Abstract -- This paper proposes a self-attentive LSTM (LSTA) model for short-term forecasting of wind and photovoltaic (PV) power and investigates the operational impact of forecast errors on coordinated wind–PV–battery scheduling. Historical 2024 generation and load data from a North American region (15-minute resolution) are used to train and evaluate the models. The LSTA integrates an attention module into the LSTM gate structure to emphasize temporally salient features in volatile renewable generation series. Predicted 24-hour profiles are incorporated into a day-ahead scheduling model that minimizes total system cost (curtailment, storage investment and charge/discharge cost). Compared with a baseline LSTM, the proposed LSTA reduces wind RMSE from 895.67 to 324.67 MW and solar RMSE from 109.58 to 58.24 MW. When used in the scheduling stage, LSTA forecasts lower the wind–solar curtailment rate (from 10.9% to 3.1%) and decrease total operational cost relative to the LSTM-based schedule. Results demonstrate that embedding attention into LSTM improves forecast fidelity for highly variable renewable outputs, which in turn yields measurable reductions in curtailment and system cost in coordinated dispatch. Key contributions include the gate-level attention design, the end-to end prediction-to-scheduling evaluation, and a quantitative assessment of forecasting accuracy on dispatch outcomes.

Wind and photovoltaic power forecasting; LSTM; Attention mechanism; Coordinated dispatch; Energy storage