Gradient Boosting Machine과 Laplace Distribution 기반의 확률적 전력수요 예측 알고리즘 연구

Abstract

In this paper, we develop three probabilistic electric load forecasting approaches: two parametric approaches and one non-parametric approach. In the parametric approach, we design the probability of load forecasts as the Laplace distribution since the empirical distribution of load forecasts has a shape of Laplace distribution. We also design the probability of load forecasts as the Gaussian distribution, since it has been widely used in other studies. We compare the forecasting accuracy of two distributions. The means of distributions are estimated by using the gradient boosting machine (GBM), and the standard deviations of distributions are estimated by analyzing forecasting errors through the cross validation. In the non-parametric approach, we find the probability of load forecasts by using the quantile regression (QR). Finally, we compare the forecasting accuracy of parametric and non-parametric approaches by measuring the accuracy on the pinball loss function. A parametric approach based on the Laplace distribution and GBM is the most accurate approach.