Determination method of typical daily load curve in planning target year based on multi-algorithm fusion

By combining autoregressive integral sliding average model and random forest algorithm in load prediction of power system, combined with Bayesian optimization and improved K-mean clustering algorithm, the load prediction complexity and data noise problems in the existing technology are solved, and the problem of load prediction is realized. High precision and high reliability load prediction.

CN118673466BActive Publication Date: 2024-11-22ZHEJIANG UNIV +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411156717.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-22
Estimated Expiration
2044-08-22

AI Technical Summary

Technical Problem

The existing power system load prediction method faces the problems of large data volume, high computational complexity, and overfitting of the model when dealing with long-term predictions, and it is difficult to effectively deal with missing values ​​and outliers in the load data. It is difficult for a single algorithm to meet at the same time. Accuracy and stability requirements for load curve prediction.

Method used

Using a multi-algorithm fusion method, the autoregressive integral sliding average model and the random forest algorithm are combined. Through the improved K-mean clustering algorithm, data cleaning and model prediction are optimized, and the Bayesian optimization is combined to improve the model parameter selection efficiency to achieve high Precision load prediction.

Benefits of technology

The data is cleaned by the local anomaly factor algorithm, combined with the autoregressive integral sliding average model and the random forest algorithm to improve the prediction accuracy, the improved K-mean clustering algorithm more accurately reflects the seasonal changes and trends of the load curve, significantly improving the load prediction of the power system accuracy and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118673466B_ABST
    Figure CN118673466B_ABST
Patent Text Reader

Abstract

The present invention provides a method for determining a typical daily load curve for a planning target year based on multi-algorithm fusion. The method first uses a local anomaly factor algorithm to clean historical data and detect outliers to identify and replace abnormal data. Subsequently, an autoregressive integral moving average model is used to predict the future daily maximum load, and a random forest algorithm is used to predict the daily normalized load curve. The selection efficiency and prediction accuracy of the algorithm model parameters are improved by Bayesian optimization. The daily load curve for the planning target year is obtained by using the maximum load value predicted by the regression integral moving average model and the normalized curve predicted by the random forest. Finally, the predicted daily load curve for the target year is clustered and analyzed on a monthly basis using an improved K-means clustering algorithm, and the mean of the largest cluster is selected as the typical daily load curve for each month. Experimental results show that this method can accurately capture the seasonal changes and trends of the load curve and significantly improve the prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of power system load forecasting, and in particular to a method for determining a typical daily load curve of a planning target year based on multi-algorithm fusion. Background Art

[0002] With the transformation of the global energy structure and the rapid development of renewable energy, load forecasting of power systems has become particularly important. Traditional load forecasting methods mainly rely on simple trend extrapolation of historical data, which is difficult to cope with the high volatility and complexity of loads in modern power systems. In recent years, the application of machine learning and deep learning methods in load forecasting has gradually increased, such as short-term load forecasting and medium- and long-term load forecasting. However, these methods face problems such as large data volume, high computational complexity, and easy overfitting of models when dealing with long-term forecasts. In addition, missing values ​​and outliers in load data can also significantly affect the prediction accuracy. Existing prediction methods, such as the autoregressive integrated moving average model, are suitable for processing trends and seasonal changes in time series data, but have poor effects on nonlinear relationships in load data. The random forest algorithm has advantages in processing nonlinear and high-dimensional data, but has shortcomings in capturing long-term trends and periodic changes in time series. Therefore, it is difficult for a single algorithm to simultaneously meet the accuracy and stability requirements of load curve prediction. Summary of the invention

[0003] In order to overcome the above challenges, this paper proposes a method based on multi-algorithm fusion, combining the autoregressive integrated moving average model and the random forest algorithm, and using their respective advantages to improve the accuracy and reliability of load forecasting. Through the improved K-means clustering algorithm, the method for determining the typical daily load curve is optimized to make it more suitable for the needs of modern power system planning.

