Convolutional Neural Network and Stochastic Variational Gaussian Process for Heating Load Forecasting

Heating load forecasting is a key task for operational planning in district heating networks. In this work we present two advanced models for this purpose, namely a Convolutional Neural Network (CNN) and a Stochastic Variational Gaussian Process (SVGP). Both models are extensions of an autoregressive linear model available in the literature. The CNN outperforms the linear model in terms of 48-h prediction accuracy and its parameters are interpretable. The SVGP has performance comparable to the linear model but it intrinsically deals with prediction uncertainty, hence it provides both load estimations and confidence intervals. Models and performance are analyzed and compared on a real dataset of heating load collected in an Italian network.