Hybrid Forecasting for Sustainable Electricity Demand in The Netherlands Using SARIMAX, SARIMAX-LSTM, and Sequence-to-Sequence Deep Learning Models

Accurate forecasting is essential for effective energy management, particularly in evolving and data-driven electricity markets. To address the increasing complexity of national energy planning in The Netherlands, this study proposes a hybrid multi-stage forecasting framework to improve both short- and long-term electricity demand predictions. We compare three model types, classical statistical (SARIMAX), hybrid statistical–deep learning (SARIMAX–LSTM), and deep learning (sequence-to-sequence), across forecasting horizons from 1 to 180 days. The models are trained on daily load data from ENTSO-E (2009–2023), incorporating exogenous variables such as weather conditions, energy prices, and socioeconomic indicators, as well as engineered temporal features such as calendar effects, seasonal patterns, and rolling demand statistics. Three feature configurations were tested: exogenous-only, generated-only, and a combined set. Internally generated features consistently outperformed exogenous inputs, especially for long-term forecasts. The sequence-to-sequence model achieved the highest accuracy at the 180-day horizon, with a mean absolute percentage error (MAPE) of approximately 1.88%, outperforming both SARIMAX and the SARIMAX–LSTM hybrid models. An additional SARIMAX-based analysis assessed the individual effects of renewable and socioeconomic indicators. Renewable energy production improved short-term accuracy (MAPE reduced from 2.13% to 1.09%) but contributed little to long-term forecasting. Socioeconomic variables had limited predictive value and, in some cases, slightly reduced accuracy, particularly over long-term horizons.