An Expectile Factor Model for Day-ahead Wind Power Forecasting

Abstract

Volatile renewable power sources will soon dominate the power portfolio of leading industrial countries to match the goals of the Paris climate accord. Wind power is the leading renewable energy source. Its production is strongly volatile due to weather dependency and highly decentralised to various locations across a country. Hence, transmission and distribution system operators (TSO, DSO) and energy market participants are exposed to volume risk. Good forecasts of future wind power production are required to match residual load supply and demand in the network at any point time to circumvent load failure or frequency jumps. We propose a flexible methodology that can be scaled from a small wind farm data set to a large fleet size to forecast aggregated, turbine or farm specific wind power production. The methodology is based on multivariate expectile regressions using a factor model in combination with a vector autoregressive (VAR) time series framework with exogenous variables for forecasting purposes. Moreover, the methodology can be extended to deploy spatial information of various wind farms. We perform our analysis on a farm-level and an aggregated TSO-level data set. Results indicate out-performance of our candidate models for one day ahead forecasts. Moreover, we show that applying our candidate model affordable daily weather forecasts give better results than employing expensive intra-day weather forecasts. The model performs similarly well for farm-level data as well as for aggregated data.

Keywords: forecast, renewable energy, wind power, factor model, penalisation, functional data analysis, expectile, multivariate regression, short-term, Markov switching, cluster