A Hybrid Deep Learning Model Integrating CNN–LSTM and Metaheuristic Optimization for Seasonal Wind Power Forecasting

This study introduces a medium-term wind power forecasting approach based on hybrid deep learning architectures integrating convolutional and recurrent neural networks with metaheuristic optimization. Three hybrid models CNN-LSTM, ISO-LSTM, and SO-LSTM were trained using 2024 daily meteorological and power data from a 52.80 MW wind farm in the Marmara Region of Turkey. Seasonal performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) metrics. Among the models, CNN-LSTM achieved the highest accuracy, particularly in summer, with significantly lower error values (MAE = 9.02 MW, MAPE = 4.45%, RMSE = 10.52 MW). ISO-LSTM demonstrated moderate accuracy, while SO-LSTM showed higher errors, especially during autumn and winter due to increased meteorological variability. The strong performance of CNN-LSTM is attributed to its efficient spatial feature extraction and ability to capture medium-term temporal dependencies. Overall, the CNN-LSTM model provides a reliable and practical forecasting tool for real-time power prediction and grid applications under varying climatic conditions.