Abstract:

Wastewater treatment plant is an important infrastructure for controlling water pollution and improving water environment. Establishing a reliable water quality assessment model is one of the research hotspots to timely feedback water quality status in the field of water environment. The Stacking theory，integrating the five algorithms（LSTM，BPNN，SVR，XGBoost，and KNN），was used to establish the prediction model of total nitrogen in effluent from a wastewater treatment plant. By analyzing the prediction effect of a single algorithm，the RMSE，MAE and R^{\mathrm{2}} of KNN are 1.10，0.995 and 0.567 respectively. It means that KNN has a lower fitting degree than other algorithms and can be used as the meta learner of the Stacking model. By comparing the difference of prediction performance among algorithms，LSTM，BPNN，SVR and XGBoost were selected as the base learners of the Stacking model. Compared with other models，the combination of LSTM，BPNN，SVR and XGBoost as the base learner and KNN as the meta learners has the highest prediction accuracy，and the RMSE，MAE and R^{\mathrm{2}} of this combination model were 1.01，0.782 and 0.702，respectively. In addition，compared with the prediction results of LSTM algorithm with the best prediction results in single algorithm，RMSE and MAE of this combination were reduced by 4.77% and 15.1%，and R^{\mathrm{2}} increased 10.9%，showing obvious prediction advantages.

Key words: wastewater treatment plant, water quality prediction, stacking integration algorithm

摘要：

污水处理厂是控制水体污染、改善水环境质量的重要基础设施，建立可靠的污水处理厂出水水质预测模型以便及时反馈污水处理状况是污水处理领域的研究热点之一。利用Stacking集成思想结合LSTM、BPNN、SVR、XGBoost、KNN 5种算法建立了污水处理厂出水总氮的预测模型。分析单一算法预测效果发现，KNN算法与其他算法相比拟合程度偏低，RMSE、MAE和 R^{\mathrm{2}}分别为1.10、0.995和0.567，可作为Stacking模型的元学习器。通过算法间预测性能差异的比较，选择LSTM、BPNN、SVR、XGBoost算法作为基学习器。与其他基、元学习器的组合相比，以LSTM、BPNN、SVR、XGBoost为基学习器，KNN为元学习器的组合方式预测准确度最高，RMSE、MAE和 R^{\mathrm{2}}分别为1.01、0.782和0.702。与使用单一算法中预测结果最好的LSTM算法的预测结果相比，该组合的RMSE、MAE分别降低了4.77%和15.1%， R^{\mathrm{2}}提升了10.9%，具有明显的预测优势。

关键词: 污水处理厂, 水质预测, Stacking集成算法

CLC Number: 

• X703