Energy Consumption Prediction Model for Smart Homes via Decentralized Federated Learning With LSTM
Dawid Połap, Gautam Srivastava, Antoni Jaszcz
Abstract
The rapid pace of development of the Internet of Things and the requirements of various devices have allowed us to perform calculations at the edge, especially in terms of consumer electronics. Such progress makes it possible to design new solutions for energy distribution and prediction for smart homes. In this paper, we propose a solution that can be used to optimize energy distribution by analyzing the energy demand in individual homes. The proposed methodology is based on edge technology, where a dedicated LSTM network with a multi-head self-attention network is trained with measurement data from different sensors for predicting energy demand. Training of this network is extended to a decentralized learning process with an additional aggregation decision module (that allows rejection of the model in case of worst adaptation to private data). In order to increase data security, we added a blockchain network with a Byzantine strategy and Proof of Stake (PoS) consensus. The solution was tested for a publicly available database in order to demonstrate the possibilities and advantages of such an architecture.