Hybrid Machine Learning-based Short-Term Electricity Price Forecasting in Smart Grids Using Weather, Demand, and Market Data: A Systemic Review

Abstract

Electricity Price Forecasting (EPF) represents one of the most important areas of activities of smart grid, as it directly affects the economic dispatch, demand responses, risk hedging, and reliability of the systems. Short term EPF is a difficult engineering problem due to the nonlinearity, volatility, and sensitivity to exogenous variables like weather and demand. This article provides a thoroughly structured and comprehensive analysis of hybrid Machine Learning (ML)-based methods of solving problems in short-term EPF and carries out the review of the literature published since 2015 and until 2025. Under the engineering view, the concept of review conceptualizes EPF as a data-driven forecasting pipeline which is composed of data acquisition, preprocessing, feature extraction, model training, and forecasting output. The paper studies the topic of hybrid architectures which combine signal decomposition methods (e.g., EMD, VMD, wavelets), sophisticated learning methods (e.g., CNN-LSTM, attention mechanisms, transformers), and ensemble policies including stacking and quantile regression averaging. These hybrid systems are evaluated on how well they can model temporal dependencies, the cross-feature interactions as well as extreme price behaviors. The review also speaks of formal evaluation practice based on conventional performance measures, such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) and probabilistic forecasting measures. The analysis of comparable results of benchmark datasets and real-world electricity markets indicates hybrid models tend to perform better than individual learners in both point and probabilistic forecasting tasks. But the difference in performances between market structures, forecasting horizons (inter and day ahead), and regime conditions through the use of strong validation procedures including rolling-origin analysis and leakage-free experimental design emerge. Moreover, the paper determines the main issues associated with the real-time implementation such as the complexity of the computation, scalability, and integration with the energy management system. It also points out the increased significance of uncertainty judgment by means of probabilistic forecasting methodology. Although this has been recently improved, the gaps in research have not been closed yet, such as cross-market generalization, transparency in what is available in the decision time and interest in evaluating operation value besides mainstream error measures. Lastly, the paper gives the future research paths on how to establish strong, interpretable and uncertainty aware hybrid EPF frameworks to increase the practical applicability of such models to the current smart grid setting.