Multi‐Step Forecasting of Wind Speed and Solar Irradiance Using Lag‐Optimized Stochastic Stacked LSTM Networks

The rapid expansion of solar PV and wind generation intensifies the challenge of day‐ahead grid planning under stochastic meteorological conditions. While the literature is dominated by sub‐day forecasting studies, day‐ahead predictions remain critical for unit commitment and grid scheduling. This study proposes a stochastic stacked Long Short‐Term Memory (LSTM) framework, tuned via Bayesian optimization, for simultaneous day‐ahead forecasting (96 step‐ahead) of wind speed and solar irradiation using high‐resolution 15‐min measurements from the Adrar 20 MW solar PV facility in Algeria. Four input strategies were rigorously evaluated: full autoregressive lags (S‐LSTM), Random Forest importance‐based lag selection (RF‐LSTM), engineered auxiliary features (AUX‐RF‐LSTM), and exogenous meteorological variables (EXG‐RF‐LSTM), benchmarked against Holt‐Winters and naïve persistence baselines. The parsimonious RF‐LSTM architecture demonstrated superior performance across all evaluation frameworks. In deterministic point forecasting, it achieved the highest accuracy for both variables (R2=85.12%$$ {R}^2=85.12\% $$ for wind speed and 97.15%$$ 97.15\% $$ for solar irradiation) while substantially reducing residual serial dependence relative to the Holt‐Winters benchmark. These gains were confirmed as statistically significant via the Harvey‐Leybourne‐Newbold corrected Diebold‐Mariano test across all pairwise comparisons. In probabilistic forecasting via Monte Carlo Dropout, the RF‐LSTM achieved near‐perfect calibration for solar irradiation (PICP=95.10%$$ PICP=95.10\% $$) and acceptable risk‐aware uncertainty boundaries for wind profiles. Interestingly, expanding the input space with engineered or exogenous features systematically degraded both point and probabilistic accuracy by introducing multicollinearity and optimization complexity.