基于CNN-LSTM-ARIMA的超短期风速预测  

Ultra-Short-Term Wind Speed Forecasting Based on CNN-LSTM-ARIMA

作　　者：王世明[1] 张少童 娄嘉奕 WANG Shiming;ZHANG Shaotong;LOU Jiayi(College of Engineering Science and Technology,Shanghai Ocean University,Shanghai 201306,China;Suzhou Jintai Ocean Engineering Research Co.Ltd.,Changshu 215500,Jiangsu,China)

机构地区：[1]上海海洋大学工程学院,上海201306 [2]苏州市津泰海洋工程研究有限公司,江苏常熟215500

出　　处：《新能源进展》2024年第6期688-695,共8页Advances in New and Renewable Energy

基　　金：国家自然科学基金项目(41976194);上海市“科技创新行动计划”软科学研究项目(23692102600);上海市“科技创新行动计划”上海工程技术研究中心项目(19DZ2254800)。

摘　　要：提升风速预测的精准度对于实时调整电力系统的管理策略及增强风电市场的竞争实力有着关键作用。提出一种基于卷积神经网络(CNN)、长短期记忆网络(LSTM)和自回归集成移动平均(ARIMA)模型的超短期风速预测方法,通过CNN卷积层捕捉时间序列数据中的模式和局部特征,利用LSTM模型对提取的特征进行学习训练,基于CNN-LSTM组合架构模型,预测未来风速并对比实际数据获得残差值,最终利用ARIMA分析历史残差来修正未来的预测误差值,实现对风速的超短期预测。以土耳其某个风电场的实际风速记录为基础,对未来10min的风速进行预测。结果表明,与CNN-LSTM、双层LSTM传统神经网络模型相比,CNN-LSTM-ARIMA模型对风速预测结果的平均绝对误差分别下降了16.40%、26.92%,能显著提高预测精度。Improving wind speed prediction accuracy is important for timely adjustments in power system scheduling plans and enhancing competitiveness in the wind energy market.This paper presents an ultra-short-term wind speed forecasting method based on an ensemble of convolutional neural network(CNN),long short-term memory network(LSTM),and autoregressive integrated moving average(ARIMA).Firstly,CNN convolutional layers capture patterns and local features in time series data.Subsequently,it utilizes LSTM models to learn and train on the extracted features.Based on the CNN-LSTM composite architecture model,it predicts future wind speeds and compares them with actual data to obtain residuals.Finally,it employs ARIMA to analyze historical residuals to correct future prediction errors,achieving ultra-short-term wind speed forecasting.Using measured wind speed data from a wind farm in Turkey as an example,it predicts wind speeds for the next 10 minutes.The results indicate that compared to traditional neural network models like CNN-LSTM and LSTM-LSTM,the CNN-LSTM-ARIMA model has reduced the mean absolute error in wind speed forecasting by 16.40%and 26.92%,respectively,significantly enhancing the prediction accuracy.

关 键 词：风速预测 卷积神经网络 长短期记忆网络 自回归集成移动平均模型 

分 类 号：TK81[动力工程及工程热物理—流体机械及工程]