Linear, fuzzy and neural networks models for definition of baseline consumption: Early findings from two test beds in a University campus in Portugal

This paper presents a comparative study of modelling a baseline of electricity consumption in two experimental spaces in a Portuguese University campus: one amphitheater and one library. Five input variables were defined for the study: occupation, day length, solar radiation, solar radiation from the previous day and heating and cooling degree days. Current performance and verification protocols accept linear regression models to quantify savings. We present neuro-fuzzy models as an alternative since energy consumption may not be described only by linear models. For each space, a linear regression model was applied, followed by three fuzzy models and one neural network model. The performance of each is assessed and compared in this work. Linear regression and fuzzy models are considered as less adequate to describe electricity consumption under the experimental setups after analyzing the performance indexes and the study of the output profiles. Neural network models give better performance indexes, although they still result in low VAFs equal to 40.4% for the amphitheater, 44.3% for the library in the warm season and 55.8% in the cold season. Constraints were identified such as the precision of the occupation characterization and new models are proposed that take into account the identification of discrete events.