Predicting foundation heat losses: neural networks versus the basesimp correlations

This paper describes a series of tests that were performed to determine whether a neural-network model could outperform a correlation-based model in representing foundation heat losses. The two models were trained with data generated by BASECALC, a finite-element-based program for modelling foundation heat losses. The two models are described along with details of the tests used to compare them. The most important conclusion of this work is that although both models accurately represent the BASECALC data, the NN model outperforms the correlation-based model in the majority of the tests. This observation has greater implications in terms of time rather than accuracy. The use of neural networks rather than correlations could significantly reduce the development time of regression-based algorithms for building energy programs. Although correlation techniques may be preferable for some applications due to their closed-form nature, neural network models should be given due consideration.