准确预测出水水质对于污水处理厂的节能降耗具有重要意义。近年来,以废水处理仿真基准模型1号(BSM1)为代表的机理模型和各种深度学习算法被广泛运用于污水处理厂出水水质预测。然而,出水水质具有复杂的非线性关系,现有的预测模型通用性较差。基于此,提出一种基于变分模态分解(VMD)和4种深度学习算法的预测框架。通过变分模态分解方法将水质序列分解后,引入综合评价指标(CEI)为分解后的子序列寻求预测性能最好的算法,最后叠加各子模型的预测值得到最终的预测结果。以湖北省武汉市的一座污水处理厂出水化学需氧量(COD)浓度为例进行实例验证,结果表明,所提出的模型较单一模型在预测性能上达到了最佳效果,均方根误差(RMSE)达到了0.485。

[Objective] This study aims to propose a novel method for predicting effluent water quality in wastewater treatment plants, in order to enhance prediction accuracy and address the inadequate generalizability of existing models, thereby providing robust support for the operational optimization of wastewater treatment plants. [Methods] The proposed prediction framework primarily includes the following steps: First, the water quality sequence was decomposed into multiple subsequences with different characteristics using the variational mode decomposition (VMD) method. Subsequently, a comprehensive evaluation indicator (CEI) was introduced, based on which the deep learning algorithm with optimal prediction performance was selected for each decomposed subsequence. Four deep learning algorithms were involved in this study. Finally, the predicted values from each sub-model were aggregated to obtain the final effluent quality prediction. Taking the effluent chemical oxygen demand (COD) concentration of a wastewater treatment plant in Wuhan, Hubei Province as the research object, the proposed prediction framework was validated through a case study. The performance of the proposed framework was evaluated by comparing the prediction performance with that of single models. [Results] The effluent COD concentration data from a wastewater treatment plant in Wuhan were used for validation. The results showed that by decomposing the COD time series into different intrinsic mode functions (IMFs) using VMD, the complexity of the COD time series was effectively reduced. This provided simplified components for subsequent prediction, enabling the prediction model to better capture underlying patterns in the data and consequently improve prediction performance. Meanwhile, by introducing the CEI, four key evaluation indicators—mean absolute error (MAE), root mean square error (RMSE), standard deviation (STD), and mean absolute percentage error (MAPE)—were successfully integrated. This allowed for a comprehensive consideration of multi-dimensional error conditions when selecting the optimal prediction algorithm for each IMF subsequence, ensuring the comprehensiveness and accuracy of the selected algorithm. Finally, predictions were made for each different IMF based on the selected algorithm with optimal prediction performance. The results showed that this method effectively improved the overall model’s prediction accuracy, with the RMSE reaching 0.485. This confirmed that the proposed prediction framework achieved significant improvement in prediction performance compared to single models, providing strong support for accurate effluent water quality prediction in wastewater treatment plants. [Conclusions] The proposed water quality prediction framework based on VMD and multiple deep learning algorithms achieves high-precision prediction of effluent COD concentration in wastewater treatment plants by reasonably decomposing the water quality sequence and adaptively selecting prediction algorithms. The framework overcomes the limitations of existing single prediction models in handling complex nonlinear relationships, providing more accurate water quality predictions to support energy-saving and consumption-reduction decision-making in wastewater treatment plants. With significant practical value, it can be further extended in the future to predict other water quality indicators and be applied to wastewater treatment plants of different scales and types, thereby promoting intelligent operation and management in the wastewater treatment industry.