[0004] The present invention proposes a method for determining a typical daily load curve in a power system planning target year based on multi-algorithm fusion, the method comprising:

[0005] S1: Obtain historical power load data and perform data cleaning;

[0006] S2: Predict the daily maximum load of the target year based on the cleaned load data through the autoregressive integrated moving average model;

[0007] S3: The cleaned load data is normalized; a random forest model is used to predict the daily normalized load curve of the target year based on the normalized load data;

[0008] S4: combining the predicted daily maximum load of the target year with the daily normalized load curve of the target year to obtain the daily load forecast curve of the target year;

[0009] S5: Using the improved K-means clustering method, the daily load forecast curve of the target year is clustered and analyzed on a monthly basis, and the mean of the largest cluster is selected as the monthly typical daily load curve of the target year.

[0010] Furthermore, in step S1, the data cleaning includes missing value processing and outlier processing:

[0011] The missing value processing includes: filling the missing value in the historical power load data with the mean of two or more data points before and after the missing value;

[0012] The outlier processing includes: using a local outlier factor algorithm to identify an abnormal data point in the historical power load data, and replacing it with the average of two data points before and after the abnormal data point.

[0013] Furthermore, the normalization specifically includes: the cleaned load data is divided into multiple groups of daily load data, the daily load data includes load data at various time points of the day, for each group of daily load data, the maximum load of the day is calculated, and the load data at various time points of the day are divided by the maximum load of the day as the normalized load data.

[0014] Further, in step S3, the prediction of the daily normalized load curve of the target year based on the normalized load data specifically includes: the random forest model includes multiple decision trees, the decision trees predict the load data of the target year based on the normalized load data, the final prediction value of the load data of the target year is the average of the prediction values ​​of all decision trees, and the final prediction value of the load data at each time point every day of the target year constitutes the daily normalized load curve of the target year.

[0015] Furthermore, the step S4 comprises: multiplying the predicted daily maximum load of the target year by the daily normalized load curve of the target year, so as to obtain the daily load prediction curve of the target year.

[0016] Furthermore, the step S5 specifically includes: for the daily load forecast curves of each month in the target year, K daily load curves with the smallest sum of Kendall correlation coefficients are selected through a local search algorithm, and the selected K daily load curves are clustered by a K-means algorithm with the selected K daily load curves as clustering centers; the silhouette coefficients of the clusters under different K values ​​are calculated, and the K value corresponding to the maximum silhouette coefficient is the final number of clusters. For the clustering results corresponding to the final number of clusters, the average value of the largest cluster is the monthly typical daily load curve of the target year.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] The local outlier factor (LOF) algorithm is used to clean historical data and detect outliers, which can effectively identify and replace abnormal data, reduce data noise, and improve the accuracy of load forecasting. The autoregressive integral moving average model is combined with the random forest algorithm, and the model parameter selection efficiency and prediction accuracy are improved through the Bayesian optimization method, so as to achieve high-precision prediction of the future daily maximum load and normalized load curve. The present invention comprehensively utilizes the autoregressive integral moving average model to process the trend and seasonal changes of time series, and combines the random forest algorithm to process the nonlinear and high-dimensional characteristics of data, and performs well in long-term and short-term load forecasting. The predicted target year daily load curve is clustered and analyzed by the improved K-means clustering algorithm, and the mean of the largest cluster is selected as the typical daily load curve for each month, which more accurately reflects the seasonal changes and trends of the load curve. The method is suitable for modern power systems with high volatility and complex load characteristics, can cope with variable load data, and provide more reliable data support for power system planning. In summary, the present invention has made significant progress in data cleaning, model optimization, multi-algorithm fusion and load curve determination, and has significantly improved the accuracy and reliability of power system load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a framework diagram of a method for determining a typical daily load curve for a target year in power system planning by integrating multiple algorithms of the present invention.

