Explainable Net-Load Forecasting for High-Penetration Solar-PV-Plus-Storage Grids: A West African Case Study

Utility scale solar photovoltaic capacity paired with battery energy storage is changing the forecasting problem facing power system operators in emerging grids. The relevant quantity is shifting from gross demand to net load, demand net of variable renewable generation, and it must be forecast with limited historical operational data. Explainable deep learning for load forecasting has advanced rapidly, but almost entirely on data rich grids in Europe, North America, and China. This paper proposes TFT lite, an explainable Temporal Fusion Transformer style architecture combining a variable selection network, an LSTM encoder decoder, and multi head temporal self-attention, for day ahead net load forecasting under high PV plus BESS penetration. The framework is validated on a case study built around the awarded engineering, procurement, and construction contract for Ghana’s Pwalugu Solar PV Plant with Integrated BESS (50 MWp DC, 44 MW AC, 15 MW and 25 MWh usable BESS, Upper East Region), calibrated against the project’s PVSyst simulation report and Comprehensive Design Description: the synthetic dataset’s annual generation of 83,332 MWh matches the documented full-plant Year-1 figure of 83,335 MWh to within 0.004 percent. On a chronologically held out test set of 164 forecast days, TFT-lite achieves RMSE and MAE of 2.310 MW and 1.719 MW against persistence, Holt Winters, XGBoost, and LSTM baselines. A five seed robustness check shows XGBoost holding a small, consistent, but not statistically significant point accuracy edge (paired t-test p equal to 0.122) with lower run-to-run variance, while TFT-lite significantly outperforms the LSTM ablation and persistence (p less than 0.001 for both), isolating the variable selection network and attention layer as the source of that gain. TFT-lite also produces calibrated 10 to 90 percent quantile forecasts and two native interpretability channels, variable selection weights and temporal attention. Explainability analysis, corroborated by SHAP analysis of the XGBoost baseline, shows solar related features dominating midday horizons and becoming irrelevant after sunset, with direct relevance to battery dispatch scheduling. Code is released to support reuse in similarly data constrained solar storage grids across Sub Saharan Africa.