Alazemi Talal, Darwish Mohamed, Radi Mohammed
Brunel University London Kingston Lane Uxbridge, Middlesex, UB8 3PH, United Kingdom.
UK Power Networks, Pocock House, 237 Southwark Bridge Rd, London, SE1 6NP, United Kingdom.
Heliyon. 2024 Feb 14;10(4):e26088. doi: 10.1016/j.heliyon.2024.e26088. eCollection 2024 Feb 29.
The use of renewable energy sources (RESs) at the distribution level has become increasingly appealing in terms of costs and technology, expecting a massive diffusion in the near future and placing several challenges to the power grid. Since RESs depend on stochastic energy sources -solar radiation, temperature and wind speed, among others- they introduce a high level of uncertainty to the grid, leading to power imbalance and deteriorating the network stability. In this scenario, managing and forecasting RES uncertainty is vital to successfully integrate them into the power grids. Traditionally, physical- and statistical-based models have been used to predict RES power outputs. Nevertheless, the former are computationally expensive since they rely on solving complex mathematical models of the atmospheric dynamics, whereas the latter usually consider linear models, preventing them from addressing challenging forecasting scenarios. In recent years, the advances in machine learning techniques, which can learn from historical data, allowing the analysis of large-scale datasets either under non-uniform characteristics or noisy data, have provided researchers with powerful data-driven tools that can outperform traditional methods. In this paper, a systematic literature review is conducted to identify the most widely used machine learning-based approaches to forecast RES power outputs. The results show that deep artificial neural networks, especially long-short term memory networks, which can accurately model the autoregressive nature of RES power output, and ensemble strategies, which allow successfully handling large amounts of highly fluctuating data, are the best suited ones. In addition, the most promising results of integrating the forecasted output into decision-making problems, such as unit commitment, to address economic, operational and managerial grid challenges are discussed, and solid directions for future research are provided.
从成本和技术角度来看，在配电网层面使用可再生能源（RESs）变得越来越具有吸引力，预计在不久的将来会大规模普及，并给电网带来诸多挑战。由于可再生能源依赖于随机能源，如太阳辐射、温度和风速等，它们给电网带来了高度的不确定性，导致功率不平衡并恶化电网稳定性。在这种情况下，管理和预测可再生能源的不确定性对于将它们成功整合到电网中至关重要。传统上，基于物理和统计的模型已被用于预测可再生能源的功率输出。然而，前者计算成本高昂，因为它们依赖于求解大气动力学的复杂数学模型，而后者通常考虑线性模型，这使得它们无法应对具有挑战性的预测场景。近年来，机器学习技术取得了进展，这些技术可以从历史数据中学习，能够分析具有非均匀特征或噪声数据的大规模数据集，为研究人员提供了强大的数据驱动工具，其性能可以超越传统方法。本文进行了系统的文献综述，以确定预测可再生能源功率输出最广泛使用的基于机器学习的方法。结果表明，深度人工神经网络，特别是能够准确模拟可再生能源功率输出自回归特性的长短期记忆网络，以及能够成功处理大量高度波动数据的集成策略，是最适合的方法。此外，还讨论了将预测输出整合到诸如机组组合等决策问题中以应对经济、运营和管理电网挑战的最有前景的结果，并提供了未来研究的坚实方向。
