Comparison of time series models for predicting net primary productivity dynamic changes of Abies fargesii Franch. on the southern slopes of Taibai Mountain

Based on meteorological data collected from 1959 to 2016 and physiological parameters of Abies fargesii Franch. forest in Taibai Mountain, we used the Biome-BGC model to calculate some results, and then analyzed the results. Then we got the annual net primary productivity (NPP) of Abies fargesii forest on the southern slopes of Taibai Mountain. Using the autoregressive integrated moving average (ARIMA) model, R language, and NAR (Nonlinear auto-regressive) dynamic neural network model to make trend fitting and short-term predictions regarding changes about NPP's dynamic change respectively, in order to establish a time series model that applied to NPP of Abies fargesii forest on the southern slopes of Taibai Mountain. Using the white noise test and other inspection methods, we evaluated the predictive results of the three models.Results indicated that:the NPP of Abies fargesii forest on the southern slopes of Taibai Mountain showed a rising trend from 2017 to 2026, and the highest value probably appeared since 1959. In forecasting future changes in Abies fargesii forest, the three prediction models demonstrated their own characteristics:the ARIMA model passed the white noise test on the NPP prediction results of the Abies fargesii forest, and given the possible results under different confidence intervals; NAR dynamic neural network model showed good fitting effect, and also passed the error autocorrelation test. The prediction results well simulated future change trends; R language can use the basic data to simulate the NPP dynamic change of Abies fargesii forest on the southern slopes of Taibai Mountain after removing abnormal data points. The results showed that the correlation between the prediction results and verification results was 0.944, and the P value of the error term was far lower than 0.01. The models constructed by the three methods showed good results in data fitting, and the prediction results were also in the credible range. Therefore, different methods can be selected according to the characteristics of data in practical work.