WiseLSTM: a depthwise CNN--LSTM framework with metaheuristic optimization for offshore wind power forecasting

Reliable one-hour-ahead wind-power forecasting can support scheduling andenergy-management decisions in renewable-to-hydrogen systems. We proposeWiseLSTM, a hybrid architecture combining two depthwise-separable temporalconvolutional blocks with unidirectional Long Short-Term Memory (LSTM) modeling.Using 40 years of hourly ERA5-derived meteorological conditions and model-derivedwind-farm power for Baie de Saint-Brieuc, evaluated with a five-fold blockedrepeated-holdout protocol, WiseLSTM achieves an RMSE of\((33.91±3.70)\) MW, MAE of \((20.40±3.00)\) MW, and\((R^2=0.9657±0.0076)\). It reduces RMSE by approximately5.0% relative to CNN--LSTM and 8.5% relative to one-hour persistence. At matchedchannel widths, the depthwise front-end uses 65.8% fewer convolutional kernelweights than standard convolutions. Particle Swarm Optimization (PSO) and WhaleOptimization Algorithm (WOA) are further used to explore five mixedhyperparameters. The WOA-selected configuration achieves an RMSE of\((33.67±1.24)\) MW and reduces the cross-fold RMSE standard deviation by 66.5% relative tountuned WiseLSTM. These results highlight the joint importance of efficient localfeature extraction, recurrent temporal modeling, and configuration stability foroffshore wind-power forecasting.