Enhancing the accuracy of global horizontal irradiance estimation model using convolutional neural network coupled with wavelet transform

• Published in: European physical journal plus 2024-10, Vol.139 (10), p.924, Article 924
• Main Authors: Gupta, Rahul, Jha, S. K., Jha, Pinaki, Chaprana, Kumud, Singh, Sarvendra Kumar
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Alternative energy sources; Applied and Technical Physics; Artificial neural networks; Atomic; Complex Systems; Condensed Matter Physics; Datasets; Deep learning; Energy industry; Estimation; Forecasting; Forecasting techniques; Irradiance; Machine learning; Mathematical and Computational Physics; Measurement techniques; Molecular; Multilayer perceptrons; Multilayers; Neural networks; Optical and Plasma Physics; Physics; Physics and Astronomy; Radiation; Regular Article; Root-mean-square errors; Solar energy; Sparsity; Support vector machines; Theoretical; Time series; Variables; Wavelet transforms
• ISSN: 2190-5444; 2190-5444
• DOI: 10.1140/epjp/s13360-024-05730-x

One of the most important tasks for managing and producing solar energy as a fossil fuel substitute is the accurate estimation of global horizontal irradiance (GHI). To improve the accuracy of GHI estimation, this work highlights the significance of investigating cutting-edge deep learning techniques and their combinations such as wavelet long short-term memory (WLSTM), wavelet multi-layer perceptron (WMLP), wavelet gated recurrent unit (WGRU), wavelet convolutional neural network is assessed for modeling global horizontal irradiance using various input combinations of climatic parameters for Rajasthan, a city situated in India. For an accurate assessment of the prediction accuracy of the proposed models, two robust statistical metrics such as mean absolute error (MAE) and root mean squared error (RMSE) are employed. With significant percentage decreases in MAE and RMSE, the results demonstrated that the WCNN technique outperformed the other models in estimating GHI. It is observed that the percentage reduction in MAE and RMSE for the standalone and wavelet-based models are (48.58, 15.67) for CNN, (63.86, 42.75) for GRU, (91.12, 22.09) for LSTM, and (33.40, 50.26) for MLP (52.80,76.34) for WLSTM, (50.14,13.63) for WMLP and (24.32,22.49) for WGRU, respectively. This study demonstrated significant improvements in prediction accuracy attained by hybridizing wavelet transform techniques with the CNN model. Based on climatic characteristics, this hybrid technique performed better at properly anticipating GHI. The results imply that more accurate and consistent GHI predictions can be achieved by utilizing the advantages of convolutional neural networks in combination with wavelet transforms.