Univariate Individual Household Energy Forecasting by Tuned Long Short-Term Memory Network

Abstract

Accurate forecast of energy consumption has proven to be invaluable to many nodes of the energy sector, enabling an efficient and cost-effective distribution of energy among consumers. However, the nonlinear and non-stationary nature of household energy usage challenges contemporary machine learning algorithms, where there is potential to yet develop robust and dependable technologies. In this work presents a novel, hybrid long short-term memory (LSTM) model tuned by Best Guided Search-Arithmetic Optimization Algorithm (BGS-AOA), that employs quasi-reflection-based learning (QRL) to overcome the exploration–exploitation imbalance of the original AOA. The introduced approach was tested on the publicly available dataset capturing energy consumption of individual London households. Due to the stochastic nature of optimizers, all experiments were carried out over the course of five independent runs. In order to evaluate metaheuristics solutions, the total MSE was utilized as the objective function for 3-step ahead forecast. Results of the experiments demonstrated superior performance of our model when compared to other metaheuristics frequently encountered in literature (ABC, FA, SSA, ChOA).