Conditional Density Forecast of Electricity Price Based on Ensemble ELM and Logistic EMOS

In recent years, probabilistic forecast of electricity price has become of particular interests to market participants as it can effectively model the uncertainties due to competitive market behaviors. Decision makers heavily rely on such forecast to formulate optimal strategies with minimal risk and maximum profits to deal with stochasticity in market and system operation. Different from the widely used volatility models with least square or maximum likelihood techniques in probabilistic forecast of prices, this paper proposes a reliable continuous ranked probability score-oriented predictive density construction strategy for day-ahead electricity prices. The proposed method effectively quantifies the uncertainty involved in extreme learning machine network by applying an ensemble structure and a logistic distribution-based ensemble model output statistics technique. Moreover, an efficient covariance structure directly determined by the empirical correlations of observed probabilistic forecast series is developed to capture the essential temporal interdependence thus to facilitate the operational scenarios' generation. Through validating on the real day-ahead market in Sweden, the proposed hybrid approach proves to provide superior full distributional forecasting skill over the existing approaches.