Estimating Energy Consumption of Educational Spaces Using Artificial Neural Networks (ANNs)

Size of classroom’s windows has significant effects on both comfort level of users and electricity consumption for lighting. Moreover, windows are the main source of energy loss in classrooms in both cooling and heating sectors.
Considering the large number of educational buildings and long life cycle of such them, choosing proper window size is crucial for energy saving in sustainable architectural design. Despite the role that windows have in energy consumption, the literatures are surprisingly limited in providing detailed recommendations for architects in determining the appropriate window size in different climates. Therefore, energy based window design has always been complicated for architects due to the number of involved different components and variables. In order to help the architectural designers, in this paper a new methodology is developed using a well-known artificial intelligence technique. In the proposed methodology, a predictive model for energy consumption cost in terms of window to wall ratio (WWR) and the window facing was created using Artificial Neural Network (ANN). The methodology consisted of a limited sets of direct numerical energy simulations for any specific climatic zone to generate the data required for training the ANN. The DOE-2is suggested in the proposed methodology for direct numerical energy simulations of the daylighting scenarios required for training the ANN. The DOE-2 is a popular and powerful computational model developed with financial support of U.S. department of energy. The trained ANN-based model provides a fast and convenient way of comparing the different daylighting scenarios in designing stage. Indeed, further calculations for direct energy simulations are not necessary and an architect can readily utilize the trained ANN-based model as a powerful tool for forecasting the total energy consumption cost. In order to show the applicability and performance of the proposed approach, 288 daylighting scenarios for a standard classroom in a warm and dry climate, Shiraz-Iran, were simulated to determine the corresponding electric and gas consumption. A square classroom of side 7.4 m is the standard classroom defined by Iranian Organization for Renovating, Developing and Equipping Schools. The DOE-2 is utilized for simulating the defined standard classroom in the study area for estimating the annual gas and electric consumption of the generated scenarios over a 50 years period. Included daylighting scenarios were randomly split into train and test sets. In this study, around 80 percent of data were used for training, and the rest were used to evaluate the performance of the trained ANN. The best training and learning functions for different number of layers and neurons was determined in a trial-error process. Correlation Coefficient (CC), Mean square error (MSE) and Root mean square error (RMSE) are the statistical indices used for training procedure. The best results were obtained with 2 hidden layers and 6 neurons per layer. The 'Levenverg-Marquardt back propagation (trainlm)' and 'perceptron weight and bias learning function (learnp)' were the best training functions found for this research. The results show that the trained ANN can accurately predict the total energy consumption cost (RMSE=0.0811, MSE=0.0066, and CC=0.9672).