Paper Conference

Proceedings of eSim 2016: 9th Conference of IBPSA-Canada


Method for validation of statistical energy models

Miroslava Kavgic, Trent Hilliard, Lukas Swan, Zheng Qin

Abstract: Advanced building control technologies such as model predictive control rely on fast executing statistical models. These models likely use large quantities of training data to accurately predict a buildings state. It is important to have evaluation techniques that are able assess accuracy of both the aggregated and timestep predictions in order to have confidence in statistical model predictions. There is presently no standardized procedure for validation and verification of statistical models against test data. This paper presents two tools that have been developed for such high-resolution (e.g. 15 minute) statistical models evaluations: a Residual Analysis tool and an Absolute Percentage Error tool. These tools uses a varying limit curve for timestep data and defines benchmarks for monthly/annual energy consumption and indoor air temperature accuracy.
Pages: 426 - 435