工业水处理 ›› 2025, Vol. 45 ›› Issue (7): 81-93. doi: 10.19965/j.cnki.iwt.2024-0515

• 试验研究 • 上一篇    下一篇

基于机器学习算法的炼化污水厂出水水质预测模型研究

陈霖1(), 刘浩威1, 王庆宏1, 冯光明2, 詹亚力1, 王强3, 陈春茂1()   

  1. 1. 中国石油大学(北京)化学工程与环境学院,石油石化污染物控制与处理国家重点实验室,北京 102249
    2. 中国石油天然气股份有限公司华北油田巴彦勘探开发分公司,内蒙古 巴彦淖尔 015000
    3. 葫芦岛市生态环境保护综合行政执法队,辽宁 葫芦岛 125080
  • 收稿日期:2024-10-14 出版日期:2025-07-20 发布日期:2025-07-22
  • 通讯作者: 陈春茂
  • 作者简介:

    陈霖(1997— ),在读博士,E-mail:

  • 基金资助:
    中国石油天然气股份有限公司十四五前瞻性基础性战略性技术研究项目(2022DJ6904)

Research on the prediction model of effluent quality in petrochemical industries wastewater treatment plants based on optimized machine learning algorithms

Lin CHEN1(), Haowei LIU1, Qinghong WANG1, Guangming FENG2, Yali ZHAN1, Qiang WANG3, Chunmao CHEN1()   

  1. 1. State Key Laboratory of Heavy Oil Processing, College of Chemical Engineering and Environment, China University of Petroleum(Beijing), Beijing 102249, China
    2. Bayan Exploration and Development Branch of North China Oilfield, NCPC, Bayannur 015000, China
    3. Huludao City Comprehensive Administrative Law Enforcement Team for Environmental Protection, Huludao 125080, China
  • Received:2024-10-14 Online:2025-07-20 Published:2025-07-22
  • Contact: Chunmao CHEN

摘要:

炼化企业生产工艺流程复杂且装置繁多,污水水质和水量波动大,下游响应调控滞后,水质超标问题难以避免,亟需构建高效水质预测模型。以广东省某炼化企业2023年全年监测池出水水质数据为基础,构建水质预测模型。结果表明:插值算法可以实现对炼化污水缺失数据的有效填充;出水硫化物(HS)、总氮(TN)、总有机碳(TOC)、五日生化需氧量(BOD5)、pH与化学需氧量(COD)未表现出明显的相关性,多参数预测模型无法捕获数据特征;选用反向传播-神经网络(BP-NN)与支持向量回归机(SVR)为基础算法构建的时间序列预测模型可以大幅提高预测准确性,变异粒子群算法(MPSO)可以实现对BP-NN权值、阈值以及SVR惩罚因子c和核函数参数g的显著优化;MPSO-BP-NN模型在测试集中对COD的预测精度最高,决定系数(R 2)和相关系数(r)分别为0.81和0.89,MAE、RMSE、MBE和MAPE分别为1.10 mg/L、1.63 mg/L、-0.25 mg/L和2.58%;现场验证结果表明MPSO-BP-NN模型有较好的稳定性和泛化能力,可以显著提升预测水质数据的时效性,为炼化污水处理系统上游工艺参数的调控提供理论指导,保障系统长周期平稳运行。

关键词: 炼化污水, 水质预测, 相关性分析, 机器学习, 算法优化

Abstract:

The production processes in petrochemical enterprises are characterized by intricate workflows and numerous installations, leading to significant fluctuations in wastewater quality and volume. Downstream response regulation often suffers from delays, making it challenging to prevent the exceeding of water quality. There is an urgent need to establish an efficient water quality prediction model. This study developed a predictive model based on annual 2023 effluent water quality data from the monitoring tank of a petrochemical enterprise in Guangdong Province. The results demonstrated that interpolation algorithms effectively imputed missing data in petrochemical wastewater. Effluent parameters, including hydrogen sulfide (HS), total nitrogen (TN), total organic carbon (TOC), five-day biochemical oxygen demand (BOD5), pH, and chemical oxygen demand (COD), exhibited no significant correlations, rendering multi-parameter prediction models ineffective in capturing data characteristics. Time-series models constructed using backpropagation neural networks (BP-NN) and support vector regression (SVR) as foundational algorithms significantly improved prediction accuracy. Modified particle swarm optimization (MPSO) effectively optimized the weights and thresholds of BP-NN, as well as the penalty factor (c) and kernel function parameter (g) of SVR. The MPSO-BP-NN model achieved the highest prediction accuracy for COD in the test set, with a coefficient of determination (R²) of 0.81 and a correlation coefficient (r) of 0.89. The mean absolute error (MAE), root mean square error (RMSE), mean bias error (MBE), and mean absolute percentage error (MAPE) were 1.10 mg/L, 1.63 mg/L, -0.25 mg/L, and 2.58%, respectively. Field validation confirmed the robustness and generalization capability of the model, with significantly enhanced the timeliness of water quality predictions. This model provides theoretical guidance for optimizing upstream process parameters and ensuring long-term stable operation in petrochemical wastewater treatment systems.

Key words: refining wastewater, water quality prediction, correlation analysis, machine learning, algorithm optimization

中图分类号: