Coupling input feature construction methods and machine learning algorithms for hourly secondary supply temperature prediction

Accurate supply temperature prediction plays a vital role in achieving meticulous management of heating station. However, there are relatively few studies on ultra-short-term supply temperature prediction at present. This paper comprehensively applied 4 feature selection methods and 3 prediction algorithms to estimate hourly secondary supply temperature. Taking a floor radiant heating system as the case, the correlation analysis (CA) based on the back propagation neural network (BPNN) model and the support vector regression (SVR) model shows that outdoor temperature, supply and return temperatures are the main input feature categories. This paper novelty proposed the categorical principal component analysis (CPCA) method, compared with the traditional principal component analysis (PCA), this method can reduce the root mean square error (RMSE) of BPNN model and SVR model by an average of 18.6% and 19.7%, respectively. The comparison of 4 historical input lengths for the long and short-term memory (LSTM) model shows that historical 12 h can fully consider the influence of building thermal inertia and heating system thermal delay for floor radiant. Further comprehensive comparison shows that the BPNN model based on correlation analysis has the best performance.