Artificial neural network structure optimisation for accurately prediction of exergy, comfort and life cycle cost performance of a low energy building

In recent years, surrogate modelling approaches have been implemented to overcome the time and computational power demands of traditional building energy modelling. Artificial neural networks (ANN), due to their potential to capture building energy systems complex interactions are regarded as powerful surrogate models; however, the definition of optimal ANN structures and hyperparameters have been overlooked causing substandard prediction performance. The aim of this study is to present a novel hybrid neuro-genetic modelling framework developed as an open source tool capable of identifying optimal multi-input/multi-output ANN structures for accurately predicting building thermodynamic performance. The ANN optimisation process uses a genetic algorithm that minimises the root mean squared error (RMSE) data difference between the target and predicted values for both the training and testing data. As a case study, an archetype social house located in different climatic regions in Mexico is used. The ANN training database has been generated by simulating a sample of high-resolution energy models considering a combination of different active and passive energy strategies (input data) while calculating building exergy destructions, occupant thermal comfort and life cycle cost (output data). After automatically evaluating thousands of different structures, the neuro-genetic tool has identified a single deep ANN structure (3 hidden layers with 18, 17, 20 neurons respectively) capable of predicting the model's high output variability, achieving a prediction accuracy >0.95 for each of the outputs. The presented framework and tool can be adapted to further optimisation stages in the building design process and to solve similar problems in other research areas.