ARIMA乘积季节模型与LSTM神经网络模型对我国麻疹发病数预测效果的比较

目的 探讨长短期记忆神经网络(long short term memory, LSTM)模型和差分整合移动平均自回归(autoregressive integrated moving average model, ARIMA)乘积季节模型在全国麻疹发病趋势预测中的应用，为麻疹的早期防控提供科学依据。方法 选取2005年1月至2016年2月全国麻疹月发病数分别构建LSTM模型和ARIMA乘积季节模型，同时运用得到的模型对2016年3月至2018年12月发病数进行预测，运用两种评价指标平均绝对百分比误差(mean absolute percentage error, MAPE)和均方根误差(root mean square error, RMSE)检验模型的外推预测精度。最后应用模型预测2019年1月至2019年5月的全国麻疹月发病数。结果 LSTM模型和最优模型ARIMA(0,1,1)(0,1,1)₁₂外推预测的均方根误差(RMSE)分别为0.25和1.54,平均绝对百分比误差(MAPE)分别为3.6%和18.7%,提示LSTM神经网络的外推预测精度优于ARIMA模...

Comparison between ARIMA model and LSTM neural network model in predicting measles incidence in China

Objective To integrate the application of long short term memory(LSTM) model and multiple season autoregressive and moving average model(ARIMA) in predicting the incidence trend of measles in China, and to provide scientific basis for the early prevention and control of measles. Methods LSTM model and ARIMA model were respectively constructed based on the monthly incidence of measles in China from January 2005 to February 2016, and the incidence of measles from March 2016 to December 2018 were predicted by using the obtained model. Two evaluation indexes, MAPE and RMSE were used to test the prediction accuracy. Finally, the model was used to predict the monthly incidence of measles in China from January 2019 to May 2019. Results The RMSE of LSTM model and the optimal model ARIMA(0,1,1)(0,1,1)12 were 0.25 and 1.54 respectively, and the MAPE were 3.6% and 18.7% respectively, suggesting that the prediction accuracy of LSTM neural network was better than that of ARIMA model. The prediction confirmed very well to field results. Conclusion The prediction effect of LSTM neural network model is better than ARIMA model, which can provide theoretical basis for the early warning and prevention and control of measles.