Federated learning based appliance-level energy demand forecasting for residential buildings

Energy demand forecasting plays a vital role to plan electricity generation effectively in Smart Grids. With increasing electricity demand from residential buildings, a deeper understanding of individual appliances' consumption patterns becomes necessary. Most of the existing studies forecast the aggregated energy consumed by all household appliances. They lack granularity about the individual appliance's energy consumption. A few other studies perform appliance-level energy demand forecasting in a single household. However, they neither generalize nor scale well, even for a single appliance type from multiple households. Moreover, they use a centralized method to train the model raising privacy concerns on sensitive data. Our solution proposes a class-based grouping approach to group appliances with similar characteristics from multiple households and performs appliance-level energy demand forecasting for sets of appliances. We design our model using an LSTM (Long Short-Term Memory) network. We employ Federated Learning (FL) to mitigate privacy concerns and reduce the communication overhead of sharing the raw data to the server. We propose an improvised distributed model optimization algorithm, Fed-Adamax, over the existing FedAvg optimization algorithm with our FL-based approach. We tested the performance of our FL-based solution using two real-world datasets. We performed experiments on appliance classes such as refrigerators, lighting devices, microwave ovens, dishwashers and air conditioners, and our FL-based approach achieves better accuracy on all appliance classes than the models designed using the existing centralized approach.Master of Engineerin