Applying Artificial Neural Networks to Estimate Bldg Energy use in Early Stages

Session 14 Paper 2, CIBSE ASHRAE Technical Symposium, Imperial College, London, UK
18th and 19th April 2012

A pilot study was conducted to evaluate the application of an artificial neural network (ANN) for forecasting building energy use based on building architectural and energy use parameters. The objectives were to ascertain the prediction performance of the ANN and to measure the causal strengths of each parameter on energy use.

Electricity and heating fuel use data from the Display Energy Certificate (DEC) scheme together with building parameters were collected for a sample comprising London university buildings. 148 datasets were compiled covering 97 different buildings.

For electricity use the ANN mean absolute percentage error reduced to 34.0%, a 31.8% reduction relative to a theoretical benchmark-based approach; for heat use it reduced to 25.1%, a 49.8% reduction against the benchmark. Prediction performance appeared to be restricted however, perhaps owing to the limited number of training patterns.

It was found that building activity, material, building services type and glazing type had the greatest causal strengths for both types of energy use; also that height was a strong determinant for heat use and summer sun hours and aspect ratio were significant for electricity use.

From the pilot study the ANN approach appears to be suitable for this application and may also offer advantages relative to other approaches such as benchmarking and simulation-based modelling. A broader follow-up study is planned accordingly and measures to develop the methodology are presented.