Journal of African Development

Deep hybrid architectures such as CNN–BiLSTM–Attention have established strong accuracy benchmarks for short-term wind speed forecasting, but their decision logic remains opaque to grid operators, wind-farm engineers, and regulatory auditors. Operational deployment of such models in the Indian power system, especially under the Central Electricity Regulatory Commission's Deviation Settlement Mechanism, increasingly demands interpretability   both to build trust and to satisfy auditability requirements. This paper applies two complementary Explainable AI (XAI) techniques SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations) to a CNN–BiLSTM–Attention model trained on 8,760 hourly SCADA observations from an operational Indian onshore wind turbine (2018 data). Attention weights are also extracted from the model itself and compared against the SHAP and LIME attributions. Results reveal strong convergence across all three interpretability methods: the four most recent hourly lags (t-1 to t-4) contribute approximately 70% of the total prediction influence, while the 24-hour lag (t-24) provides a secondary contribution reflecting diurnal cyclicity. These patterns are operationally meaningful and align with the known atmospheric physics of short-term wind dynamics. The paper additionally demonstrates a deployment-ready pipeline in which SHAP, LIME, and attention weights are integrated into a real-time operator dashboard for Indian wind-farm monitoring. The findings establish that the CNN–BiLSTM–Attention model is not only accurate but also interpretable, transparent, and audit-ready..