Hybrid Wavelet-SVR, machine learning, and deep learning models for predicting clean energy markets amidst global events

Abstract

This study investigates hybrid machine learning models combined with wavelet transforms for predicting clean energy market dynamics from 01.04.2014 to 02.05.2024. Models such as support vector regression (SVR), artificial neural networks (ANNs), eXtreme Gradient Boosting (XGBoost), gradient boosting machine (GBM), long short-term memory (LSTM), and convolutional neural network (CNN) are compared to forecast the Nasdaq Clean Edge Green Energy Index (NasdaqClean). Discrete wavelet transform (DWT) and continuous wavelet transform (CWT) are used for feature extraction and visualizations, capturing both short-term fluctuations and long-term trends. Shapley additive explanations (SHAP) and permutation feature importance (PFI) assess feature contributions. Analysis across sub-periods, including the Paris Agreement, COVID-19, and the Russia–Ukraine conflict, reveals that different models perform optimally in different periods. Specifically, Wavelet-SVR emerges as the most accurate model in the entire dataset, before the Paris Agreement and Paris Agreement periods, demonstrating strong predictive power by reducing noise and enhancing feature extraction. LSTM performs best during COVID-19, capturing long-term dependencies and volatile market dynamics. Meanwhile, CNN yields the most accurate predictions during the Russia–Ukraine conflict, effectively identifying spatial patterns in the dataset.