基于特征工程与仿生优化算法构建河流溶解氧预测模型  

River Dissolved Oxygen Prediction Model Based on Feature Engineering and Bionic Optimization Algorithm

作　　者：李鹏程[1,2] 苏永军 王钰 贾悦 LI Peng-cheng;SU Yong-jun;WANG Yu;JIA Yue(School of Geographic Sciences,Hebei Normal University,Shijiazhuang 050010,Hebei Province,China;Cangzhou Remote Sensing and Intelligent Water Conservancy Technology Innovation Center of Hebei University of Water Resources and Electric Engineering,Cangzhou 061001,Hebei Province,China)

机构地区：[1]河北师范大学地理科学学院,河北石家庄050010 [2]河北水利电力学院沧州市遥感与智慧水利技术创新中心,河北沧州061001

出　　处：《中国农村水利水电》2025年第2期37-44,共8页China Rural Water and Hydropower

基　　金：河北省教育厅科学研究项目资助(BJ2021056);河北省省级水利科技计划项目(2024-39);河北省水利工程局集团有限公司科技计划项目(2024-08-CX-002)。

摘　　要：河流水体中溶解氧骤增或耗竭均会引发系列环境污染、物种多样性破坏等问题,准确预测河流溶解氧(DO)浓度对河流水环境治理具有重要意义。为提高模型输入特征的可解释性及模型精度,获取河流DO浓度最优预测模型,研究利用黄河流域山西境内水质监测站点数据,以双向长短期记忆网络(BiLSTM)为基础,结合卷积神经网络模型(CNN)和注意力机制(Attention Mechanism),基于随机森林模型(RF)进行特征优选,建立RF-CNN-BiLSTM-Attention(RF-CBA)模型,进一步利用吸血水蛭优化算法(BSLO)、黑翅鸢优化算法(BKA)、白鲨优化算法(WSO)等仿生优化算法,构建了BSLO-RF-CBA、BKA-RF-CBA、WSO-RF-CBA共3种优化模型,并与深度学习中CNN-A、LSTM-A、BiLSTM-A、CBA、RF-CBA模型对比,分析得到河流溶解氧预测结果,以平均绝对误差(MAE)、均方根误差(RMSE)、均方误差(MSE)、决定系数(R2)、全绩效指标(GPI)和相对误差(MAPE)评价不同模型精度,结果表明:(1)RF模型通过对影响河流DO特征值进行排序、筛选,可消除冗余特征对水质预测模型的影响,提高预测精度。(2)利用仿生算法优化RF-CBA模型的神经元数量、学习率、正则化系数等参数,模型模拟精度进一步提升,总体上捕捉到了DO波动的时间序列特征,模型表现出强稳定性和泛化能力。(3)BSLO-RF-CBA模型模拟精度最高,对DO变化捕捉能力突出,具有更强的捕获全局依赖关系的能力,推荐用于河流溶解氧预测模型。该模型具备扩展至不同河流溶解氧等污染物浓度预测的能力,为河流水体污染预警与系统化管理提供技术支撑。Sudden increase or depletion of dissolved oxygen in river water bodies can cause a series of environmental pollution,species diversity destruction and other problems,and accurate prediction of dissolved oxygen(DO)values in rivers is of great significance to the management of river water environment.In order to improve the interpretability of model input features and model accuracy,and to obtain the optimal prediction model of river DO values,this study utilized the data from water quality monitoring stations in the Yellow River Basin in Shanxi,and used the bidirectional long and short-term memory network(BiLSTM)as the basis,combined with the convolutional neural network model(CNN)and the Attention Mechanism to optimize the feature selection,and conducted feature optimization based on the Random Forest Model(RF),established RF-CNN-BiLSTM-Attention(RF-CBA)model,and further utilized the bionic optimization algorithms such as Bloodsucking Leech Optimization Algorithm(BSLO),Black-winged Kite Optimization Algorithm(BKA),and White Shark Optimization Algorithm(WSO).A total of five optimization models,BSLO-RF-CBA,,BKA-RF-CBA,and WSO-RF-CBA,were constructed and compared with the CNN-A,LSTM-A,BiLSTM-A,CBA,and RF-CBA models of deep learning,and analyzed to obtain the prediction results of the river dissolved oxygen with the Mean Absolute Error(MAE),Root Mean Square Error(RMSE),Mean Square Error(MSE),Coefficient of Determination(R2),Global Performance Indicator(GPI),and Relative Error(MAPE)to evaluate the different model accuracies,and the results showed that:①The RF model can eliminate the effect of redundant features on the water quality prediction model and improve prediction accuracies by sorting and filtering the influence of the feature values affecting the DO in rivers.②Using the bionic algorithm to optimize the number of neurons,learning rate,regularization coefficient and other parameters of the RF-CBA model,the model simulation accuracy was further improved,the overall time series characteristics of DO fluctuati

关 键 词：溶解氧 双向长短期记忆网络机 特征优选 仿生优化算法 耦合模型 

分 类 号：TV11[水利工程—水文学及水资源] X832[环境科学与工程—环境工程]