Forecasting day-ahead electricity prices: A comparison of time series and neural network models taking external regressors into account

Abstract

The amount of renewable energies in electricity production has increased significantly in the last decade, resulting in more variability of the day-ahead electricity spot price. The Electricity Price Forecast (EPF) has to adapt to the new situation by applying flexible models. However, the numerous available forecasting methods differ widely, with no distinct candidate offering the best solution. Against this background, we conduct a comparative study of four different approaches to forecasting the German day-ahead electricity spot price. In addition to the prominent Seasonal Integrated Auto-Regressive Moving Average model ((S)ARIMA(X)) and the Long-Short Term Memory (LSTM) neural network models, we employ a Convolutional Neural Network LSTM (CNN-LSTM) and an extended two-stage multivariate Vector Auto-Regressive model (VAR) approach as hybrid models. For better performance, we include common external influences such as the consumer load, fuel and CO2 emission prices, average solar radiation and wind speed in our analysis. We analyse hourly data for twelve samples from October 2017 to September 2018. Each model is implemented to deliver price forecasts at three horizons, i.e., one day, seven days and thirty days ahead. While the LSTM model achieves the best forecasting performance on average, the two-stage VAR follows closely behind and performs exceedingly well for shorter prediction horizons. Further, we provide evidence that a combination of both forecasting methods outperforms each of the single models. This indicates that combining advanced methods could lead to further improvements in electricity spot price forecasts.