摘要
提出了一种将Transformer与卷积神经网络(CNN)相结合的城区地下水位预测模型。Transformer模型能够提取地下水位在时间序列上包含的关键信息,有效提升了模型的长时间预测能力;CNN能获取相邻监测站点地下水位数据之间的空间关联信息,使信息的提取更加丰富。使用开源地下水位数据集对模型进行训练,并进行仿真验证。仿真结果表明,在预测未来12个时刻的地下水位值时,CNN-Transformer模型预测结果整体的均方根误差值相比于循环神经网络(RNN)系列模型从0.2507米降到0.1427米,在未来第12个时刻的均方根误差也仅为0.2309米,验证了上述模型能实现长时间高精度的地下水位预测。
A method based on Transformer and Convolutional Neural Network(CNN)is proposed for urban groundwater level prediction.The key information can be extracted from the time series of the groundwater level by the Transformer model,thus improving the long-term predictive ability of the model.The spatial correlated information of the groundwater levels at adjacent monitoring stations can be obtained by CNN,which enriches the extracted information.We use the open-source groundwater level data set to train the model and perform simulation verification.The simulation results show that the overall root mean square error value of the predicted groundwater level at the next 12 consecutive moments by the CNN-Transformer model is reduced from 0.2507 m to 0.1427 m compared to that by the Recurrent Neural Network(RNN)series model,and the root mean square error at the 12th moment in the future is only 0.2309 m.The result indicates that the CNN-Transformer groundwater level prediction model can realize longterm and high-precision groundwater level prediction.
作者
冯鹏宇
金韬
沈一选
但俊
FENG Peng-yu;JIN Tao;SHEN Yi-xuan;DAN Jun(College of Information Science and Electronic Engineering,Zhejiang University,Hangzhou Zhejiang 310027,China)
出处
《计算机仿真》
北大核心
2023年第4期492-498,共7页
Computer Simulation
基金
国家自然科学基金项目(61675180)
企业合作项目“液位监测系统研发”(校合-2021-KYY-546001-0003)。
关键词
地下水位预测
深度时序模型
卷积神经网络
Groundwater level prediction
Deep time series model
Convolutional neural networks
