基于CNN-LSTM-Attention和自回归的混合水位预测模型  

A hybrid water level forecasting model based on CNN-LSTM-Attention with Autoregressive

作　　者：吕海峰 涂井先 林泓全 冀肖榆[1] LYU Haifeng;TU Jingxian;LIN Hongquan;JI Xiaoyu(Guangxi Key Laboratory of Machine Vision and Intelligent Control,WuZhou University,Wuzhou 543002,Guangxi,China)

机构地区：[1]梧州学院广西机器视觉与智能控制重点实验室,广西梧州543002

出　　处：《水利水电技术（中英文）》2024年第6期16-31,共16页Water Resources and Hydropower Engineering

基　　金：国家自然科学基金项目(62262059);广西高校中青年教师科研基础能力提升项目(2024KY0692)。

摘　　要：【目的】水位预测对交通运输、农业以及防洪措施具有重要影响。精确的水位值可用于提升水道运输的安全及效率、降低洪水风险,同时也是保障区域可持续发展的必要条件。【方法】提出一种CRANet的混合水位预测模型,以卷积神经网络(CNN)、长短期记忆网络(LSTM)、注意力机制以及自回归(AR)组件为基础,旨在应对时间序列数据中存在的线性与非线性问题,缓解自回归及ARIMA模型的缺陷。其应用不仅在于为航运调度提供决策支撑,加强导航安全效率,同样能提升防洪减灾的能力。其中,CNN和LSTM组件有效地针对数据集内的局部和全局关系进行捕捉,AR组件则能充分考虑数据的时间序列特性。同时,通过注意力机制,模型能够优先考虑相关特性,提高预测效果。【结果】研究成果所提出的模型已成功应用于中国西江梧州站的水位预测,在测试集上预测未来3 h级别水位的MAE、RMSE和R^(2)分别为0.086、0.114 5和0.950 8。【结论】结果表明所提出的CRANet模型在水位预测方面的高可用性、准确度与稳健性,相较于AR、SVR、CNN、LSTM等模型具有更优的MAE、RMSE和R^(2)。[Objective]Water level forecasting in multivariate time series is crucial for various applications such as transportation,agriculture,and flood control.Accurate prediction of water levels in the Xijiang River is essential for enhancing the safety and efficiency of waterway transportation,reducing flood risks,and promoting sustainable development in the region.However,water level forecasting involves a combination of linear and nonlinear problems,which traditional method like autoregressive and ARIMA models may struggle to handle effectively.[Methods]To address this challenge,a novel hybrid water level forecasting model called the Convolutional Recurrent Attention Autoregressive network(CRANet)is proposed.The strengths of Convolutional Neural Network(CNN),Long Short-Term Memory(LSTM),Attention mechanism,and Autoregressive(AR)component are combined by CRANet.By integrating these components,both local and global dependencies within the water level dataset are efficiently captured by CRANet.Spatial and temporal patterns are excellently captured by the CNN and LSTM components,while the time-series nature of the data is accounted for by the AR component.Furthermore,the model′s ability to prioritize relevant features is enhanced by the attention mechanism,leading to further improvements in its forecasting performance.[Results]The proposed CRANet model has been successfully applied to water level forecasting at Wuzhou Station in the Xijiang River,China.On the test set,the MAE,RMSE,and R^(2) for forecasting future water levels at a 3-hour interval are observed to be 0.086,0.1145,and 0.9508,respectively.[Conclusion]The result indicate that the proposed CRANet model demonstrates high availability,accuracy,and robustness in water level forecasting,exhibiting superior MAE,RMSE,and R^(2) compared to other baseline models such as AR,SVR,CNN,LSTM and et al.

关 键 词：时间序列 水位预测 CNN LSTM ATTENTION 影响因素 洪水 西江 

分 类 号：P339[天文地球—水文科学]