Multi-source Transfer Learning in Reinforcement Learning-based Home Battery Controller

Abstract

On the path to decarbonization, batteries play an important role, as they can help manage uncertainties associated with renewable energy sources such as PV. One of the main challenges of developing a controller for home batteries is its scalability and transferability to other households. In this paper, we propose a multi-source transfer learning framework to scale a pretrained reinforcement learning (RL)-based policy for controlling different unseen households. Our proposed data-driven framework tackles two major disadvantages of vanilla RL methods, i.e., data-intensive training and scalability. In our proposed framework, initially, a global RL agent is trained on multiple source households. Thereafter, the pretrained global agent is finetuned on data from each target household to obtain an individual RL controller for each. We assess the performance of our proposed framework using real data from 100 Belgian households with different load patterns and weather conditions. We benchmark our proposed framework against a rule-based baseline. The results show that our finetuned controllers outperform the baseline rule-based controller by ~8%. Furthermore, compared to agents trained locally from scratch, our finetuned agents require significantly fewer training episodes to learn a good control policy on unseen, target households, validating the scalability of our proposed framework.