Comparison of machine-learning models for predicting short-term building heating load using operational parameters.

• The importance of 14 parameters for short-term heating load were determined by Relief. • 15 machine-learning models were compared from aspects of accuracy, stability and computation time. • GPR and SVM were recommended for small and large building load datasets, respectively. Short-term building energy consumption prediction is of great significance to the optimal operation of building energy systems and conservation. Machine-learning models are widely used due to their high prediction accuracy and efficiency in dealing with high-dimensional nonlinear problems. To compare the performance of different machine-learning models in building energy consumption prediction, this paper analyzes 15 machine-learning models, including a multi-layer perception, a radials basis function neural network, a generalized regression neural network, an extreme learning machine, a support vector machine, a least-square support vector machine, a Gaussian process regression, a regression tree, a model five tree, a random forest, a gradient boosting decision tree, an extreme gradient boosting tree, a light gradient boosting machine, a categorical gradient boosting tree and a multi-adaptive regression spline, from the aspects of model prediction accuracy, model stability (over-fitting) and calculation speed. The operation parameters determined by ReliefF algorithm were used as input parameters. The results showed that the prediction accuracy of all models is higher in the training phases, with R2 values greater than 0.90, while the prediction accuracy in testing phases was much lower. In terms of prediction accuracy and model stability, the Gaussian process regression model had the best overall performance among the 15 models, while the support vector machine had the faster calculation speed with acceptable prediction accuracy. For small datasets, the Gaussian process regression model is recommended, and the support vector machine should be preferred for large datasets. The results of this paper can provide a basis for model selection and the establishment of combined models for predicting building energy consumption. [ABSTRACT FROM AUTHOR]