Enhanced Wind Speed Prediction Using Dual-Memory LSTM: A Novel Approach to Temporal Dynamics

Abstract

Accurate wind speed forecasting is crucial for optimizing wind energy integration, ensuring grid stability, and advancing renewable energy systems. This study introduces the dual-memory long short-term memory (DMLSTM) model, an innovative extension of conventional LSTM architectures designed to explicitly capture both short- and long-term temporal dependencies. By incorporating separate memory cells and dynamic gating mechanisms, DMLSTM overcomes the limitations of traditional models—such as autoregressive integrated moving average (ARIMA), artificial neural networks (ANN), and baseline LSTM—in capturing the non-linear, stochastic, and hierarchical patterns inherent in wind speed data. Using a comprehensive meteorological dataset from Tetouan City comprising wind speed, temperature, and humidity, data normalization and outlier handling were applied to ensure high data quality. The DMLSTM model was trained using a sliding window approach to map historical sequences to future wind speeds. When evaluated against ANN, ARIMA, and baseline LSTM models using root mean square error (RMSE) and mean absolute error (MAE), DMLSTM achieved the lowest error rates across all metrics. Visual comparisons further demonstrate the model's robustness in capturing abrupt changes and complex temporal dynamics. Overall, these findings underscore DMLSTM's potential as a reliable tool for renewable energy forecasting and sustainable grid integration.