Using an ensemble learning framework to predict residential energy consumption in the hot and humid climate of Iran

Abstract

To achieve the desirable energy conservation targets, it is important to use forecasting tools to predict the energy consumption based on several expected parameters. This study aims at developing an Ensemble Learning (ENL)-framework for predicting residential electricity consumption. Fifty parameters classified into categories of environmental, context and building, and occupant related factors in addition to time related and some extra inputs were considered for model development. A Random Forest (RF) algorithm was trained and tested for each one of the main input groups. Then, based on feature importance analysis, a combination of the most important parameters from these separated models was used to build a general model with 34 inputs. The results show that RF models based on building (RFC&B) or occupant parameters (RFOcc) are superior to the model based on environmental ones (RFEnv). Furthermore, RF model with a combination of environmental, building, and occupant-related inputs (RFCombined) has the best prediction performance with 0.99, 0.91, 101.80, 282.53, 159.53, and 456.31 values for RTrain2, RTest2, MAETrain, MAETest, RMSETrain, and RMSETest, respectively. Even with reducing inputs of the combined model to nine parameters, the model still shows an acceptable performance with R2 values of 0.98 and 0.89 for train and test sets. Moreover, comparison of feature selection results for RFCombined with the performance of separated models reveals that although RFC&B and RFOcc are superior to RFEnv, the importance of environmental parameters cannot be denied which indicates the strong role of providing thermal comfort in building energy consumption.