Predicting the cooling and heating loads of energy efficient buildings : a hybrid machine learning approach

Abstract

Accurate prediction of heating load (HL) and cooling load (CL) in residential buildings is essential for efficient building energy management. In this study, a hybrid machine learning (ML) model is developed to estimate the HL and CL of the energy efficient residential buildings from eight input parameters: the relative compactness, the roof area, wall area, surface area, glazing area, the overall height, the orientation, and the glazing area distribution. This is a novel model based on the multi-output radial basis function neural networks (RBFNNs) optimized using mountain gazelle optimizer (MGO). The performance of MGO-RBFNN is then compared with two metaheuristics based RBFNN models, namely particle swarm optimization (PSO) and Genetic algorithm (GA), and benchmarked against generalized regression neural network, extreme learning machine, and a gradient based RBFNN. The results show that the proposed multi-output optimized RBFNN models considerably improved both the training and testing accuracy performance of the original RBFNN. The MGO-RBFNN model (testing performance: determination coefficient (R2) of 0.99 and 0.97; root mean squared (RMSE) error of 0.84, and 3.09; and mean absolute percentage error (MAPE) of 3.90, and 13.02, respectively, for HL and CL) outperformed both the PSO and GA-optimized RBFNN, and other compared ML models. The relevancy factor analysis demonstrated that building energy consumption is positively correlated with relative compactness, roof area, glazing area, overall height, and glazing area distribution, whereas orientation, wall area, and surface area have a negative effect on it. Moreover, overall height and relative compactness are the most influential variables affecting building energy consumption.