Unsupervised energy prediction in a smart grid context using reinforcement cross-buildings transfer learning

In a future Smart Grid context, increasing challenges in managing the stochastic local energy supply and demand are expected. This increased the need of more accurate energy prediction methods in order to support further complex decision-making processes. Although many methods aiming to predict the energy consumption exist, all these require labelled data, such as historical or simulated data. Still, such datasets are not always available under the emerging smart grid transition and complex people behaviour. Our approach goes beyond the state-of-the-art energy prediction methods in that it does not require labelled data. Firstly, two Reinforcement Learning algorithms are investigated in order to model the building energy consumption. Secondly, as a main theoretical contribution, a Deep Belief Network (DBN) is incorporated into each of these algorithms, making them suitable for continuous states. Thirdly, the proposed methods yield a cross-building transfer that can target new behaviour of existing buildings (due to changes in their structure or installations), as well as completely new types of buildings. The methods are developed in the MATLAB environment and tested on a real database recorded over seven years, with hourly resolution. Experimental results demonstrate that the energy prediction accuracy in terms of RMSE has been significantly improved in 91.42% of the cases after using a DBN for automatically extracting high-level features from the unlabelled data, compared to the equivalent methods without the DBN pre-processing.