基于双层数据分解混合模型预测鄱阳湖COD

化学需氧量（Chemical Oxygen Demand, COD）是衡量水质状况的最重要参数之一，反映水体受还原性物质污染的程度。该研究采用改进的完全集合经验模式分解（Improved Complete Ensemble Empirical Mode Decomposition With Adaptive Noise, ICEEMDAN）、变分模式分解（Variational Mode Decomposition, VMD）相结合的双层数据分解算法，并利用双向长短期记忆（Bidirectional Long Short-term Memory, BLSTM）神经网络，提出了一种混合模型IVB（Improved Complete Ensemble Empirical Mode Decomposition With Adaptive Noise-Variational Mode Decomposition - Bidirectional Long Short-term Memory），并以鄱阳湖高锰酸盐指数（Permanganate index, CODMn）监测数据为研究对象，进行案例研究。结果表明，IVB模型具有良好的预测性能：1 d以后的CODMn预测中，IVB模型的平均绝对百分比误差为2.21%，与单一BLSTM神经网络模型相比降低了10.57个百分点，而与IB (Improved Complete Ensemble Empirical Mode Decomposition With Adaptive Noise - Bidirectional Long Short-term Memory)模型相比降低了4.62个百分点；7 d以后的CODMn预测中，IVB模型的平均绝对百分比误差为8.18%，与单一BLSTM神经网络模型相比降低了16.34个百分点，而与IB模型相比降低了4.68个百分点。这项研究表明，所开发的IVB模型可以用作水资源管理的有效分析与决策工具。

COD forecasting of Poyang lake using a novel hybrid model based on two-layer data decomposition

Poyang Lake is the largest freshwater lake in China. However, the ecosystem around the Poyang Lake has been threatened by water pollution in recent years. The chemical oxygen demand (COD) has been one of the most indicative parameters to evaluate the water quality, indicating the degree of water pollution from the organics and reductants in environmental chemistry. Generally, the high accuracy COD refers to the amount of oxygen that can be consumed by reactions in a measured solution at monitoring stations. But, it is still lacking on the predict ability of water quality in advance. Furthermore, the water body has been polluted for the subsequent treatment, due to the current or overdue data from the water quality monitoring stations. An early warning of water pollution is a high demand before the pollution occurs. An accurate and rapid COD prediction of water quality still remains a challenge, due to the high dynamic characteristics in a short time, indicating the unstable prediction performance for the time series with many peak points. In this study, a two-layer decomposition approach was employed to combine the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), variation mode decomposition (VMD), and bidirectional long short-term memory (BLSTM) neural network for the decomposed subseries prediction, in order to develop a new hybrid model ICEEMDAN-VMD-BLSTM (IVB). First, the ICEEMDAN model was used to decompose the original COD time series into the several components, and then the VMD model was utilized to decompose the component with the highest frequency during data processing. Second, the BLSTM neural network was used to predict each component. Last, all forecasted components were reconstructed to obtain the final COD forecast value. A case study was conducted using CODMn monitoring data from August 1, 2017 to April 30, 2020 at Poyang Lake. A hybrid model was proposed to predict the CODMn time series after data processing. In addition, several competitor models were also used to compare with the proposed hybrid model. Experiment result shows that the IVB model presented a high consistency between the predicted and actual values, indicating the better forecast performance than the rest. The mean absolute percentage errors (MAPE) were 2.21%, and 8.18%, respectively, for the 1 and 7 d ahead prediction using the IVB model. Especially, the MAPEs in the IVB model were reduced by 10.57 percentage point and 4.62 percentage point for 1 d ahead prediction, while 16.34 percentage point and 4.68 percentage point for 7 d ahead prediction, compared with the BLSTM and IB model. In the case of unstable data with the rapid changing points, the IVB model also showed a relatively stable performance, indicating more stable in extreme cases. Consequently, the IVB model can be expected to serve as a promising new forecast model for the efficient decision-making tool in water resource management.