基于VMD-CNN-LSTM模型和迁移学习框架的风暴潮预测研究

Storm Surge Prediction Based On VMD-CNN-LSTM Model and Transfer Learning Framework

针对数值风暴潮预测模型耗时长及可操作性差的问题,基于“分解-预测-集成”的思想,融合变分模态分解算法(VMD)、“点-窗”采样模型、卷积神经网络(CNN)、长短期记忆网络(LSTM)及数值集成方法,构建了风暴潮组合预测模型。首先,采用VMD分解风暴潮增水时序数据,得到多个分量;其次,利用“点-窗”采样模型对分量数据进行采样,构建输入矩阵;之后,将输入矩阵输入到CNN-LSTM组合预测框架中进行预测;最后,集成各分量预测结果,得到风暴潮最终预测结果。实证结果表明,VMD-CNN-LSTM预测模型与目前广泛采用的单一模型和其他组合预测模型相比,具有更高的预测精度。为了将模型更好地运用于实际工程中,采用迁移学习方法,将大数据训练模型“迁移”到小数据领域,结果表明:即使新样本数据量较少,迁移后的模型仍具有较好的泛化能力。

Aiming at the problems of long time and poor operation of the numerical storm surge prediction model,a combined prediction model of storm surge was constructed based on the idea of“decomposition-prediction-integration”,combining variational mode decomposition algorithm(VMD),“point-window”sampling model,convolutional neural network(CNN),long and short-term memory network(LSTM)and numerical integration method.To begin with,VMD is used to decompose the storm surge time series data to obtain multiple components;then,a“point-window”sampling model is used to sample the component data and construct the input matrix;subsequently,the input matrix is input to the CNN-LSTM combined prediction framework for prediction;ultimately,the prediction results of each component are integrated to obtain the final storm surge prediction results.The empirical results show that the VMD-CNN-LSTM prediction model has higher prediction accuracy compared with the widely used single model and other combined prediction models.In order to better apply the model to practical engineering,the transfer learning method is used to“transfer”the large data training model to the small data domain,and the results show that the transferred model has better generalization ability even if the amount of new sample data is small.