基于时序-空间双流卷积神经网络的炼化污水厂进水COD预测模型

doi:10.11949/0438-1157.20251151

摘要/Abstract

摘要：

炼化污水厂进水化学需氧量（COD）波动幅度大，影响污水处理工艺的稳定运行和水质达标，但现有预测方法难以整合其时序及空间关联特征，无法实现高效预测。本研究提出了一种时序-空间双流卷积神经网络（TS-DSCNN）模型。该模型通过并行双流架构，分别利用一维卷积核挖掘COD时序规律，并挖掘瞬时水质参数电导率（EC）、氨氮（NH₃-N）、悬浮固体（SS）与COD的关联关系，引入注意力机制自适应融合双流特征。实验结果表明，TS-DSCNN模型测试集决定系数（R²）达0.89，均方根误差（RMSE）为16 mg/L，性能显著优于长短时记忆网络（R² = 0.81、RMSE = 20 mg/L）、卷积神经网络（R² = 0.78、RMSE = 19 mg/L）等基准模型。消融实验结合沙普利可加解释（SHAP）验证了时序流的主导作用与双流架构的有效性，时序-空间特征深度融合（R² = 0.89）优于单一空间流（R² = 0.75）、单一时序流（R² = 0.81）及简单特征拼接模型（R² = 0.84）。SHAP分析与注意力权重分布（时序流权重集中于0.8~1.0，空间流权重集中于0.0~0.2）相互印证，明确时序流为预测主导因素、空间流为辅助因素。30天现场验证表明模型可有效跟踪COD波动趋势，多数时段平均绝对百分比误差低于10%，仅在水质剧烈波动时存在局部误差，整体稳定性与泛化性良好。本研究可以为污水处理过程的智能预警与优化调控提供有效技术支撑。

关键词: 石油, 污染, 水质预测, 神经网络, SHAP分析

Abstract:

Significant fluctuations in the influent Chemical Oxygen Demand (COD) at petrochemical wastewater treatment plants jeopardize process stability and effluent compliance, while existing prediction methods are unable to synergistically capture both its temporal dynamics and spatial correlations. To address this issue, this study proposes a Temporal-Spatial Dual-Stream Convolutional Neural Network (TS-DSCNN) model. Employing a parallel dual-stream architecture, the model leverages 1D convolution to respectively extract the temporal patterns of COD and the correlation features of key instantaneous water quality parameters (electrical conductivity, ammonia nitrogen, and suspended solids). Additionally, an attention mechanism is introduced to adaptively fuse the features from the two streams. Experimental results demonstrate that the TS-DSCNN model achieves a coefficient of determination (R²) of 0.89 and a root mean square error (RMSE) of 16 mg/L on the test set, outperforming benchmark models significantly—including Long Short-Term Memory (R² = 0.81, RMSE = 20 mg/L) and Convolutional Neural Network (R² = 0.78, RMSE = 19 mg/L). Ablation experiments and Shapley Additive Explanation (SHAP) interpretability analysis jointly verify the dominant role of the temporal stream and the effectiveness of the dual-stream architecture: the deep fusion of temporal-spatial features (R² = 0.89) outperforms the single spatial stream (R² = 0.75), single temporal stream (R² = 0.81), and simple feature concatenation model (R² = 0.84). SHAP analysis is consistent with the attention weight distribution (temporal stream weights concentrated in the range of 0.8~1.0, spatial stream weights in 0.0~0.2), confirming that the temporal stream serves as the dominant factor in prediction while the spatial stream plays an auxiliary role. A 30-day field verification shows that the model can effectively track the fluctuation trend of COD, with the mean absolute percentage error below 10% in most periods. Only localized errors occur during severe water quality fluctuations, indicating good overall stability and generalization ability. This study provides effective technical support for intelligent early warning and optimal regulation in wastewater treatment processes.

Key words: petroleum, pollution, water quality prediction, neural network, SHAP analysis

中图分类号:

X703

陈霖, 胡方杰, 王庆宏, 陈春茂. 基于时序-空间双流卷积神经网络的炼化污水厂进水COD预测模型[J]. 化工学报, DOI: 10.11949/0438-1157.20251151.

Lin CHEN, Fangjie HU, Qinghong WANG, Chunmao CHEN. Prediction model for COD in petrochemical wastewater influent based on temporal-spatial dual-stream convolutional neural network[J]. CIESC Journal, DOI: 10.11949/0438-1157.20251151.

