Abstract: As traditional methods are difficult to extract multi-source hydrological features and solve the problem of feature redundancy,a monthly runoff prediction model based on improved Temporal-Convolutional-Network (TCN) and Long Short Term Memory (LSTM) was proposed.The model constructed a multi-convolution kernel parallel network to extract multi-source timing features while maintaining the causal convolution characteristics.Dilated convolution was introduced to extract higher order hydrological features,improving the processing efficiency of memory units within a long period.The introduction of residual links enabled the complete features of the bottom stage to be transmitted across stages,which enriched the feature results and optimized the overall network structure,and Lushui River Basin was taken as an example for verification.The experimental results showed that the model was superior to other comparative models in terms of computational efficiency,accuracy and network structure,which verified the effectiveness of the model in hydrological prediction of the basin.

Key words: runoff forecast model, time convolutional neural network, long and short term memory neural network, multi-source hydrological data

摘要： 为提升水文模型预测精度和计算效率,解决传统方法难以浓缩多源水文特征,以及长时间跨度下数据量纲、值域不同造成的特征冗杂问题,提出基于改进时间卷积网络(TCN)和长短时记忆网络(LSTM)的月径流量预测模型。通过构造多卷积核并行网,以提取多源时序特征,并保持原因果卷积特性。引入扩张卷积抽取高阶水文特征,提升长时间跨度记忆单元处理效率。利用残差链接方式跨层传输底层完整特征,丰富特征结果,同时优化整体网络学习过程,并以泸水河流域为例进行验证。实验结果表明,该模型在计算效率、精度、网络结构上均优于其他对比模型,从而验证了其在该流域水文预测的有效性。

关键词: 径流预测模型, 时间卷积神经网络, 长短时记忆神经网络, 多源水文数据