Meteorological statistics include daily maximum temperature, minimum temperature, average temperature, and relative humidity data;

[0046] 3) The 24-hour load data of the day before the forecast date, meteorological statistics data, and meteorological statistics data of the forecast date are used as model inputs, and the 24-hour load data of the forecast date is used as model outputs. A pair of data samples are added to the dataset, and finally, after min-max normalization, the sample dataset D is obtained.

[0047] 4) In the sample dataset D, the last pair of samples is the test set, and the remaining samples are the training set;

[0048] In step two above, an LSSVM model is established to perform regression prediction on the dataset D constructed in step one, and the regression results are calculated over the time set T. k The LSSVM prediction outputs the mean absolute error. Based on the trend of the mean absolute error, the entire 24 hours of the day are finally divided into N time sets T1, T2, ..., T3 with no repeating elements. k The specific process is as follows:

[0049] 1) LSSVM prediction output in time set T k The formula for calculating the mean absolute error index is:

[0050]

[0051] Among them, T k Let n be the k-th time set, and let n be the set T. k Length, y tp For the actual value, y tm This is a predicted value;

[0052] 2) Initialization parameters k=1 and time set T for time period division k ={t start ,t start +1,t start +2,t end}, T k The elements in the set represent specific times within a 24-hour period, with the first element being t. start =1, representing 1:00:00, the last element t end =t start +3 represents 4:00:00;

[0053] 3) The LSSVM model uses a radial basis function kernel, based on the dataset D constructed in step one.

[0054] Conduct training;

[0055] 4) Based on step 2, 1), calculate the LSSVM prediction output in time set T. k The mean absolute error index MAE0;

[0056] 5) Try moving to set T k Add t end +1 element;

[0057] 6) If t end +1>24, then T k Once the construction is complete, exit the loop and output the time period segmentation results T1, T2, ..., T k Otherwise, calculate MAE1 according to step 2(1).

[0058] 7) If MAE1≤MAE0, then T k =T k ∪{T end +1},T end =T end +1;

[0059] 8) If MAE1>MAE0, then T k Construction complete, k = k + 1, T start =T end +1;

[0060] 9) Initialize the new time set T k ={t start ,t start +1,t start +2,t end}, T start =T start +3;

[0061] 10) If the new time set T k If any element is greater than 24, exit the loop and output the time period division results T1, T2, ..., T k Otherwise, repeat steps 2.4) to 2.9);

[0062] In step three above, N PSO-LSSVM models are established based on the N time sets obtained in step two. Each LSSVM model is trained on the dataset D constructed in step one. The PSO model predicts the output of the LSSVM model in the corresponding time set T. n The process of iteratively optimizing the control parameters of the LSSVM model by minimizing the mean absolute error index as the fitness function to obtain N sets of optimal LSSVM parameters is as follows:

[0063] 1) The LSSVM control parameters that need to be optimized for each PSO model are the penalty factor C and the kernel parameter σ;

[0064] 2) The PSO models of each group predict the output of the LSSVM model at the corresponding time set T. n The mean absolute error index is minimized as the fitness function, as shown below:

[0065]

[0066] Among them, T k The time set responsible for the k-th PSO-LSSVM model, where n is the set T. k Length, y tp For the actual value, y tm This is a predicted value;

[0067] In step four above, the N sets of optimal LSSVM parameters obtained in step three are substituted into N LSSVM models and used to predict the dataset D constructed in step one. The prediction outputs of the N LSSVM models are then concatenated based on the N time sets obtained in step two to obtain the final short-term load forecast result. The process is as follows:

[0068] 1) The output of each model group is the 24-hour load data for the prediction day of the test set;

[0069] 2) According to the time set T k The elements represent specific hourly times within a 24-hour period. The load data corresponding to the k-th PSO-LSSVM model at that time is extracted into set L. k ;

[0070] 3) The final model prediction output is L = L1∪L2···L k .

