Comparative analysis of deep neural network architectures for renewable energy forecasting: enhancing accuracy with meteorological and time-based features

• Published in: Discover Sustainability
• Main Author: Khan S.; Mazhar T.; Khan M.A.; Shahzad T.; Ahmad W.; Bibi A.; Saeed M.M.; Hamam H.
• Format: Article
• Language: English
• Published: Springer Nature 2024
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85213685937&doi=10.1007%2fs43621-024-00783-5&partnerID=40&md5=f45c8c868b50c1efa158f9dcdfef51ae

This study evaluates and differentiates five advanced machine learning models—LSTM, GRU, CNN-LSTM, Random Forest, and SVR—aimed at precisely estimating solar and wind power generation to enhance renewable energy forecasting. LSTM achieved a remarkable Mean Squared Error (MSE) of 0.010 and R2 score of 0.90, highlighting its proficiency in capturing intricate temporal relationships. GRU closely followed, demonstrating potential as a viable option due to its remarkable combination of computational efficiency and accuracy (MSE = 0.015, R2 = 0.88). In datasets abundant in spatial correlations, the CNN-LSTM hybrid demonstrated its utility by providing novel insights into spatial–temporal patterns; nonetheless, it lagged considerably in accuracy, with a mean squared error (MSE) of 0.020 and a R2 of 0.87. Conversely, traditional models demonstrated a reliable albeit less dynamic ability to elucidate the complexities of renewable energy data; for instance, Random Forest exhibited a mean squared error (MSE) of 0.025, while Support Vector Regression (SVR) recorded an MSE of 0.030. The results affirm that deep learning architectures, particularly LSTM, offer a transformative method for renewable energy forecasting, hence enhancing accuracy and reliability in energy management systems. © The Author(s) 2024.