A Novel Hybrid Ensemble Wind Speed Forecasting Model Employing Wavelet Transform and Deep Learning

Abstract

Efficient wind speed forecasting is crucial for operations, optimizations, and decision-making interventions in wind energy systems. However, capturing nonlinearity and relevant information from the wind speed data poses challenges in developing efficient wind speed forecasting models. The present study proposes a novel hybrid ensemble wind speed forecasting model based on signal decomposition, deep learning model, and hyperparameter optimization for short-term applications to improve the model performances. This study comprises a novel architecture, a novel hybrid ensemble wind speed forecasting model, a two-level optimization strategy, and a transfer learning approach. The present study consists of three stages: model development, validation, and transfer learning. The proposed model employs wavelet transform, deep learning models such as Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Convolutional Neural Network (CNN), and a combined model using Convolutional Neural Network and Bidirectional Long Short-Term Memory (CNN-BiLSTM) and meta-heuristic optimization algorithms. The novel architecture of the CNN-BiLSTM model is capable of exhibiting better results than baseline models. Artificial Bee Colony (ABC) and the Differential Evolution (DE) algorithms are explored to optimize the model hyperparameters. The ensemble weights of the proposed model are optimized through a DE algorithm. The model implementation is presented through a transfer learning technique using pre-trained models from the model development and validation phases. The model comparison results indicate that the proposed models outperform these models. The transfer learning results of Proposed Model-1 (PM-1) are Root Mean Squared Error (RMSE)- 0.1943 m/s, Mean Squared Error (MSE)- 0.0378 m/s, Mean Absolute Error (MAE) 0.1542 m/s, coefficient of determination (R²)- 0.9883, and Index of Agreement (IA)- 0.9997. The Proposed Model-2 (PM-2) is 0.1554 m/s (RMSE), 0.0241 m/s (MSE), 0.1263 m/s (MAE), 0.9915 (R²), and 0.9998 (IA). The proposed model architecture and the transfer learning are viable approaches for wind speed forecasting applications.