Answers are grounded in this paper and cite their sources.
Accurate forecasting and scheduling are essential for sustainable management of decentralized electricity systems. We introduce a framework for evaluating statistical, machine learning (ML), and deep learning (DL) architectures for particularized electricity demand forecasting in rural Energy Communities (ECs). The models are evaluated under a four-fold expanding-window protocol with seasonal test windows spanning early to mature operation, and selected by a Wilcoxon signed-rank tournament across 24 candidates. ML models statistically outperform all DL variants at all six supply points as conventionally deployed. Under a feature-matched ablation the advantage holds at five of six supply points, the sixth favoring a univariate network on the four-fold mean. Removing calendar and weather inputs improves DL accuracy in 15 of 18 cases, so much of the DL disadvantage reflects input configuration rather than architecture. RMSE-based and statistical selection agree in only 8% of cases, yet yield comparable operational cost, so neither is unreliable. Daily DL retraining reduces Mean Absolute Error by 5%–7%, insufficient to close the gap with the best ML models. Day-ahead forecasts then feed a linear programming optimizer that schedules battery storage, grid interaction, and photovoltaic self-consumption under real-world constraints, running on a Raspberry Pi 5 and an NVIDIA Jetson Orin Nano in under 211 ms. Over a 123-day, four-season evaluation on an operational rural EC in northern Spain, the installed PV, batteries, and charge controller convert electricity expenditure into net revenue. Using only day-ahead forecasts, the optimizer saves a further 7% of the grid-only cost relative to the default charge controller.
No takes yet. Share an insight, caveat, or question.
Cosío et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: