Manu, Abdurrahman and Haja Mohideen, Ahmad Jazlan and Akhter, Muhammad Naveed and Ahmad Huzri, Muhammad Danial and Mohd Ibrahim, Azhar and Mohd Zaki, Hasan Firdaus and Sreeram, Victor
(2025)
Industrial implementation of machine learning forecasting algorithms for large-scale Solar Heat for Industrial Process (SHIP) Systems.
IEEE Access, 13 (2025).
pp. 204978-204995.
ISSN 2169-3536
Abstract
Natural gas boilers are commonly used to fulfill the heated water and steam requirements needed for running specific processes inside oleo-chemical factories. The vast usage of natural gas boilers results in significant greenhouse gas emissions. Therefore, solar heat for industrial process (SHIP) systems are being introduced around the world as a clean alternative source of energy for supplying heated water and steam. Optimizing these SHIP industrial processes and consequent energy management requires accurate thermal energy forecasting. In this study the thermal energy forecasting of a large-scale SHIP system was performed. This SHIP system was installed at an oleochemical factory in Johor Bahru, Malaysia. A deep learning method (LSTM) was proposed for an hour ahead forecasting of the SHIP system output on a quarterly basis for a one-year duration. The performance of LSTM was compared with other machine learning methods namely Support Vector Regression, Random Forest, Decision Tree and XGBoost. Finally, three hybrid models (PSO-LSTM, GA-LSTM, SSA-LSTM) were developed using different parameter optimization techniques to tune the hyperparameters of the developed deep learning method (LSTM) in order to enhance its forecasting accuracy. The models were evaluated based on their predictive accuracy using metrics such as R2, MAE, correlation coefficient (r), and MSE. While all models demonstrated comparable accuracy in terms of R2, MAE, and r, the PSO-LSTM model notably excelled in reducing MSE, suggesting a superior capability in managing large prediction errors and outliers. However, in terms of runtime, the GA-LSTM model significantly outperformed the others. These findings indicate that while the models share similar predictive accuracies, their practical application might be differentiated by considerations of computational efficiency.
Actions (login required)
 |
View Item |
Download Statistics
Download Statistics