A deep bi-directional long-short term memory neural network-based methodology to enhance short-term electricity load forecasting for residential applications

Abstract

Unexpected fluctuations associated with electricity load consumption patterns pose a significant threat to the stability, efficiency, and sustainability of modernized energy systems. Therefore, there is an eminent need for sophisticated Short-Term Load Forecasting (STLF) models to mitigate the impact of these uncertainties. In this paper, a novel methodology that aims to enhance the prediction accuracy of the STLF model is developed, tested, implemented, and investigated. The proposed methodology simultaneously considers optimizing both the input feature sets and the prediction methods. The results indicate that the proposed deep bidirectional long short-term memory neural network-based approach improves the prediction accuracy by nearly 95% in comparison with various competitive benchmarks which focus only on the prediction algorithm. This improvement can be attributed to the significant effect of considering both the input feature set and the learning-based model hyperparameters optimization instead of the traditional practice focusing only on the prediction algorithm.