【摘要】 为了精准预测污水中溶解氧(DO)浓度值，通过局部异常因子(LOF)算法对深圳某污水处理厂5个月的数据进行分析。利用集合经验模态分解(EEMD)-长短期记忆(LSTM)神经网络模型，对曝气控制系统的出水水质影响较大的DO浓度进行准确预测。首先，通过LOF算法剔除数据中的异常值。然后，使用EEMD算法筛选出输入数据中强相关的特征子序列。最后，将特征子序列输入LSTM模型中以得到DO预测值。试验结果表明，LOF-EEMD-LSTM模型的准确率可达95.4%、平均绝对误差(MAE)为0.036、均方误差(MSE)为0.003 8、均方根误差(RMSE)为0.061 4、平均绝对百分比误差(MAPE)为0.046。以上指标相比于反向传播(BP)神经网络、随机森林、LSTM、LOF-LSTM、EEMD-LSTM和变分模态分解-最小二乘支持向量机(VMD-LSSVM)预测模型皆有明显的提升。所提模型的预测精度较高，具有较高的实用价值。更多还原

【Abstract】 In order to accurately predict the concentration value of dissolved oxygen(DO) in wastewater, five months of data from a wastewater treatment plant in Shenzhen are analyzed through the local outlier factor(LOF) algorithm. The model of ensemble empirical mode decomposition(EEMD)-long short-term memory(LSTM) neural network are used to accurately predict the concentration of DO, which is a major influence on the effluent quality of the aeration control system. Firstly, outliers in the data are removed by the LOF algorithm. Then, the EEMD algorithm is used to filter out the strongly correlated feature subsequences in the input data. Finally, the feature subsequences are input to the LSTM model to obtain the predicted values of DO. The experimental results show that the accuracy rate of the LOF-EEMD-LSTM model can reach 95.4%, the mean absolute error(MAE) is 0.036, the mean square error(MSE) is 0.003 8, the root mean square error(RMSE) is 0.061 4, and the mean absolute percentage error(MAPE) is 0.046. The above metrics are significant improvement that comparable to those of back propagation(BP) neural network, random forest, LSTM, LOF-LSTM, EEMD-LSTM and variational mode decomposition-least squares support vector machine(VMD-LSSVM) prediction models. The proposed model has higher prediction accuracy and has higher practical value.更多还原