A power transaction auxiliary decision-making method based on GMM and LightGBM

Abstract

The power transaction auxiliary decision-making method based on GMM and LightGBM predicts the wind power by Gaussian mixture, ridge regression algorithm and Light Gradient Boosting Machine algorithm model, improves the prediction accuracy of decision makers on power demand, market price and renewable energy supply, optimizes transaction decision and resource scheduling, and is used for power transaction auxiliary decision-making. The data set collection time is split, and the numerical features and time features are respectively constructed into numerical class new features and time scale features; the numerical class new features are clustered by the GMM algorithm according to the hour, day and month time scale, and then the clustered features are transmitted to the ridge regression algorithm to fit the preliminary power prediction results for each cluster feature to constitute the step A data set, and the time scale features are screened by the original data set to constitute the step B data set to keep the data characteristics and prevent overfitting, and the prediction value in the step A and the step B data set constitute the final LightGBM prediction data set.

Claims

1. A power transaction auxiliary decision-making method based on GMM and LightGBM, characterized in that, This method uses the GMM algorithm, Ridge Regression algorithm, and LightGBM algorithm to predict ultra-short-term wind power. First, the acquisition time in the original dataset is split into month, day, and hour, and the acquisition time item is deleted, separating numerical features from time features. The time features include month, day, and hour. Then, new features for ultra-short-term prediction are constructed, including numerical features and time-scale features. Numerical features are reconstructed by building feature interaction terms and squared terms. The cross term is formulated as: ; The formula for the square term is: ; Where data is the dataset, i is any feature in the dataset, and j is the index of any feature in the dataset other than i; The final dataset was obtained through the following steps: In the first part, the numerical new features are clustered by the GMM algorithm according to the hourly, daily and monthly time scales. Then, the three clusters of features are transmitted to the Ridge Regression algorithm. The Ridge Regression algorithm fits the preliminary power prediction results for each cluster of features, and obtains three preliminary prediction results, which constitute the first part of the dataset. The second part constructs time-scale features and merges the original dataset features and time-scale features to jointly screen for feature importance, forming a second dataset that preserves the characteristics of the original data and prevents overfitting. The predicted values ​​from the first part of the dataset are combined with the second part of the dataset to form the final dataset; The final dataset is fed into the LightGBM algorithm for ultra-short-term wind power final prediction, providing data support for power trading auxiliary decisions.

2. The power trading auxiliary decision-making method based on GMM and LightGBM according to claim 1, characterized in that, Before splitting the acquired dataset, iterate through the dataset and fill in outliers and missing values ​​by averaging adjacent units.

3. The power trading auxiliary decision-making method based on GMM and LightGBM according to claim 1, characterized in that, The GMM algorithm performs feature clustering according to hourly, dayly, and monthly time scales, using hour, day, and month as representations. The probability distribution model satisfied by the Gaussian mixture model is as follows: ； in For the mixed probability density, It is a d-dimensional Gaussian probability density. These are the weights, mean, and covariance matrix of the k-th Gaussian model, respectively. The GMM algorithm fits k mixtures of Gaussian distributions by solving for the maximum expectation.

4. The power trading auxiliary decision-making method based on GMM and LightGBM according to claim 3, characterized in that, The GMM algorithm process includes: Step 1: Initialize and set the number of categories K; Step 2: Calculate the posterior probability ,in ； Step 3: Update parameters ； Where n is the sample number, n=1,2,…,N, and N is the total number of samples; Step 4: Calculate the log-likelihood function: ; Step 5: Determine whether the Gaussian parameters or likelihood function have converged. If they have converged, end the iteration and assign the sample to the class with the highest likelihood function. If they have not converged, return to Step 2 to continue the iteration. After clustering, the GMM algorithm outputs three clusters of feature values ​​that are closely related to hour, day, and month, respectively.

5. The power trading auxiliary decision-making method based on GMM and LightGBM according to claim 4, characterized in that, The ridge regression algorithm is used to fit preliminary power prediction results for each of the three clustered feature values. The regression coefficients of the multiple linear regression equation constructed by the ridge regression algorithm satisfy the following equation: ； in For regression variables 1-th order matrix for The transpose of the matrix, For the dependent variable 1st order vector For regression coefficients Rank vector; Ridge regression algorithm satisfies the following equation: ； in Let λ be the p-order identity matrix and λ be the ridge parameter. A non-negative factor on the main diagonal element.

6. The power trading auxiliary decision-making method based on GMM and LightGBM according to claim 1, characterized in that, The second part of the dataset construction method involves statistically analyzing the actual power characteristics by month, day, and hour. Specifically, the maximum, minimum, median, standard deviation, and average values ​​of the actual power are calculated for each month, day, and hour. These values ​​are then added to the original feature set as new feature values. The new statistical features and the features in the original dataset are used together to calculate the feature importance with the prediction target, i.e., the actual power. The top 30 features with the highest importance are selected as a new feature set using the forward selection method. The second part of the dataset, consisting of the top 30 features, and the first part of the dataset obtained through the GMM algorithm and the ridge regression algorithm are used to construct the final dataset.

7. The power trading auxiliary decision-making method based on GMM and LightGBM according to claim 1, characterized in that, LightGBM algorithm formula: ； Where F(x) is the output ultra-short-term wind power prediction value; Let t be the output value of the t-th weak regression tree, and T be the total number of weak regression trees; The input to the LightGBM algorithm is the final dataset, which consists of two parts: the first part is the first dataset, which is the three preliminary power prediction results obtained by fitting the features of each cluster using the Ridge Regression algorithm after clustering by the GMM algorithm according to hourly, daily, and monthly time scales; the second part is the second dataset, which is the top 30 features of importance obtained by combining the original dataset with the actual power features statistically analyzed by hourly, daily, and monthly time scales and filtering them by feature importance. In this algorithm, five-fold cross-validation is used on the training set of LightGBM, and Bayesian hyperparameter tuning algorithm is used to optimize the hyperparameters in LightGBM.

Images