In recent years, increasing climate change has triggered the development of various techniques and strategies for predicting building energy consumption and thermal comfort. However, many features could affect prediction efficiency, and the coupling of appropriate learning algorithms with different sampling methods and input parameters requires further research. In this study, we evaluated 25 features of an office building model. Then, the comprehensive effects of five different sampling methods and cumulative contribution rates (CCR) of input parameters on the prediction performance of building carbon emission (BCE) and indoor discomfort hour (IDH) were explored in 10 machine learning algorithms. The results indicated that the Sobol sampling method could achieve the best prediction effect in the combinations of different contribution rates of features and machine learning algorithms. Meanwhile, an artificial neural network (ANN) was the best learning algorithm when the CCR was 100%. However, the optimal machine learning method corresponding to each CCR stage differed. When the CCR was reduced to 50%, only three influence factors were considered most important, and K-nearest neighbors (KNN) was the best prediction algorithm in this scenario, indicating that the input parameters were reduced by 88%, and the R2 value was only reduced by 6.1%. Therefore, we proposed a new strategy idea for the research of building performance prediction by determining the most important building impact factors and suitable machine learning algorithms, which could simplify the prediction process and improve prediction efficiency while reducing building parameters.
All Science Journal Classification (ASJC) codes