Coal consumption prediction model of space heating with feature selection for rural residences in severe cold area in China

Identification of optimal subsets of input variables is a primary task in data-driven prediction modeling for the coal consumption for space heating (CCSH) of rural residences. However, most related predictive models of CCSH in rural residences ignore the nonlinear relationship between the factors and CCSH, and are short of feature selection process. This paper proposed an enhanced CCSH prediction model with learning-based optimized feature selection based on the measured weekly CCSH during real operation in Chifeng, Inner Mongolia, China. Partial least squares regression and random forest were employed to rank the features, and ten models with various input subsets were established by support vector regression. The prediction accuracy of the ten models was compared and the optimal features were examined based on the coefficient of variation of the root mean square error (CVRMSE), coefficient of determination (R-2) and model generalization ability. Furthermore, the residual errors between predicted and measured CCSH are distributed around zero evenly and extracted from the normal distribution for the optimized model. Particularly, we employed the best model to predict the aggregate CCSH at the district level. The prediction model with the optimal inputs was verified to be reasonable and accurate at the individual and district scales.