[0071] Based on the above steps, select the daily 24-hour power load, daily maximum temperature, daily minimum temperature, daily average temperature and daily average humidity data for a certain region from July 7th to July 21st, 2021 (a total of 15 days), construct a multi-input multi-output sample dataset D, use the first 13 pairs of samples as the training set and the 14th pair of samples as the test set to train the model, and use it to predict the 24-hour load data for the 15th day.

[0072] Based on the above time period division method, a total of 4 time sets were obtained, and the division results are shown in the table below:

[0073]

[0074] Based on the time period segmentation results, four PSO-LSSVM models were established. Each PSO population had a size of 20 and 100 iterations. The LSSVM control parameters to be optimized were the penalty factor C and the kernel parameter σ, both with a search range of [0.01, 100]. The corresponding time set T... n The mean absolute error index was minimized as the fitness function. Simultaneously, a PSO-LSSVM prediction model was established, with the mean absolute error index of the overall prediction output minimized as the fitness function, without considering time period divisions for control. The obtained optimal parameters for the four groups and the optimal parameters for the control group are shown in the table below:

[0075]

[0076] After substituting the optimal control parameters of LSSVM, the four models and the control group predict the output based on dataset D. The predicted output obtained by concatenating the four models, the predicted output of the control group, and the actual values ​​are compared as follows: Figure 2As shown in the prediction output image, the short-term power load prediction output of the multi-group PSO-LSSVM model considering time period division is closer to the original data than the output of the single-group PSO-LSSVM prediction model. The model performance is evaluated using Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). The formulas for calculating MAE and MAPE are as follows:

[0077]

[0078]

[0079] Among them, y tp For the actual value, y tm These are the predicted values; the calculation results are shown in the table below:

[0080] method MAE(MW) MAPE (%) Embodiments of the present invention 9.4639 0.4609 control group 16.2191 0.7845

[0081] As can be seen from the table above, the short-term power load forecasting method using multiple PSO-LSSVM models that consider time period divisions outperforms the control group method that uses only one PSO-LSSVM model in terms of both mean absolute error (MAE) and mean absolute error percentage (MAPE).

[0082] Based on the above, the advantages of this invention are that the time period division and multi-group PSO-LSSVM model collaborative prediction method proposed in this invention performs better than a single PSO-LSSVM model. This is because multiple PSO-LSSVMs can better capture the characteristics between input and output data within the corresponding time set region, flexibly adjust LSSVM control parameters, improve local prediction accuracy, and thus obtain more accurate load prediction results, providing a way to optimize short-term load prediction accuracy. At the same time, it also provides decision support for regional power grid dispatch and stable operation of the power market.

