Umme Mumtahina

Load forecasting is an integral part of the power industries. Load-forecasting techniques should minimize the percentage error while prediction future demand. This will inherently help utilities have an uninterrupted power supply. In addition to that, accurate load forecasting can result in saving large amounts of money. This article provides a systematic review based on the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) framework. This article presents a complete framework for short-term load forecasting using metaheuristic algorithms. This framework consists of three sub-layers: the data-decomposition layer, the forecasting layer, and the optimization layer. The data-decomposition layer decomposes the input data series to extract important features. The forecasting layer is used to predict the result, which involves different statistical and machine-learning models. The optimization layer optimizes the parameters of forecasting methods to improve the accuracy and stability of the forecasting model using different metaheuristic algorithms. Single models from the forecasting layer can predict the results. However, they come with their limitations, such as low accuracy, high computational burden, stuck to local minima, etc. To improve the prediction accuracy, the hyperparameters of these models need to be tuned properly. Metaheuristic algorithms cab be used to tune these hyperparameters considering their interdependencies. Hybrid models combining the three-layer methods can perform better by overcoming the issues of premature convergence and trapping into a local minimum solution. A quantitative analysis of different metaheuristic algorithms and deep-learning forecasting methods is presented. Some of the most common evaluation indices that are used to evaluate the performance of the forecasting models are discussed. Furthermore, a taxonomy of different state-of-the-art articles is provided, discussing their advantages, limitations, contributions, and evaluation indices. A future direction is provided for researchers to deal with hyperparameter tuning.