State transition stochastic model for predicting off to on cooling schedule in dwellings as implemented using a multilayered artificial neural network

Abstract

Our previous study (Tanimoto, J. and Hagishima, A. 2005. State transition probability for the Markov model dealing with on/off cooling schedule in dwellings. Energy and Buildings, 37, 181–187) proposed a set of state transition probabilities for the Markov chain dealing with the on/off cooling schedule in dwellings. The probability of turning on an air conditioner was defined in the form of a sigmoid function by the indoor globe temperature. Obviously, a real stochastic event of shifting from the off to on state is affected by not only indoor thermal quality parameters but also by other complex factors, such as the presence of family members, time of the day and whether it is a weekday or holiday. In this article, we report an alternate model, based on a multilayered artificial neural network (MANN), for predicting the off to on cooling schedule. We gathered field measurement data on family dwellings during the summer of 2008 by deploying hygrothermometers with recording functions to measure the room temperature and the globe and blowout air temperature of the air conditioner. The MANN used has nine nodes in both its input and hidden layers and a single node in its output layer, which implies that the state is either shifting from off to on (1) or not (0). The information provided to the input layer nodes includes the time of the day, whether it is a weekday or holiday, the probability of the presence of inhabitants and the predicted percentage of dissatisfied (PPD) people. PPD, derived from PMV theory, is applied as a representative parameter of the indoor thermal quality, in place of the globe temperature, since it accounts for various influences. The field measurement datasets were divided into two parts: teaching data and data for validation. A model trained by the teaching data was confirmed to reproduce the state transition characteristic of the validation period, which seems complex and is determined by the behaviour of various inhabitants. The performance of the model in reproducing this behaviour is improved over that of the previous model derived from the Markov chain.