图/表 12

图1 TS-DSCNN模型框架

Fig.1 TS-DSCNN Model Framework

图2 特征Spearman相关性分析结果

Fig.2 Results of Spearman Correlation Analysis of Features

图3 ACF及PACF分析结果

Fig.3 Results of the ACF and PACF Analysis

图4 数据增广结果

Fig.4 Data Augmentation Results

图5 模型预测效果对比结果

Fig.5 Comparison of Model Prediction Performance

图6 模型预测结果误差分布范围

Fig.6 Distribution Range of Errors in Model Predictions

图7 TS-DSCNN模型预测结果

Fig.7 TS-DSCNN Model Prediction Results

图8 TS-DSCNN模型双流注意力权重分布

Fig.8 TS-DSCNN Model Dual-Stream Attention Weight Distribution

图9 模型消融实验结果

Fig.9 Ablation Experimental Results of the Model

图10 模型可解释性分析结果

Fig.10 Model Interpretability Analysis Results

图11 TS-DSCNN模型注意力权重及SHAP值分布

Fig.11 Distribution of Attention Weights and SHAP Values in the TS-DSCNN Model

图12 TS-DSCNN模型现场验证结果

Fig.12 Field Validation Results of the TS-DSCNN Model

参考文献 34

[1]	叶黄凡. 重质油炼制污水污染特性及其处理工艺的综合评价研究[D]. 北京: 中国石油大学(北京), 2021.
	Ye H F. Characterization of heavy oil refinery wastewaters and comprehensive evaluation of the treatment processes[D]. Beijing: China University of Petroleum (Beijing), 2021.
[2]	陈霖, 刘浩威, 王庆宏, 等. 基于机器学习算法的炼化污水厂出水水质预测模型研究[J]. 工业水处理, 2025, 45(7): 81-93.
	Chen L, Liu H W, Wang Q H, et al. Research on the prediction model of effluent quality in petrochemical industries wastewater treatment plants based on optimized machine learning algorithms[J]. Industrial Water Treatment, 2025, 45(7): 81-93.
[3]	杨百玉. 炼化点源污水DOM组成及其在预处理过程中的转化行为[D]. 北京: 中国石油大学(北京), 2024.
	Yang B Y. Characteristics of dissolved organic matter in petrochemical point-source wastewaters and their transformation during pretreatment units[D]. Beijing: China University of Petroleum (Beijing), 2024.
[4]	寇悦. 炼油化工污水溶解性有机物组成及其在污水处理厂的转化行为[D]. 北京: 中国石油大学(北京), 2024.
	Kou Y. Characteristics of dissolved organic matter in petrochemical wastewaters and their transformation along treatment processes[D]. Beijing: China University of Petroleum (Beijing), 2024.
[5]	胡海良. 基于节能环保的石油化工废水处理技术[J]. 化工设计通讯, 2023, 49(6): 12-14, 52.
	Hu H L. Petrochemical wastewater treatment technology based on energy conservation and environmental protection[J]. Chemical Engineering Design Communications, 2023, 49(6): 12-14, 52.
[6]	何为, 岳留强, 唐智和, 等. 基于CNN-LSTM混合神经网络的炼化污水处理场COD排放浓度预测[J]. 西安石油大学学报(自然科学版), 2025, 40(1): 121-129.
	He W, Yue L Q, Tang Z H, et al. Prediction of COD emission concentration of petroleum refining wastewater treatment plant based on CNN-LSTM hybrid neural network[J]. Journal of Xi'an Shiyou University (Natural Science Edition), 2025, 40(1): 121-129.
[7]	唐安中, 焦赟仪, 杨斌, 等. 石化行业污水处理厂智能化建设与探索[J]. 工业水处理, 2022, 42(7): 186-191.
	Tang A Z, Jiao B Y, Yang B, et al. Intelligent construction and exploration of petrochemical industry sewage treatment plant[J]. Industrial Water Treatment, 2022, 42(7): 186-191.
[8]	Cechinel M A P, Neves J, Fuck J V R, et al. Enhancing wastewater treatment efficiency through machine learning-driven effluent quality prediction: a plant-level analysis[J]. Journal of Water Process Engineering, 2024, 58: 104758.
[9]	Duarte M S, Martins G, Oliveira P, et al. A review of computational modeling in wastewater treatment processes[J]. ACS ES&T Water, 2024, 4(3): 784-804.
[10]	Ackermann A E F, Fani V, Bandinelli R, et al. Short-term prediction in an emergency healthcare unit: comparison between ARIMA, ANN, and logistic map models[J]. Forecasting, 2025, 7(3): 52.
[11]	Azuma D, Meléndez R, Ptaszynski M, et al. SVM, BERT, or LLM? a comparative study on multilingual instructed deception detection[J]. AI, 2025, 6(9): 239.
[12]	Liu W, Wang G Y, Fu J Y, et al. Water quality prediction based on improved wavelet transformation and support vector machine[J]. Advanced Materials Research, 2013, 726/727/728/729/730/731: 3547-3553.
[13]	Huang S Z, Song G Q, Chen W, et al. Time series–based detection on tailgating fare evasions using human pose estimation[J]. Journal of Transportation Engineering, Part A: Systems, 2022, 148(7): 04022035.
[14]	Xiao Y W, Xiu J H, Zhang Y C, et al. Multi-scale transition network approaches for nonlinear time series analysis[J]. Chaos, Solitons & Fractals, 2022, 159: 112026.
