DEVELOPMENT OF MACHINE LEARNING-BASED PREDICTION MODELS TO FORECAST THE ENERGY CONSUMPTION OF PCM-INTEGRATED BUILDINGS

Abstract

A change toward more effective management strategies is required to reduce the increasing energy demand. Since buildings account for a considerable amount of the world's total energy usage (almost 40%), they provide significant opportunities for improvement of building design, using advanced methods to optimize the design features of buildings to decrease energy utilization. Hence, accurate machine learning predictions for the energy consumption of buildings incorporated with PCM can be critical to constructing energy-efficient buildings using few resources. Therefore, it is pertinent to consider the wide range of building and environmental parameters to evolve a robust forecasting model that can forecast the optimized building design characteristics for maximum energy savings according to the available climatic conditions. Prior to this research, there was a scarcity regarding the development of a best-performing machine learning-based prediction model to predict the energy consumption of buildings incorporated with PCM. Nevertheless, some building and environmental parameters can be less influential through the evaluation of energy use for PCM-incorporated buildings, leading to a decrease in prediction model accuracy and generalizability while simultaneously increasing the utilization of excessive resources. For this reason, developing a prediction model that considers only influential building and environmental parameters using feature selection methods can increase the model's reliability, accuracy, and generalizability with less complexity through model development process. For the first time, this research presents a feature selection process-based metaheuristic framework to develop the forecasting model for predicting the energy consumption of buildings incorporated with commercially available PCMs, considering only influential building and environmental parameters. The building’s layout, thermophysical characteristics, and energy efficiency measures were considered building parameters, and BSh climatic conditions were considered as environmental parameters to evaluate the energy consumption of PCM-incorporated buildings through the energy simulation process. This provided an energy consumption database of 194,400 data points according to the defined range of each parameter as a variable. Afterward, feature selection methods, including Pearson, Spearman, and Kendall correlation methods, were utilized to recognize the influential building and environmental parameters for formulating several machine learning-based prediction models. Using a reduced energy consumption database (162000 data points) for influential parameters, multiple GEP, SVM, and MEP-based estimating models were formulated, considering the deviations of their hyperparameter values to obtain the best-performing prediction model. The model evaluation and validation process revealed that the GEP-based (GEP10) forecasting model was the most impeccable and robust forecasting model for predicting the energy consumption of PCM-incorporated buildings in the selected climate. Throughout training, validation, and testing, this model (GEP10) achieved a R² value of more than 95%, indicating an accurate prediction. Further, the parametric study of the recognized estimating model showed that the variation of each building and environmental parameter value resulted in the change of PCM-integrated building energy consumption rendering to the system’s physical-boundary conditions. Finally, the energy-saving assessment of the GEP10 showed that the commercially available PCM with a melting temperature of 26°C was the optimum PCM for integration with buildings to reduce energy consumption by approximately 16% in the chosen climate zone. In conclusion, the proposed framework evolved a robust forecast model to reliably estimate the energy use of PCM-integrated buildings in the chosen climate. This framework can also be utilized to develop the best-performing forecasting model for other climates.