Title: N-BEATS neural network for mid-term electricity load forecasting

Abstract: We address the mid-term electricity load forecasting (MTLF) problem. This problem is relevant and challenging. On the one hand, MTLF supports high-level (e.g. country level) decision-making at distant planning horizons (e.g. month, quarter, year). Therefore, financial impact of associated decisions may be significant and it is desirable that they be made based on accurate forecasts. On the other hand, the country level monthly time-series typically associated with MTLF are very complex and stochastic -- including trends, seasonality and significant random fluctuations. In this paper we show that our proposed deep neural network modelling approach based on the N-BEATS neural architecture is very effective at solving MTLF problem. N-BEATS has high expressive power to solve non-linear stochastic forecasting problems. At the same time, it is simple to implement and train, it does not require signal preprocessing. We compare our approach against the set of ten baseline methods, including classical statistical methods, machine learning and hybrid approaches on 35 monthly electricity demand time series for European countries. We show that in terms of the MAPE error metric our method provides statistically significant relative gain of 25% with respect to the classical statistical methods, 28% with respect to classical machine learning methods and 14% with respect to the advanced state-of-the-art hybrid methods combining machine learning and statistical approaches.