Data-driven modeling of solar power output using a CNN–LSTM approach

Abstract

Short-term photovoltaic (PV) power forecasts are essential for storage dispatch, reserve scheduling, and grid safety, yet remain challenging under rapid irradiance ramps and seasonal regime shifts. We present a compact, causal CNN–LSTM architecture that couples local temporal pattern extraction with long-range sequence memory, augmented by physics-aware features (solar geometry, plane-of-array irradiance, clear-sky indices) and strict leakage safeguards. Using a one-hour-ahead task, we evaluate on a 2023 Accra, Ghana simulation study built with PVWatts v8 driven by NSRDB PSM v3.2 (60 kWp DC, 55 kW AC). Metrics are reported in kW and normalized to DC capacity, with daylight/overall splits for fairness. The proposed model achieves RMSE = 0.127 kW, MAE = 0.092 kW, and $$R^2$$ = 0.956 on the test split, reducing RMSE by 21.6% vs. the best CNN-only variant and by 12.4% vs. the best LSTM-only variant. Ablations show that engineered features and a Box–Cox target transform improve stability and accuracy; a 24 h look-back provides the best accuracy–latency balance, and moderate convolutional width ( $$k=5$$ , $$m=32$$ ) with $$d=128$$ LSTM units is near-Pareto-optimal (about 0.093 M parameters and 2.20 M MACs per step). Baselines (persistence, clear-sky-scaled smart persistence, and GBRT) are included to contextualize deterministic accuracy and skill. We also provide error anatomy by hour and season to highlight residual risks at dawn/dusk and during fast cloud transients. While results are strong, they reflect a simulation (plain PVWatts; no row-to-row shading or sensor noise). We outline a path to operational validation on measured plant AC data across seasons/sites and discuss extensions to probabilistic forecasting with calibrated intervals.