Hybrid method for building energy consumption prediction based on limited data

The blossoming of building related data has led
to the rapid development of machine learning methods in
building energy consumption prediction. This has also allowed
for the strengths and brilliance of machine learning methods
over popular statistical methods such as seasonal autoregressive
integrated moving average (SARIMA) to be exposed. However,
for some old buildings that cannot provide sufficient data, it
would be intractable and inefficient to apply machine learning
methods to predict energy consumption. In this study, a hybrid
method based on SARIMA and support vector machine (SVM)
was proposed to predict the energy consumption of a relatively
old educational building that solely had electricity consumption
data. The performance of proposed method was compared with
SARIMA. The results showed that SARIMA accurately
captured and predicted linear aspects of the building energy.
Although SVM is proficient for capturing inherent non-linearity
within limited data, the lack of input variables such as occupant
behaviours often restrict SVM accuracy. Multiple comparisons
between 1-year and 2-year training data indicated that
extending time spans of training data only marginally improves
prediction performance. In this study, the accuracy was
impeded by lack of adequate information about the building
closure during festive periods.