Energy price prediction using data-driven models: A decade review

The accurate prediction of energy price is critical to the energy market orientation, and it can provide a reference for policymakers and market participants. In practice, energy prices are affected by external factors, and their accurate prediction is challenging. This paper provides a systematic decade review of data-driven models for energy price prediction. Energy prices include four types: natural gas, crude oil, electricity, and carbon. Through the screening, 171 publications are reviewed in detail from the aspects of the basic model, the data cleaning method, and optimizer. Publishing time, model structure, prediction accuracy, prediction horizon, and input variables for energy price prediction are discussed. The main contributions and findings of this paper are as follows: (1) basic prediction models for energy price, data cleaning methods, and optimizers are classified and described; (2) the structure of the prediction model is finely classified, and it is inferred that the hybrid model and prediction architecture with multiple techniques are the focus of research and the development direction in the future; (3) root mean square error, mean absolute percentage error, and mean absolute error are the three most frequently used error indicators, and the maximum mean absolute percentage error is less than 0.2; (4) the ranges of data size and data division ratio for energy price prediction in different horizons are given, the proportion of the test set is usually in the range of 0.05-0.35; (5) the input variables for energy price prediction are summarized; (6) the data cleaning method has a more significant role in improving the accuracy of energy price prediction than the optimizer. (C) 2020 Elsevier Inc. All rights reserved.

Automated summary

This paper provides a systematic decade review of data-driven models for energy price prediction, discussing different types of energy price prediction models, data cleaning methods, model structure, and accuracy. Key findings include the classification of basic prediction models, the inference of the future research direction towards hybrid models and prediction architectures with multiple techniques, and the significance of data cleaning methods in improving prediction accuracy.