【摘要】 为精确预测企业废水污染物排放量，以江苏省某电子制造企业为例，加入污染物历史排放特征，基于每小时工况数据和化学需氧量排放数据(12 127组样本),构建基于极端梯度提升(XGBoost)的实时预测模型。结果表明，在用电数据的基础上增加污染物历史排放数据作为特征变量，能够有效提升模型预测精度。相较于轻量级梯度提升机(LightGBM)、随机森林(Random Forest)、梯度提升回归树(GBR)等机器学习算法，XGBoost预测效果更优，其拟合优度(R~2)为0.95,均方根误差(RMSE)为171.32 g/h,平均绝对百分比误差(MAPE)为13.55%,平均绝对误差(MAE)为49.53 g/h。该模型短期预测效果(第1周的RMSE=82.84 g/h, MAE=38.83 g/h, MAPE=16.38%)优于长期预测效果(第3周的RMSE=112.46 g/h, MAE=55.32 g/h, MAPE=16.89%),对于中长期预测需要通过迭代更新来保持模型预测精度。该方法可以快速预测企业未来污染物排放量，支撑企业合理安排生产计划，为环境管理决策提供技术支持。更多还原

【Abstract】 Accurate prediction of wastewater pollutant emissions contributes to the rational planning of production schedules and the reduction of pollution emissions. In this study, taking a company in Jiangsu Province as an example, a total of 12 127 samples were used to build an XGBoost prediction model based on hourly operating data and hourly Chemical Oxygen Demand emissions data. The differences in model accuracy were compared under different feature selection scenarios and with other machine learning algorithms. The results showed that for long-term pollutant prediction, the prediction accuracy of the model could be effectively improved by adding historical pollutant emission data as feature variables on top of electricity consumption data. Compared with machine learning algorithms such as LightGBM, Random Forest, and GBR, XGBoost showed the best prediction with an R~2 of 0.95, RMSE of 171.32 g/h, MAPE of 13.55%, and MAE of 49.53 g/h. The XGBoost performed better in the first three weeks than after three weeks, and is good for short-term predictions. It requires iterative updating to maintain model prediction accuracy for medium and long-term predictions. This method enables the precise prediction of future pollutant emissions from enterprises and provides support for environmental management decision-making.更多还原