On Predicting Energy Consumption in Administrative Buildings: Neural Network Applications

Abstract

The objective of this paper is to analyze and predict the energy consumption in an air conditioned administrative building by using Artificial Neural Networks (ANN). The model is constructed by a simulation package called Design Builder where EnergyPlus is the core program which is provided with actual weather data for Cairo and simulated to produce the system performance hourly data (Total energy consumption of the Building, Chiller energy consumption .etc). All this data are introduced to the ANN which use this data for training and forecasting. Different experiments and cases are done to evaluate the performance at different types of inputs. It was found that when handling data associated with energy in administrative buildings there are difficulties where many variables become involved. Different combinations of Input parameters and different structures of ANN are used. A feed forward backpropagation with two hidden layers (tan-sigmoid and linear activation function) is the structure of ANN which is trained by Levenberg Marquardt algorithm. The method of choosing the ANN structure has provided a successful way for selecting the number of input variables. The proposed method of selecting the number of neurons in the hidden layer is introduced by plotting the errors and performance parameters versus the number of neurons for each case and comparing the error values and selecting the lowest RMSE at each case to get the best performing size. It was found that changing the input data for the ANN has a great effect on its predicting ability. Different performance parameters are defined for assessment of the ability of the trained neural network such as Root Mean Square Error and coefficient of variation where at the best case they became 4.337 KW and 0.045. This will introduce the ANN capability as an additional tool for solving design issues and will help in getting an instantaneous opinion on the effect of a proposed change in a design related to building energy systems prediction and modeling for an air-conditioned administrative building instead of complex equations and regression analysis. This model can also be potentially used for optimization of building control systems.