[0020] Figure 2 This is a diagram showing the difference between outliers and replaced data in data cleaning according to an embodiment of the present invention.

[0021] Figure 3 This is a graph showing the prediction results of the maximum daily load test set predicted by the autoregressive integrated moving average algorithm in one embodiment of the present invention.

[0022] Figure 4 This is a graph showing the prediction results of the random forest algorithm prediction curve morphology test set in one embodiment of the present invention.

[0023] Figure 5 This is a graph of prediction results of a single algorithm test set in one embodiment of the present invention.

[0024] Figure 6 This is a graph of prediction results of the test set of the autoregressive integrated moving average and random forest hybrid prediction method in one embodiment of the present invention.

[0025] Figure 7 This is a typical daily load curve diagram for each month in 2028 predicted in one embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described below in conjunction with the accompanying drawings and examples.

[0027] This example is based on the actual 24-hour daily load data of a certain regional power grid from 2019 to 2023. The power load in this region is mainly industrial load, accounting for 80%. The average annual load growth rate has reached 10% in the past five years. The daily load curve has the characteristics of typical noon and nighttime double peaks, and the peak-to-valley difference within the day is about 8%. The typical daily load curve determination method for the target year of power system planning using multi-algorithm fusion is used to predict the typical daily load curve for each month in 2028, with a data collection time interval of 1 hour.

[0028] like Figure 1 As shown in the figure, the method for determining the typical daily load curve of the target year of power system planning based on multi-algorithm fusion mainly includes the following steps:

[0029] Step 1: Data collection and cleaning

[0030] Collect historical power load data and perform data cleaning. In this embodiment, the historical power load data is the actual daily 24-hour load data of a regional power grid from 2019 to 2023. In the process of power load data collection, data missing or abnormal conditions may occur. In order to improve data quality, data cleaning is required. Data cleaning includes missing value processing and outlier processing.

[0031] (1) Missing value processing

[0032] First, we deal with missing values ​​in the data by using the mean filling method. Assume that the data set is ,in is the number of data points in the historical power load data. If the i-th data point in the historical power load data is a missing value, then for the missing value , using the mean of two or more data points before and after the data point In this example, the mean of the two data points before and after is used to fill in the missing values. The calculation formula is as follows:

[0033]

[0034] in is the jth data point in the historical power load data. The missing value replacement formula is as follows:

[0035]

[0036] (2) Outlier processing

[0037] Use the Local Outlier Factor (LOF) algorithm to identify and replace abnormal data. The LOF algorithm calculates the local outlier factor for each data point, including the following sub-steps:

[0038] Calculate the Euclidean distance between data points. The i-th data point in the historical power load data and the jth data point The Euclidean distance is defined as:

[0039]

[0040] in Represents data points and data points The Euclidean distance of For data points The kth eigenvalue of For data points is the kth eigenvalue of , where m is the number of eigenvalues.

[0041] Calculate for each data point Distance. For data points , The distance is its The Euclidean distance of the nearest neighbors is denoted by .

[0042] Calculate the reachable distance of each data point. and its neighboring data points , define the reachable distance as:

[0043]

[0044] Calculate the local reachability density. For data points , whose local reachable density Defined as:

[0045]

[0046] in For data points The set of k nearest data points of For collection The number of elements in .

[0047] Calculate the local outlier factor (LOF). Its local outlier factor LOF is defined as:

[0048]

[0049] When the LOF value of a data point is greater than the threshold, the data point is considered an outlier. The replacement formula is as follows:

[0050]

[0051] in and For two data points adjacent to the outlier, the threshold of the LOF value is preferably 1-1.5. In this embodiment, the threshold is 1.5. Figure 2 The interpolation of the data with outliers and replacement is shown.

