Machine Learning-Based Approaches for Electricity Demand Forecasting: A Systematic Review

Abstract

Electricity load forecasting is critical for efficient power system planning, network stability and informed policy development. With power systems becoming increasingly complex due to integration of renewables, time-varying tariffs and changing demand patterns, traditional statistical approaches are insufficient to model the nonlinear multi-seasonal dynamics of the modern load series. This work offers an in-depth critical review of the application of machine learning (ML) and deep learning (DL) models for electricity demand forecasting, with a particular emphasis on research published from 2018 to 2024 in IEEE Transactions, Applied Energy, Energy Reports and other journals indexed in Scopus. Techniques such as supervised machine learning (ML) algorithms (Support Vector Regression, Random Forest, XGBoost and LightGBM) and deep learning (DL) networks (Long Short-Term Memory, Gated Recurrent Units, Convolutional Neural Networks and Transformers) are evaluated. Ensembles of signal decomposition (Variational Mode Decomposition, Wavelet), deep recurrent networks and attention mechanisms are demonstrated to consistently deliver Mean Absolute Percentage Errors (MAPE) of less than 1.5% on standard datasets. The review critically evaluates the approaches to feature engineering, evaluation and benchmarking in 40 representative papers. Key research opportunities are highlighted: the almost complete lack of dedicated, validated forecasting models for Indian states like Maharashtra; the inadequate incorporation of macroeconomic factors; and the need for Explainable AI (XAI) models in operational energy forecasting. Potential future developments such as IoT-based smart metering, Temporal Fusion Transformers and online learning are considered.