Effectiveness of neural networks and transfer learning for indoor air-temperature forecasting

Starting in 2007, EU set energy efficiency improvement targets in sectors with high energy-saving potential such as buildings. ICT allows innovative opportunities for energy consumption forecast to integrate with new control policies such as Demand/Response and Demand Side Management to reduce energy waste. However, such technologies must overcome challenges such as the lack of accurate historic data required for predictions. This article proposes an innovative methodology supporting the energy management of HVAC systems, through Smart Building indoor air-temperature forecast. The applicability of innovative neural networks for time-series predictions is explored. These neural networks are first trained on an artificial but realistic dataset based on BIM simulations with real meteorological data. The inference phase is then carried out on a second dataset collected by IoT devices. Finally, Transfer Learning techniques are exploited to improve the performances predictions. Fanger’s model is applied to validate results, showing consistent levels of accuracy and comfort.