[0052] Step 2: Hybrid algorithm predicts the daily load curve of the target year

[0053] (1) Maximum daily load forecast

[0054] For the load data collected and cleaned in step 1, find the historical daily maximum load data, input the historical daily maximum load data into the autoregressive integral moving average model, and the model predicts the daily maximum power load data of the target year. The autoregressive integral moving average model formula is as follows:

[0055]

[0056] in, is the number of autoregressive terms, is the number of moving average terms, is the power load data at time t, is the power load data at time ti, is a constant term, is the autoregression coefficient, is the moving average coefficient, , is the error term between time t and time tj.

[0057] Bayesian optimization is used to determine the parameters of the autoregressive integrated moving average model. The objective function in the optimization process is to minimize the root mean square error (RMSE) between the predicted value and the actual value. The formula for the root mean square error is as follows:

[0058]

[0059] Among them is Actual value, is the predicted value, is the number of data points in the historical power load data.

[0060] Figure 3 The forecast results of the test set for the maximum daily load predicted by the ARIMA model are shown.

[0061] (2) Normalized load curve prediction

[0062] This step uses the random forest algorithm to predict the daily normalized load curve, that is, the load curve shape. First, the historical power load data after cleaning in step (1) is normalized. Suppose the cleaned historical power load data is ,in It represents the daily load data after cleaning, and the maximum daily load is , the daily normalized load curve is obtained by normalization , the normalization formula is as follows:

[0063]

[0064] Construct N decision trees and extract multiple sample subsets from the original data set through bootstrap sampling.

[0065] During the construction of each decision tree, a subset of features is randomly selected for node splitting;

[0066] The normalized load curve is predicted by multiple decision trees, and the average of all decision tree prediction values ​​is taken as the final prediction result.

[0067] Bayesian optimization is also used to determine the optimal parameters of the random forest model, and the root mean square error (RMSE) is used as the objective function in the optimization process.

[0068] Figure 4 The prediction results of the test set for the random forest algorithm predicting the shape of the normalized load curve are shown.

[0069] (3) Daily load curve prediction

[0070] The daily load forecast curve for the target year is obtained by multiplying the daily maximum load predicted by the autoregressive integrated moving average model with the normalized load curve predicted by the random forest algorithm:

[0071]

[0072] in It represents the final load forecast at time t in the target year. The final load forecast data on the same day of the target year constitute the daily load forecast curve of the target year. represents the target year daily maximum load corresponding to time t, Represents the daily normalized load curve of the target year corresponding to time t. Figure 6 The forecast results of the test set for the autoregressive integrated moving average model and the random forest hybrid forecasting method are shown.

[0073] Step 3: Selection of typical daily load curve

[0074] In this step, the K-means clustering algorithm is used to perform cluster analysis on the daily load curve of the target year.

[0075] Specifically, the number of clusters K is set, the Kendall correlation coefficient between the daily load curves is calculated, and the K curves with the smallest correlation corresponding to different K values ​​are found through the local search algorithm as the cluster center, and finally the mean of the largest cluster cluster is used as the typical daily load curve for each month. In this embodiment, the K value is set to 2-5, and the silhouette coefficient of each K value is calculated respectively, and the K value corresponding to the largest silhouette coefficient is used to determine the number of clusters. The specific steps are as follows:

[0076] (1) Randomly select a starting load curve from the daily load forecast curves of each month in the target year and add it to the selected curve set. middle.

[0077] (2) For the remaining load forecast curves, calculate their differences from the selected curve set The minimum correlation coefficient of all curves in is:

[0078]

[0079] in is the minimum correlation coefficient between curve j and all curves in the selected curve set, It means to find the Kendall correlation coefficient between curve j and curve k.

[0080] (3) Randomly select a curve j from the unselected curves and add it to the set of selected curves and updates all remaining curves and the selected curve set The minimum correlation coefficient of all curves in :

[0081]

[0082] Each time a new curve is selected, the quality of the current solution is evaluated and compared with the previous solution. If the current solution is better, it is accepted; otherwise, other solutions in the neighborhood are searched. By continuously updating the solution, it is possible to avoid falling into the local optimum.

