Genetic algorithms based optimization of artificial neural network architecture for buildings' indoor discomfort and energy consumption prediction

Growing concerns about energy consumption reduction and comfort improvement inside buildings make it more and more necessary to be able to predict with fine precision building's heating loads and indoor discomfort. This article proposes a method based on genetic algorithms (GAs) to optimize the architecture, training parameters and inputs of an artificial neural network (ANN). The ANN is doomed to predict energy consumption and indoor discomfort in future work on the development of an on-line method for control setting optimization. Simple and advanced controllers were used in this study: ON-OFF, PID and fuzzy controllers. Validation of the optimized ANN showed good prediction accuracy, as regression coefficients (2) for consumption and discomfort were respectively greater than 0.77 and 0.84 for the three tested controllers. Various prediction distances and ANN training data quantities were tested. Conclusion is that prediction at a 2-hour distance and a 3-day quantity of data are the best tested optimization conditions.