Modular electrical demand forecasting framework — A novel hybrid model approach

Abstract

In the face of a changing European power market, accurate electric load forecasts are of significant importance for power traders, power utility and grid operators to reduce costs for ancillary services. The following case study, based on publicly available load data, focuses on a novel approach to combine different forecasting methodologies and techniques from the area of computational intelligence. The proposed hybrid model blends input forecasts from artificial neuronal networks, multi variable linear regression and support vector regression machine models with fuzzy sets to intraday and day ahead forecasts. The forecasts are evaluated with commonly used metrics (mean average percentage error - MAPE & normalized rooted mean square error - NRMSE) to allow a comparison to other case studies. The results from the input forecasting models range from a yearly MAPE of 3.1% for the artificial neuronal network to 2.51% for the support vector machine. The blended forecast from the proposed hybrid model results in a MAPE of 1.2% for one hour and a MAPE of 2.03% for 24 hours ahead forecasts.