Enhancing Wind Forecasting Accuracy Using Decomposition-Based Hierarchical LSTM Model

This study proposes a Decomposition-based Hierarchical Long Short-Term Memory (LSTM) framework to enhance wind speed forecasting accuracy. The methodology integrates a dual-stage signal decomposition strategy, Singular Spectrum Analysis (SSA) followed by Variational Mode Decomposition (VMD), to generate a Hierarchical Time Series (HTS) structure comprising 25 components. Each component is forecasted using LSTM networks across three horizons (10-, 30-, and 60-minutes ahead), and the forecasts are reconciled using methods such as Bottom-Up (BU) and Minimum Trace with Shrinkage (MINTS). Comparative analysis with models like Support Vector Regression (SVR), eXtreme Gradient Boosting (XGBoost), and Ridge Regression demonstrates that LSTM_MINTS and LSTM_BU significantly outperform alternatives, achieving up to 88% improvement in accuracy. Statistical validation using the Diebold-Mariano (DM) test confirms the superiority of the proposed approach. The results highlight the effectiveness of hierarchical reconciliation in improving LSTM-based forecasts for non-stationary wind speed data, offering a robust solution for renewable energy planning and management.