基于自适应变分模态分解和长短期记忆网络的月径流预报

准确可靠的月径流预报是流域水旱灾害防治及水资源合理配置的重要依据。原始径流时间序列包含多种频率成分,将时间序列数据分解预处理技术和机器学习模型相结合的混合模型已被用于捕捉径流动态过程。然而,将数据分解技术直接应用于整个时间序列是一种不切实际的方法,会导致部分信息从测试阶段传输到模型的训练过程中。为此,设计了一个用观测数据更新历史样本的自适应动态分解策略,提出基于自适应变分模态分解和长短期记忆网络的分解-预测-集成月径流预测混合模型。首先,采用自适应分解策略对径流时序数据进行变分模态分解,得到不同频率成分的子序列;其次,为每个分解子序列构建长短期记忆神经网络径流预测模型,并采用贝叶斯优化算法优选模型超参数;然后,将子序列的预测结果集成得到径流的最终预测结果;最后,以金沙江上游石鼓水文站月径流预报为研究实例,对比传统的分解策略(“捆绑分解”)和分解方法(离散小波变换和集成经验模态分解),验证所提混合模型的有效性和可行性。结果表明,所提混合模型在数据分解预处理中避免了引入未来信息,并能够进一步提升径流预报精度。

Monthly runoff prediction based on self-adaptive variational mode decomposition and long short-term memory network

Accurate and reliable monthly runoff prediction is an important basis for rational allocation of water resources.The original runoff time series contains a variety of frequency components, and a hybrid model combining time series data decomposition technology and machine learning model has been used to capture the runoff dynamic process.However, it is an impractical method to apply the data decomposition technology directly to the whole time series, which will cause some information to be transferred from the test period to the training process of the model.Therefore, a self-adaptive dynamic decomposition strategy is adopted to update the historical samples with the observed data, and a decomposition-prediction- integration hybrid model for monthly runoff prediction based on variational mode decomposition and long short-term memory network is proposed.First, a self-adaptive dynamic decomposition strategy is adopted to decompose the variational modes of runoff time series data.Second, long short-term memory neural network models based on Bayesian optimization are constructed to identify the input-output relationship hidden in each mode or subcomponent.Then, the final prediction result of runoff is obtained by integrating the prediction results of submodules.Finally, taking the monthly runoff prediction of the Shigu hydrological station in the upper reaches of the Jinsha River as an example, the effectiveness and feasibility of the proposed hybrid model are verified by comparing with the traditional decomposition strategy("bundled decomposition") and decomposition methods(discrete wavelet transform and ensemble empirical mode decomposition).The results show that the proposed model avoids the use of future information in the processing of data decomposition, and can further improve the accuracy of runoff prediction.