Non-Parametric Probabilistic Demand Forecasting in Distribution Grids; Kernel Density Estimation and Mixture Density Networks

Abstract

This paper presents the application of nonparametric approaches for probabilistic demand forecasting of residential customers in power distribution grids. Two short-term forecasting models based on non-parametric kernel density estimation (KDE) and mixture density networks (MDN) are developed that estimate the conditional probability of the customers aggregate demand at any given temperature and time. From conditional probability distributions and by given forecast temperatures, the demand probability distribution is obtained., which in turn can derive expected or quantile demand forecast. In addition., the impact of seasonality associated with the data on demand forecasting are taken into account. Criteria including Relative Root Mean Square Error (RRMSE) and Mean Absolute Percentage Error (MAPE) as well as Q-Q plot technique and D-M statistic test are utilized to assess the models performance., and to compare them with each other and with benchmark models.