Exploring Federated Learning for Energy Consumption Forecasting in Smart Homes

With the increasing deployment of smart meters in homes, there is a growing need to enhance privacy in the management of energy consumption data. This paper investigates the application of Federated Learning (FL) as a decentralised approach to energy consumption prediction. Unlike centralised models that require the collection of data from multiple house-holds, FL preserves privacy by keeping data localised while aggregating learning parameters across households. Utilising the "Smart Meters in London" dataset, we compare the performance of locally trained Long Short-Term Memory models, a FL model, and a centralised model using all data from each household. To better reflect real-world applications, we adopt a sliding window approach to training, which requires fewer rounds of communication and only one model update per iteration, in contrast to traditional FL methods that involve multiple training rounds within each iteration. Contrary to expectation, our results show that the FL model outperforms both the local and centralised models in terms of prediction accuracy and generalisation. The study highlights the advantages of FL in balancing privacy with predictive performance, providing insights into its practical implementation in smart home energy management. These results provide valuable guidance for energy companies and stakeholders seeking to adopt FL approaches, particularly in scenarios where both privacy and accurate prediction are required.