A neural modeling approach applied to energy conservation retrofits

This paper describes an artificial neural network- (ANN) modeling approach forpredicting the energv anddemand savings resulting from energy conservation measure (ECM) retrofits in select buildings. Simulated data sequences were used to minimize the experimental uncertainty in the initial model, and to provide post period data for training. A university building was chosen to provide the data set for this study. The building was modeled and calibrated using the DOE-2.IE building energy analysis program. Following the retrofit implementations, year-long sequences of hourly consumption data were generated, along with the corresponding climate data. Multi-layer feedforward networks were developed for the building and each ECM cosidered The input parameters and the network architecture were selected to optimize-the training time and generalization. The initial results show that the ANN method successfuliv predicted the response from a loads,system, and plant level retrofit, given a time-dependnt (dynamic) prediction problem. This paper will report the performance of the model, its- usefulness as a reliable predictor of building energy consumption, and explore the possibility for continued work to develop a library of ANN ECM models.