基于数据分解与NARX优化的滇池CODMn时间序列预测OA

The permanganate index(CODMn)is one of the important indicators for measuring the degree of pollution of water bodies by reducing substances.To improve the prediction accuracy of CODMn,a WPT-SHIO-NARX CODMn time series prediction model is proposed,which combines wavelet packet transform(WPT),success history intelligent optimization(SHIO)algorithm,and nonlinear autoregressive neural network(NARX).Firstly,WPT is used to decompose the CODMn time series into one periodic component and three fluctuation components;Then,the principle of SHIO is briefly introduced,and it is used to optimize hyperparameters such as NARX input delay order;Finally,based on the hyperparameters obtained through optimization,the WPT-SHIO-NARX model is established to predict the periodic and fluctuation components of CODMn.After reconstruction,the final prediction results are obtained.Comparative analyses are made with WPT-particle swarm optimization(PSO)-NARX,WPT-genetic algorithm(GA)-NARX,WPT-NARX,SHIO-NARX,WPT-SHIO extreme learning machine(ELM),and WPT-SHIO-BP neural network models.The models are validated using weekly CODMn monitoring data from 2004 to 2015 at the Xiyuan Tunnel and Guanyin Mountain sections of Dianchi Lake.The results show that the WPT-SHIO-NARX model has good predictive performance,with mean absolute percentage error(MAPE)of 0.108%and 0.045%,0.151%and 0.165%for the next 1 week and 2 weeks(half a month)of CODMn prediction at Xiyuan Tunnel and Guanyin Mountain,respectively.The MAPE for the next 4 weeks(January)of CODMn prediction is 1.383%and 0.809%,and the MAPE for the next 8 weeks(February)of CODMn prediction is 6.180%and 4.573%,respectively.The prediction accuracy is higher than other comparative models;WPT can decompose CODMn time series data into more regular subsequence components,improving the model's prediction accuracy;SHIO can effectively optimize NARX hyperparameters,significantly improving NARX performance,with optimization effects superior to GA and PSO;the NARX network has delay and feedback mechanisms,making it more suitable for time series prediction,and its predictive performance is better than that of ELM and BP networks.