Politecnico di Torino - Corso Duca degli Abruzzi, 24 - 10129 Torino, ITALY

+39 011 090 6100 info@tech-share.it


energy consumption monitoringenergy efficiency buildingspredictive control


The invention relates to a method for optimising the energy consumption of a building, based on Machine Learning and System Identification type modelling algorithms. The purpose of this solution is to integrate and interface with exogenous switched auto-regressive (SARX) predictive control methods, minimising complexity and enabling closed-loop control of buildings for energy efficiency and environmental comfort.

Technical features

Predictive models have been widely used over the years to design optimal control strategies to save energy. Dynamic models are applicable for predicting the energy needs of a building, such as Model Predictive Control (MPC). However, two problems arise: in the application of MPC, making a mathematical model of the building is time-consuming and expensive, especially for large/multiple buildings; data on the materials used for construction are often not readily available (old buildings), and assumptions have to be made that require expert know-how. Thus, the invention (TRL 5) concerns an alternative data-driven methodology to the previous ones, aimed at refining the SARX model (switched exogenous auto-regressive predictive control method), with the objective of reducing the number of sub-models and recognising linear or quasi-linear models, without changing the regression tree RT structure, and thus the partitioning. Drastically reducing the number of discrete modes of the SARX model improves modelling accuracy and computational cost at runtime, mitigating overfitting problems.

Possible Applications

  • Method implemented to optimise the energy consumption of a building equipped with energy control equipment, by controlling this equipment.


  • Very high experimental performance, especially when applied to real systems whose model is unknown and quite complex to derive using physics-based approaches.
  • Improved modelling accuracy and computational cost at runtime.
  • Reduction in the number of sub-models of a SARX model without compromising model accuracy and mitigating the overfitting problem.