Chiller Load Forecasting Using Hyper-Gaussian Nets

Abstract

Energy load forecasting for optimization of chiller operation is a topic that has been receiving increasing attention in recent years. From an engineering perspective, the methodology for designing and deploying a forecasting system for chiller operation should take into account several issues regarding prediction horizon, available data, selection of variables, model selection and adaptation. In this paper these issues are parsed to develop a neural forecaster. The method combines previous ideas such as basis expansions and local models. In particular, hyper-gaussians are proposed to provide spatial support (in input space) to models that can use auto-regressive, exogenous and past errors as variables, constituting thus a particular case of NARMAX modelling. Tests using real data from different world locations are given showing the expected performance of the proposal with respect to the objectives and allowing a comparison with other approaches.