Application of neural networks to predict indoor air temperature in a building with artificial ventilation: impact of early stopping

Indoor air temperature prediction can facilitate energy-saving actions without compromising the indoor thermal comfort of occupants. The aim of this study was to analyse the performance of various artificial neural networks with a view to proposing an optimal approach for predicting the indoor temperature of a tertiary building. The multi-layer perceptron (MLP), convolutional neural network (CNN), long short-term memory (LSTM) models and the convolutional neural network- long short-term memory (CNN-LSTM) combination were used and coupled with the optimisation algorithms (Adaptive Moment Estimation (Adam), Stochastic Gradient Descent (SGD)) and the independent hyper-parameters early stopping and dropout. The parameters used are outdoor ambient temperature, outdoor relative humidity, indoor relative humidity, wet bulb temperature, black globe temperature, indoor air temperature and mean radiant temperature. The data is collected in an artificially ventilated building in Yaoundé, Cameroon. A numerical code was developed in Python to run the simulations. In order to study the impact of the inputs parameters on the prediction, two scenarios were distinguished in this work: (1) all the parameters are input to the network, (2) only the parameters whose absolute value of the correlation coefficient was greater than or equal to 0.5 were used. The impact of early stopping is assessed by distinguishing two case studies: the first without early stopping, the second with early stopping. The results showed that without early stopping, the MLP, CNN, LSTM and CNN-LSTM networks are adequate for predicting the temperature with the second scenario, mainly with both the SGD and Adam algorithms, and CNN-LSTM is the most appropriate model because the mean square error (MSE) and mean absolute error (MAE) values obtained in this case were closer to 0. With early stopping, the learning time is reduced and the learning curves are improved; the models optimised better with the SGD algorithm in general, but the best neural network model was obtained with the Adam algorithm and the LSTM network for the performances MSE = 0.0005, MAE = 0.0130 with the second scenario.