[0083] (4) Repeat step (3) until the number of curves in the selected curve set S reaches the value K and the predetermined stop condition is met (such as reaching the maximum number of iterations or the improvement of the solution is less than the threshold). At this time, the sum of the correlation coefficients of all curves in the selected curve set S is the sum of the correlations between the selected curves.

[0084] (5) Repeat the above operation several times and select the set of K daily load curves with the smallest sum of correlations as the clustering center.

[0085] (6) Finally, the selected K daily load curves are used as clustering centers and the K-means algorithm is used to complete clustering.

[0086] (7) Repeat steps (1) to (6) for different K values, calculate the silhouette coefficient for each K value, use the silhouette coefficient to judge the clustering effect of different K values, determine the number of clusters based on the K value corresponding to the maximum silhouette coefficient, and use the average value of the largest cluster after clustering as the monthly typical daily load curve; the silhouette coefficient calculation formula is as follows:

[0087]

[0088] in, Indicates the intra-cluster similarity, i.e., the daily load curve sample The average distance to other samples in the cluster, Represents the inter-cluster dissimilarity, i.e., the sample The average distance to all samples in the nearest other clusters. For the silhouette coefficient of the entire daily load curve data set, the average value of the silhouette coefficients of all samples can be taken. The larger the average value, the better the clustering effect. The corresponding K value when the average value of the silhouette coefficient is the largest is used to determine the number of clusters.

[0089] Using the above method, a case study is set up to analyze the actual daily load data of a regional power grid from 2019 to 2023, and predict the typical daily load curve in 2028. The data cleaning process is as follows Figure 2 The maximum load prediction results are shown in Figure 3 As shown in Figure 2, the normalized load curve prediction results are as follows: Figure 4 As shown, the prediction results of a single algorithm are as follows Figure 5 As shown in Figure 2, the prediction results of the hybrid algorithm are as follows: Figure 6 As shown in the figure, the forecast results of typical daily load curves for each month in 2028 are as follows Figure 7 This example includes:

[0090] 1) Data cleaning

[0091] The missing values ​​are filled using the mean method, and the outliers are detected and replaced using the LOF algorithm. The difference between the detected outlier data and the replacement data is as follows: Figure 2 shown.

[0092] 2) Maximum load prediction

[0093] The daily maximum load data from 2019 to 2023 were extracted, and the autoregressive integrated moving average model was used for prediction. The model parameters were determined by Bayesian optimization. The prediction results are as follows: Figure 3 shown.

[0094] 3) Daily load curve shape prediction

[0095] The historical daily load curve is normalized, and the random forest algorithm is used for prediction. The model parameters are determined by Bayesian optimization. The prediction results are as follows: Figure 4 shown.

[0096] 4) Daily load curve prediction

[0097] The daily maximum load predicted by the autoregressive integrated moving average model is multiplied by the normalized load curve predicted by the random forest to obtain the complete load curve for each day of the target year. The prediction results are as follows: Figure 6 shown.

[0098] 5) Determination of typical daily load curve

[0099] The improved K-means clustering algorithm is used to perform cluster analysis on the daily load curve in 2028, and the mean of the largest cluster is selected as the typical daily load curve for each month. The results are as follows: Figure 7 As shown. The obtained monthly typical daily load curve is consistent with the daily load characteristics of the region. The annual load peak occurs in the fourth quarter, which is consistent with the actual situation, and the overall load increase level is consistent with the trend of historical data. Compared with the traditional power system planning method of only predicting the maximum load or selecting the typical daily load curve based on experience, the method for determining the typical daily load curve of the planning target year based on multi-algorithm fusion proposed in the present invention further improves the accuracy and scientificity of the prediction, and can effectively improve the planning effect of the power system.

