Comparison of the effects of combined model and single model in HFRS incidence fitting and prediction / 公共卫生与预防医学

ABSTRACT

Objective To compare the prediction effect of combined model and single model in HFRS incidence fitting and prediction, and to provide a reference for optimizing HFRS prediction model. Methods The province with the highest incidence in China (Heilongjiang Province) in recent years was selected as the research site. The monthly incidence data of HFRS in Heilongjiang Province from 2004 to 2017 were collected. The data from 2004 to 2016 was used as training data, and the data from January to December 2017 was used as test data. The training data was used to train SARIMA , ETS and NNAR models, respectively. The reciprocal variance method and particle swarm optimization algorithm (PSO) were used to calculate the model coefficients of SARIMA, ETS and NNAR, respectively, to construct combined model A and combined model B. The established models were used to predict the incidence of HFRS from January to December 2017. The fitted and predicted values of the five models were compared with the training data and test data, respectively. Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Standard Deviation (RMSE), and Mean Error Rate (MER) were used to evaluate the model fitting and prediction effects. Results The optimal SARIMA model was SARIMA(1,0,2)(2,1,1)12. The optimal ETS model was ETS(M, N, M), and the smoothing parameter =0.738，=1*10. The optimal NNAR model was NNAR(13,1,7)12. The residuals of the three single models were white noise (P>0.05). The expression of combined model A was ŷ=0.134*ySARIMA+0.162*yETS+0.704*yNNAR; the expression of combined model B was ŷ=0.246*ySARIMA+0.435*yETS+0.319*yNNAR. The MAPE, MAE, RMSE, and MER fitted by SARIMA, ETS, NNAR, combined model A and combined model B were 24.10%, 0.11, 0.17, 23.29%; 17.14%, 0.08, 0.14, 17.96%; 6.33%, 0.02, 0.03, 4.25%; 9.03%, 0.03, 0.05, 7.51%; 13.16%, 0.06, 0.09, 12.33%, respectively. The MAPE, MAE, RMSE, and MER predicted by the five models were 18.70%, 0.05, 0.06, 19.62%; 23.83%, 0.06, 0.07, 24.49%; 28.30%, 0.07, 0.10, 29.21%; 21.69%, 0.06, 0.08, 22.63%; 17.39%, 0.05, 0.07, 18.76%, respectively. Conclusion The fitting and prediction effects of the combined models are better than the single models. The combined model based on PSO to calculate the weight of the single model is the optimal model.