Deep Transfer Learning in Smart Building and Links with Edge Computing

This paper presents an industrial case for deep-transfer learning on data forecasting in a real tertiary building. In reaction to the ever-growing quantity of devices and data, we investigate the interest in transfer learning AI algorithms between devices and provide figures about expected data forecast gains. The use case highlights the aspect of performance expectations and evaluation, showing the effect of the most popular policies for AI algorithm transfer. The main contribution includes using the Matrix Profile concept as a new technique to evaluate domain transfers and their implementation on edge devices. We highlight the potential of new layers on top of a frozen pre-trained neural network to simultaneously reduce the overall training time of AI algorithms and the forecasting error.