Advanced Statistical Models for Modeling Hot Water Consumption Using a Connected Boiler

In this paper we propose to carry out a proof of concept on the prediction of hot water by using the following four techniques: a supervised learning, a semi-supervised learning, a clustering and finally, we propose a new prediction approach based on the use of the Dempster Shafer algorithm (DST). We proposed several parameterizations of our algorithms used in order to obtain a better prediction of the consumption of hot water. In particular, to calculate the mass functions of the Dempster Shafer algorithm, we have subdivided the space of the features that were extracted from the data into cells. In order to simplify the calculations, we applied a criterion based on the use of correlation coefficients that make it possible to eliminate the least informative focal elements from the frame of discernment. The results show that the prediction of hot water consumption has reached more than 95% and 96% of classification accuracy using DST and Deep Neural Network algorithms (DNN) respectively. This study also shows that the use of the Dempster Shafer theory is effective especially since it allows us to take into account the uncertainty on the data coming from the Chaudière sensors that we used.