Comparing the accuracy of ANN and ANFIS models for predicting the thermal data

Abstract

The study aims to propose a suitable prediction model to deliver the full heating season’s thermal performance dataset by using short-term measured data during the system operation period. Two machine learning-based models, BackPropagation Artificial Neural Network and Adaptive Neuro-Fuzzy Inference System are compared by utilizing the measured data of indoor temperature and relative humidity. The independent variables of the prediction are obtained from the weather data, in addition to the building energy simulation model. Conversely, the data of the dependent variable are obtained from the real measurements from inside of the building for 31,5 days of the heating season, starting from February 22 nd , which is called the first heating season. Moreover, the entire heating season of the building is evaluated between November 15 th and March 21 st , which is called the second heating season when the building’s monthly consumption exceeds 14 kW/m 2 . The first prediction approach is the feed-forward Artificial Neural Network (ANN) with Back Propagation Learning System (BPS). Four ANN models are structured by input-output and one hidden layer is performed. The second prediction approach is the Adaptive Neuro-Fuzzy Inference System (ANFIS). The Sugeno ANFIS method is utilized in this prediction work. Eight ANFIS models are structured by 6 layers are performed to achieve the prediction. Besides, the main motivation for approaching ANFIS is to avoid the stochasticity of the measured temperature and humidity data. The prediction results are compared with the measured data of the second heating season. The comparison showed that the ANFIS model is more efficient since it achieved an 85% accuracy rate for the indoor temperature and 81% for the humidity prediction. While the ANN prediction accuracy is 81%, 80% relatively for the temperature and humidity. Then the comparison is scaled by selecting the most ordinary period in the measured data to be the data sample that will be used in the comparison. The second comparison showed that the ANFIS model is once again better than the ANN model since the ANFIS prediction accuracy becomes 88% for temperature and 90% for humidity, while the ANN prediction accuracy becomes 83% for temperature and 87% for humidity. Nevertheless, the stochasticity of the measured affected the prediction results in accuracy rates. Hence, according to the achieved accuracy rates, both the ANFIS and ANN approaches are highly validated in this type of prediction.