[0100] The above description of the embodiments is to facilitate those skilled in the art to understand and apply the present invention. The present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A method for determining typical daily load curves in a planning target year based on multi-algorithm fusion, characterized in that: The steps include: S1: Obtain historical power load data and perform data cleaning; S2: Predict the daily maximum load of the target year based on the cleaned load data through the autoregressive integrated moving average model; S3: normalization of the cleaned load data; The random forest model is used to predict the daily normalized load curve of the target year based on the normalized load data; S4: combining the predicted daily maximum load of the target year with the daily normalized load curve of the target year to obtain the daily load forecast curve of the target year; S5: For the daily load forecast curves of each month in the target year, K daily load curves with the smallest sum of Kendall correlation coefficients are selected through a local search algorithm, and clustering is performed through the K-means algorithm with the selected K daily load curves as clustering centers; the silhouette coefficients of clusters under different K values ​​are calculated, and the K value corresponding to the maximum silhouette coefficient is the final number of clusters. For the clustering results corresponding to the final number of clusters, the average value of the largest cluster is the typical monthly daily load curve of the target year.

2. The method for determining the typical daily load curve of the planning target year based on multi-algorithm fusion according to claim 1 is characterized in that: In step S1, the data cleaning includes missing value processing and outlier processing: The missing value processing includes: filling the missing value in the historical power load data with the mean of two or more data points before and after the missing value; The outlier processing includes: using a local outlier factor algorithm to identify an abnormal data point in the historical power load data, and replacing it with the average of two data points before and after the abnormal data point.

3. The method for determining the typical daily load curve of the planning target year based on multi-algorithm fusion according to claim 2 is characterized in that: The method of using the local outlier factor algorithm to identify abnormal data points in the historical power load data specifically includes: calculating the local outlier factor of each data point in the historical power load data, and if the local outlier factor is greater than a preset threshold, the data point is considered to be an abnormal data point.

4. The method for determining the typical daily load curve of a planning target year based on multi-algorithm fusion according to claim 1 is characterized in that: The method of predicting the daily maximum load of the target year based on the cleaned load data specifically includes: for the cleaned load data, searching for the historical daily maximum load data, inputting the daily maximum load data into a pre-trained autoregressive integrated moving average model, and using the model to predict the daily maximum power load data of the target year.

5. The method for determining the typical daily load curve of the planning target year based on multi-algorithm fusion according to claim 4 is characterized in that: The autoregressive integrated moving average model determines the optimal parameters through Bayesian optimization, and the optimization objective function is the root mean square error between the predicted value and the actual value.

6. The method for determining the typical daily load curve of a planning target year based on multi-algorithm fusion according to claim 1 is characterized in that: In step S3, the normalization specifically includes: The cleaned load data is divided into multiple groups of daily load data, and the daily load data includes load data at various time points of the day. For each group of daily load data, the maximum load of the day is calculated, and the load data at various time points of the day are divided by the maximum load of the day to serve as normalized load data.

7. The method for determining typical daily load curve of a planning target year based on multi-algorithm fusion according to claim 1 is characterized in that: In step S3, the daily normalized load curve of the target year is predicted based on the normalized load data, specifically including: the random forest model includes multiple decision trees, the decision trees predict the load data of the target year based on the normalized load data, the final prediction value of the load data of the target year is the average of the prediction values ​​of all decision trees, and the final prediction value of the load data at each time point every day of the target year constitutes the daily normalized load curve of the target year.

8. The method for determining the typical daily load curve of the planning target year based on multi-algorithm fusion according to claim 7 is characterized in that: The random forest model determines the optimal parameters through Bayesian optimization, and the optimization objective function is the root mean square error between the predicted value and the actual value.

9. The method for determining typical daily load curve of a planning target year based on multi-algorithm fusion according to claim 1 is characterized in that: The step S4 comprises: multiplying the predicted daily maximum load of the target year by the daily normalized load curve of the target year, so as to obtain the daily load prediction curve of the target year.