Long-term prediction of temperature data for the assessment of future trends of the building heating and cooling loads

Abstract

The present work focuses on the long term prediction of temperature data employing neural network models. Primarily, a benchmarking auto regressive model is developed. Then, different neural networks are developed regarding the network type, the training function and the training intervals. Temperature predictions are calculated for ten and for five year intervals. Each model’s results are compared with the corresponding real temperature data, in terms of mean, maximum and minimum temperature values, cooling degree days and frequency distribution. The best predicted temperature data are used as outdoor temperature for the heating and cooling loads calculations of a typical office building. The building simulation model which is used for the energy demand calculations is the open source ESP-r model. The results indicate a relative accurate potential of the neural networks for the simulation of the mean temperature data and prediction of the cooling degree days. Regarding the high temperature values and the maximum peaks, the neural network models are unable to reach precise values, due to the lack of similar training data. As a result, the cooling loads calculated from neural network predictions are underestimated, while the heating loads prediction is more accurate.