Transfer Learning for Energy Consumption Forecasting in Smart Buildings

Energy consumption forecasting is crucial in smart buildings so as to effectively manage energy usage and minimize the expenses. The conventional methods of machine learning are not always efficient to achieve generalization when applied on as-built buildings of different types because of architectural or occupancy differences or environmental indices. The major focus of this paper is to discuss the possibility of using transfer learning in improving energy consumption forecasting in smart buildings given the fact that huge datasets are available to pre-trained models. Real data containing a wide range of buildings’ characteristics and random occupancy profiles is used in addition to simulated data to extend the real-world data and to further adjust models for certain cases. The proposed work includes testing of some new transfer learning algorithms that target at adaptability of the models and quality of their predictions. To compare their efficiency in various scenarios, these algorithms are assessed via several measures, thereby indicating the opportunities for enhancing the performance of forecasts with the help of transfer learning in various types of buildings. The work’s results show that the proposed transfer learning method acts as a very solid solution for energy forecasting in smart buildings, opening the door for more dynamic and efficient energy management in these contexts.