运用LXSAWOA-MBP神经网络模型对污水处理设施出水中总氮含量的预测

Prediction of Total Nitrogen Content in Effluent from Wastewater Treatment Plants Using the LXSAWOA-MBP Neural Network Model

摘要: 为了提升污水处理厂排放水质预测的准确性，并增强鲸鱼优化算法(WOA)在搜索过程中的精确度，本文提出了一种结合拉普拉斯交叉算子(LX)与模拟退火算法(SA)对WOA进行改进的新方法，即LXSAWOA。基于此优化算法，构建了一个用于水质预测的多层BP神经网络模型。以云南省普洱市江城县某污水处理厂出水总氮浓度预测为例进行了实证分析。选取六个典型测试函数对提出的LXSAWOA进行了仿真验证，并将其结果与仅使用LX改进的WOA以及原始WOA进行了对比。通过应用LXSAWOA来确定MBP神经网络中隐藏层节点数量的最佳配置，建立了含有2至5个隐藏层的不同版本LXSAWOA-MBP预测模型，并将它们的性能与单隐藏层结构下的LXSAWOA-BP、LXSAWOA-SVM、权重阈值调整后的LXSAWOA-BP及普洱PSO-SVR模型进行了比较。实验结果显示，LXSAWOA不仅在寻找最优解方面优于LXWOA和标准WOA，展示了较强的全局搜索能力；而且当应用于MBP神经网络时，对于实际案例中总氮浓度的预测误差绝对平均值控制在1.76%~1.38%之间，优于其他三种模型及PSO-SVR模型的预测效果，证明了LXSAWOA能够有效优化MBP神经网络架构中的隐藏层节点数。本研究所提出的方法为水质预测以及其他相关领域的预测研究提供了有价值的参考。

Abstract: To enhance the accuracy of effluent quality prediction in wastewater treatment plants and improve the precision of the Whale Optimization Algorithm (WOA) during its search process, this paper proposes a novel method that combines the Laplace Crossover (LX) and Simulated Annealing (SA) to optimize WOA, termed LXSAWOA. Based on this optimization algorithm, a multi-layer Backpropagation (BP) neural network model for water quality prediction was constructed. An empirical analysis was conducted using the prediction of total nitrogen concentration in the effluent of a wastewater treatment plant in Jiangcheng County, Pu’er City, Yunnan Province, as a case study. Six typical test functions were selected to simulate and validate the proposed LXSAWOA, and the results were compared with those from WOA improved only by LX (LXWOA) and the standard WOA. By applying LXSAWOA to determine the optimal configuration of the number of hidden layer nodes in the MBP neural network, different versions of the LXSAWOA-MBP prediction model with 2 to 5 hidden layers were established. Their performance was compared with that of the single-hidden-layer LXSAWOA-BP, LXSAWOA-SVM, LXSAWOA-BP with adjusted weights and thresholds, and the Pu’er PSO-SVR model. Experimental results demonstrate that LXSAWOA not only outperforms LXWOA and the standard WOA in finding the optimal solution, showcasing strong global search capabilities, but also, when applied to the MBP neural network, achieves a mean absolute prediction error for total nitrogen concentration in the practical case study ranging between 1.76% and 1.38%. This performance is superior to that of the other three models and the PSO-SVR model, proving that LXSAWOA can effectively optimize the number of hidden layer nodes in the MBP neural network architecture. The method proposed in this study provides a valuable reference for water quality prediction and forecasting research in other related fields.