Aminhossein Jahanbin
Accurate multi-horizon state-of-charge (SOC) prediction for hybrid hydrogen (H 2 ) and battery energy storage systems (BESS), ranging from hourly to weekly or monthly horizons, is essential for effective net-zero energy management in PV - driven buildings, particularly under cold climate conditions. Short - term forecasts ( e.g. , 6 h or day-ahead) facilitate optimal intra - day energy shifting, while medium- and long-term predictions ( e.g., seasonal) enable informed scheduling and strategic storage dispatch between BESS and H 2 . In this context, this study proposes a multi-horizon machine learning (ML) architecture for forecasting SOCs and orchestrating storage shifts in bi-level H 2 –BESS systems, supporting net-zero energy management in PV - driven buildings. The framework synergistically integrates varied timescale surrogate modeling — from short- (granular) to long - term (aggregated) temporal resolutions—with desirability-based multi-criteria optimization (MCO) outputs and dynamic simulation-driven learning. To effectively capture nonlinear dependencies and temporal correlations, four joint-output ML regressors—LGB, RF, LSTM, and SVR—are trained on high-resolution datasets comprising temporal descriptors, PV generation, load profiles, PV and BESS surplus, and H 2 –BESS operational variables. The MCO outputs confirm the reliability of the established model in identifying optimal component configurations that achieve an energy autonomy ratio (EAR) of 1.0 while minimizing life - cycle cost (LCC). High-fidelity datasets for ML training are obtained from several-year dynamic simulations in five case-study cities, resulting in 131,406 hourly samples. Across all temporal horizons, tree-based ensemble models consistently outperform LSTM and SVR, with RF delivering the highest accuracy for SOC H2 (hourly R 2 = 0.9956, RMSE = 0.020) and exhibiting progressively tighter U 95 uncertainty bands at coarser aggregations, whereas SVR shows the weakest performance at fine resolutions. SHAP and correlation analyses reveal that electrolyzer operation, load, and seasonal effects predominantly shape SOC trajectories. Beyond the methodological advances, the proposed framework offers a practical tool to enhance system flexibility, facilitate net-zero compliance, and support both operational control and long-term planning of hybrid energy storage systems.