Stacked ensemble forecasting of electricity price, demand, and solar generation to support battery management

The increasing adoption of decentralized photovoltaic systems and storage batteries has significantly changed the paradigm of participation in the Iberian electricity market (MIBEL). In this context, this paper proposes a predictive-analytical approach to support hourly energy management decisions by producers with batteries, namely in the choice between self-consumption, storage or sale of surplus energy. To this end, three models are developed and validated for: (i) hourly forecast of the electricity price in the MIBEL market; (ii) forecast of the electricity consumption of a real institutional customer; and (iii) forecast of photovoltaic production. The models combine machine learning and deep learning techniques, namely XGBoost and LSTM networks, integrated through an ensemble stacking with a linear regression meta-model. Models are evaluated under a strict leakage-free protocol with an explicit temporal train/calibration/test split, expanding-window cross-validation, a seasonal-persistence baseline, and conformal prediction intervals. The price forecast is accurate and robust (test R2=0.94, above the baseline R2=0.39); photovoltaic production is moderately predictable (R2≈0.69 over daylight hours), while consumption for this irregular institutional load is harder (R2≈0.40). The prescriptive optimization module that would consume these forecasts is left as future work; accordingly, this paper delivers and rigorously validates the forecasting stage intended to support such optimization.