[0083] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A PSO-LSSVM short-term load forecasting method considering time period division, comprising the following steps: The step one is acquiring; the step two is calculating; the step three is establishing; and the step four is substituting; characterized in that: In the step one, the power load data and the meteorological statistical data are acquired to construct a sample data set D; Wherein in the above step two, a LSSVM model is established to make regression prediction on the data set D constructed in step one, to calculate the average absolute error of LSSVM prediction output under the time set T k , and according to the variation trend of the average absolute error, the whole day of 24 hours is finally divided into N time sets T1, T2, …, T k ; a LSSVM model is established to make regression prediction on the data set D constructed in step one, to calculate the average absolute error of LSSVM prediction output under the time set T k , and according to the variation trend of the average absolute error, the whole day of 24 hours is finally divided into N time sets T1, T2, …, T k The specific process is as follows: 1) LSSVM prediction output in time set T k The calculation formula of the average absolute error index of T is as follows: Among them, T k Let T be the k-th time set, and n be the set T. k Length, y tp For the actual value, y tm This is a predicted value; 2) Periodicity initialization parameters k = 1 and time set T k = {t start , t start +1, t start +2, t end +3}, T k set elements represent certain hour points in 24 hours, the first element t start = 1 represents 1 :00:00, and the last element t end = t start +3 represents 4:00:00; 3) The LSSVM model adopts a radial basis kernel function and is trained according to the data set D constructed in the step one; 4) Calculate the mean absolute error indicator MAE0 of the LSSVM predicted outputs at the time set T according to step two 1) k ; 5) try to add t to the set T k +1 element to T end ; 6) If t end +1>24, then T k After the construction, exit the loop and output the period division results T1, T2, …, T k , otherwise, calculate MAE1 according to step two 1). 7) if MAE1 < MAE0, then T k = T k ∪ {T end +1}, T end = T end +1; 8) if MAE1 > MAE0, then T k Build complete, k = k + 1, T start = T end + 1; 9) initialize new time set T k = {t start , t start +1, t start +2, t end +3}, T start = T start +4; If any element in the new time set T k is greater than 24, exit the loop and output the time period division result T k 1, T2, …, T k n, otherwise repeat steps 24) to 29). In the step three, N groups of PSO-LSSVM models are established according to the N time sets obtained in the step two, each group of LSSVM models is trained by the data set D constructed in the step one, and the PSO model is optimized by the average absolute error index of the LSSVM model prediction output in the corresponding time set T k The average absolute error index is minimized as the fitness function, the LSSVM model control parameters are iteratively optimized, and N groups of LSSVM optimal parameters are obtained. In the step four, the N groups of optimal LSSVM parameters obtained in the step three are substituted into the N LSSVM models to predict the data set D constructed in the step one, the prediction output results of the N LSSVM models are spliced according to the N time sets obtained in the step two, and the final short-term load prediction result is obtained. 2.The PSO-LSSVM short-term load forecasting method considering time period division according to claim 1, characterized in that: In the step one, the power load data and the meteorological statistical data are acquired to construct a sample data set D, and the specific process is as follows: 1) The power load data interval is 1 hour; 2) The meteorological statistical data include daily maximum temperature, minimum temperature, average temperature and relative humidity data; 3) The 24-hour load data of the day before the day to be predicted, the meteorological statistical data and the meteorological statistical data of the day to be predicted are taken as the model input, the 24-hour load data of the day to be predicted is taken as the model output, a pair of data samples are constructed and added to the data set, and finally the sample data set D is obtained after min-max normalization processing; 4) In the sample data set D, the last pair of samples is the test set, and the remaining samples are the training set. 3.The PSO-LSSVM short-term load forecasting method considering time period division according to claim 1, characterized in that: In the third step, N groups of PSO-LSSVM models are established according to the N time sets obtained in the second step, each group of LSSVM models is trained by the data set D constructed in the first step, and the PSO model is optimized by minimizing the average absolute error index of the LSSVM model prediction output in the corresponding time set T n The process of obtaining the optimal parameters of the N groups of LSSVM is as follows: 1) The LSSVM control parameters to be optimized by each group of PSO models are the penalty factor C and the kernel parameter σ; 2) Each group PSO model minimizes the mean absolute error index of the LSSVM model prediction output at the corresponding time set T n as the fitness function as follows: Wherein, T k is the time set responsible for the kth group of PSO-LSSVM model, n is the length of set T k tp is the actual value, y tm is the predicted value.​ 4.The PSO-LSSVM short-term load forecasting method considering time period division according to claim 1, characterized in that: In the step four, the N groups of optimal LSSVM parameters obtained in the step three are substituted into the N LSSVM models to predict the data set D constructed in the step one, the prediction output results of the N LSSVM models are spliced according to the N time sets obtained in the step two, and the final short-term load prediction result is obtained. 1) The output of each group of models is the 24-hour load data of the test set on the day to be predicted; 2) According to time set T k The element represents some certain hour time in 24 hours, and the load data corresponding to the time of the kth group of PSO-LSSVM model is extracted to set L k ; 3) The final model predicts the output result as L = L1 U L2...L k .

Images