Short-term offshore wind speed forecast by seasonal ARIMA-A comparison against GRU and LSTM
Short-term offshore wind speed forecast by seasonal ARIMA-A comparison against GRU and LSTM
复制标题
DOI：
10.1016/j.energy.2021.120492
复制
发表时间：
2021-04-07
期刊：
影响因子：
9
通讯作者：
Feng, Ziming
中科院分区：
文献类型：
--
作者：
Liu, Xiaolei;Lin, Zi;Feng, Ziming
海上风力发电是世界上发展最快的能源之一，既环保又具有经济竞争力。短期时间序列风速预测对于正确和有效的海上风能评估非常重要，反过来又有利于风电场所有者，电网运营商以及最终客户。在这项研究中，提出了一个季节性自回归综合移动平均（SARIMA）模型来预测苏格兰沿海/近海地区每小时测量的风速。所使用的数据集包括三个风速时间序列，这些数据是在不同海拔高度从沿海桅杆上收集的，该桅杆是为示范海上风力涡轮机设计的。为了验证SARIMA的性能，将开发的预测模型与新开发的基于深度学习的门控循环单元（GRU）和长短期记忆（LSTM）算法进行了进一步的比较。尽管最近计算能力的发展已经引发了更先进的机器学习算法，但所提出的SARIMA模型在预测未来海上风速随时间序列滞后的准确性方面表现出色。在基于相应数据集和评估的预测范围的所有三种测试预测模型中，SARIMA模型提供了最高的准确性和健壮性。(c) 2021 Elsevier Ltd.版权所有。
Offshore wind power is one of the fastest-growing energy sources worldwide, which is environmentally friendly and economically competitive. Short-term time series wind speed forecasts are extremely sig-nificant for proper and efficient offshore wind energy evaluation and in turn, benefit wind farm owner, grid operators as well as end customers. In this study, a Seasonal Auto-Regression Integrated Moving Average (SARIMA) model is proposed to predict hourly-measured wind speeds in the coastal/offshore area of Scotland. The used datasets consist of three wind speed time series collected at different ele-vations from a coastal met mast, which was designed to serve for a demonstration offshore wind turbine. To verify SARIMA's performance, the developed predictive model was further compared with the newly developed deep-learning-based algorithms of Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM). Regardless of the recent development of computational power has triggered more advanced machine learning algorithms, the proposed SARIMA model has shown its outperformance in the accu-racy of forecasting future lags of offshore wind speeds along with time series. The SARIMA model pro-vided the highest accuracy and robust healthiness among all the three tested predictive models based on corresponding datasets and assessed forecasting horizons.(c) 2021 Elsevier Ltd. All rights reserved.