Evolutionary Ensemble LSTM based Household Peak Demand Prediction

Abstract

The popularization of electric vehicle, the commercialization of micro-generation, and the advance of local storage lead great challenges to the local power grid on household and neighbourhood level. A potential solution is to construct a home/neighbourhood energy management system (HEMS) to coordinate all available electrical equipment together using AI. As a portion of HEMS, peak demand prediction is critically important on triggering load scheduling among the household power environment to achieve better electricity usage curve. Long short-term memory (LSTM) network as an eminent type of machine learning method is generally considered to be capable on forecasting based on time series data including temporal dynamic behaviours with unknown lags. Various LSTM networks are adopted in existing researches to provide predictions in energy informatics field. However, the presented network structures are commonly selected through empirical or enumerative approaches. The utilized networks are generally carefully tuned as case by case studies. In this article, an evolutionary ensemble LSTM (EELSTM) method is proposed to pool LSTM networks with the same structure or with similar structures to obtain a more reliable prediction automatically. Experimental study demonstrates that networks with suitable structures and initialization are selected out through the learning process. A better performed peak demand prediction is achieved comparing with single LSTM unit network. In addition, the evolutionary parameters have variant impacts on the model performance.