[15]	Yakut Y B. Short-term voltage loss prediction in PEMFCs using single-step and multi-step LSTM models[J]. International Journal of Hydrogen Energy, 2025, 144: 652-664.
[16]	Liu Y Y, Zheng H, Zhao J S. Enhanced water quality prediction by LSTM and graph attention network (L-GAT): an analytical study of the Pearl River Basin[J]. Water Research X, 2025, 28: 100383.
[17]	Yang Y H, Zhang P T, Wang Y L. An attention-based parallel model with sliding window decomposition algorithm for water quality prediction[J]. Journal of Water Process Engineering, 2025, 78: 108751.
[18]	Chen H L, Yang J B, Fu X H, et al. Water quality prediction based on LSTM and attention mechanism: a case study of the Burnett River, Australia[J]. Sustainability, 2022, 14(20): 13231.
[19]	Rene E R, Saidutta M B. Prediction of BOD and COD of a refinery wastewater using multilayer artificial neural networks[J]. Journal of Urban and Environmental Engineering, 2008, 2(1): 1-7.
[20]	Wu X, Chen M, Zhu T Y, et al. Pre-training enhanced spatio-temporal graph neural network for predicting influent water quality and flow rate of wastewater treatment plant: improvement of forecast accuracy and analysis of related factors[J]. Science of the Total Environment, 2024, 951: 175411.
[21]	Huang S, Xia J, Wang Y L, et al. Water quality prediction based on sparse dataset using enhanced machine learning[J]. Environmental Science and Ecotechnology, 2024, 20: 100402.
[22]	Li J W, Liu J, Tauqeer I M, et al. Water chemical oxygen demand prediction based on a one-dimensional multi-scale feature fusion convolutional neural network and ultraviolet-visible spectroscopy[J]. RSC Advances, 2025, 15(24): 19431-19442.
[23]	Liu W J, Wang J, Li Z H, et al. A hybrid improved dual-channel and dual-attention mechanism model for water quality prediction in nearshore aquaculture[J]. Electronics, 2025, 14(2): 331.
[24]	Liu L B, Tian X Y, Ma Y G, et al. Online soft measurement method for chemical oxygen demand based on CNN-BiLSTM-Attention algorithm[J]. PLoS One, 2024, 19(6): e0305216.
[25]	Cai W S, Li Y K, Shao X G. A variable selection method based on uninformative variable elimination for multivariate calibration of near-infrared spectra[J]. Chemometrics and Intelligent Laboratory Systems, 2008, 90(2): 188-194.
[26]	Flores J H F, Engel P M, Pinto R C. Autocorrelation and partial autocorrelation functions to improve neural networks models on univariate time series forecasting[C]//The 2012 International Joint Conference on Neural Networks (IJCNN). June 10-15, 2012, Brisbane, QLD, Australia. IEEE, 2012: 1-8.
[27]	Demir V, Zontul M, Yelmen İ. Drug sales prediction with ACF and PACF supported ARIMA method[C]//2020 5th International Conference on Computer Science and Engineering (UBMK). September 9-11, 2020, Diyarbakir, Turkey. IEEE, 2020: 243-247.
[28]	Boschetti A, Massaron L. Python Data Science Essentials[M]. Birmingham: Packt Publishing - ebooks Account, 2015.
[29]	Priyadarshini I. An explainable autoencoder-based feature extraction combined with CNN-LSTM-PSO model for improved predictive maintenance[J]. Computers, Materials & Continua, 2025, 83(1): 635-659.
[30]	Wang Q G, Li X, Qin Q. Feature selection for time series modeling[J]. Journal of Intelligent Learning Systems and Applications, 2013, 5(3): 152-164.
[31]	Qian R, Xiong X, Zhou J H, et al. CSA-SA-CRTNN: a dual-stream adaptive convolutional cyclic hybrid network combining attention mechanisms for EEG emotion recognition[J]. Brain Sciences, 2024, 14(8): 817.
[32]	Matsunaga S, Yamaoka H, Kawasaki M, et al. On the experimental study of tuyere model ablation and its heat transfer analysis[J]. Transactions of the Iron and Steel Institute of Japan, 1976, 16(6): 332-337.
[33]	Sujitha S M S, Subiramoniyan S, Mahil J, et al. Metaheuristic-optimized swin transformer with SHAP explainability for keratoconus classification from corneal topography maps[J]. International Ophthalmology, 2025, 45(1): 396.
[34]	Huang C, Xu J J, Qiu H, et al. Developing nurse-accessible hypertension prediction tools for low-income populations: a comparative analysis of machine learning algorithms with SHAP interpretation[J]. International Journal of Nursing Practice, 2025, 31(5): e70060.