基于长短时记忆改进的Informer模型的污水COD与总氮预测

Prediction of wastewater COD and total nitrogen based on Informer model improved by Long Short-Term Memory

摘要

【目的】监测和管理污水处理厂的出水水质指标对于环境保护至关重要。然而,污水处理具有高度可变性以及难以预测的复杂生化过程。【方法】通过建立一个基于长短时记忆网络(Long-Short Term Memory,LSTM)-Informer的污水处理厂出水COD和总氮预测模型,首先采用长短时记忆模块提取进水水质指标中的时间依赖关系,然后将提取的时间特征融入Informer网络模块中解决长期依赖的问题。并通过对几种典型模型(Convolutional Neural Network,CNN、LSTM、CNN-LSTM、Informer)性能评估指标的分析,比较了各个模型的拟合效果和预测效果。【结果】试验结果表明:LSTM-Informer模型能够准确预测出水COD、TN浓度,其在试验数据集的预测值和实际值的均方误差分别为0.630mg·L^-1和0.105 mg·L^-1,均方根误差分别为0.79 mg·L^-1和0.32 mg·L^-1,平均绝对误差分别为0.60 mg·L^-1和0.24 mg·L^-1,平均绝对百分比误差分别为2.16%和2.26%,决定系数R²分别为0.97和0.90,各项指标均优于几种典型神经网络模型;在统计学结果显著性检验中,验证在显著性水平为0.05时,p值均大于其余对比模型。【结论】结果表明:基于长短时记忆网络捕获短期时间依赖于Informer的多头稀疏自注意力获取长期时序特征所建立的LSTM-Informer模型为污水出水水质提供了长期监测和预测保证,其准确性在评价指标与统计学检验中都取得了一定优势,综合来讲可以为污水处理厂提前预警水质变化,对实现智能化控制和优化处理模式具有重要意义。

Abstract

[Objective] Monitoring and managing effluent quality indicators of wastewater treatment plants is critical for environmental protection. However, wastewater treatment involves highly variable and unpredictable complex biochemical processes. [Methods] A prediction model for effluent COD and total nitrogen in wastewater treatment plants was established based on a Long Short-Term Memory(LSTM)-Informer. The LSTM module was first used to extract temporal dependencies from the influent quality indicators. Subsequently, the extracted temporal features were integrated into the Informer module to address the issue of long-term dependency. Additionally, through the analysis of performance evaluation indicators of several typical models(Convolutional Neural Network, CNN, LSTM, CNN-LSTM, and Informer), the fitting performance and prediction performance of each model were compared. [Results] The experimental result showed that the LSTM-Informer model could accurately predict the effluent COD and TN concentrations. On the experimental dataset, the mean squared errors between the predicted and actual values were 0. 630 mg · L^-1 and 0. 105 mg · L^-1, respectively. The root mean square errors were 0. 79 mg·L^-1 and 0. 32 mg · L^-1, the mean absolute errors were 0. 60 mg · L^-1 and 0. 24 mg · L^-1, the mean absolute percentage errors were 2. 16% and 2. 26%, and the coefficients of determination(R²) were 0. 97 and 0. 90, respectively. All these indicators outperformed those of several typical neural network models. In the statistical significance test, it was verified that the p-values of the LSTM-Informer model were all greater than those of the other comparative models at a significance level of 0. 05. [Conclusion] The result demonstrate that the LSTM-Informer model, which captures short-term temporal dependencies via the LSTM network and extracts long-term temporal features through the multi-head sparse self-attention mechanism of the Informer, provides a reliable basis for the long-term monitoring and prediction of wastewater effluent quality. Its accuracy demonstrates advantages in both evaluation indicators and statistical tests. Overall, this model can provide early warning of water quality changes for wastewater treatment plants, which is crucial for achieving intelligent control and optimizing treatment processes.

关键词

污水处理 / 神经网络 / LSTM-Informer模型 / COD预测 / 总氮预测 / 注意力机制 / 长时序预测 / 水质预警

Key words

wastewater treatment / neural network / LSTM-Informer model / COD prediction / total nitrogen prediction / attention mechanism / long-term time series prediction / water quality early warning

基金资助

国家重点研发计划(2018YB1700200)

2023 年度辽宁省研究生教育教学改革研究项目(2023-132)