[Objectives]  To determine the feasibility of using meteorological factors from different sites to develop models to predict the annual population dynamics of insect pests. [Methods]  Meteorological data, including temperature, humidity, rainfall and sunshine duration in Changsha from 2000 until 2013, and from Chenzhou, Guangzhou and Nanchang from 2000 until 2015, and data on the abundance of Spodoptera litura (Fabricius) in Chenzhou from 2000 until 2015, were analyzed using stepwise regression and Chi-square tests to develop effective, multivariate forecasting models. [Results]  Explanatory variables were selected based on a 16-year-long S. litura dataset from Chenzhou city, 170 meteorological factors for the corresponding years in Chenzhou and Changsha, and 104 meteorological factors for corresponding years in Guangzhou and Nanchang. For regression models and Chi-square tests, alpha was ≤ 0.05 and variance inflation factors (VIF) ≤ 5. Seventy-four significant explanatory variables were obtained and 16 effective models were created. The results of Chi-square tests indicate that χ² for back-substitution results from all 16 models was between χ²<χ²_0.995, in other words, there was no significant difference between the back-substituted predicted values and the observed values. We obtained four groups of Chi-square cumulative values; the first group was within χ²<χ²_0.995, the second group within χ²<χ²_0.950, the fourth group within χ²<χ²_0.500, and the third group within χ²<χ²_0.250. [Conclusion]  It is accurate and feasible to predict the annual occurrence dynamics of pests on the basis of meteorological factors from different locations obtained in previous years. Compared with predictive methods using only data from the current year and local meteorological factors, the main advantages of this approach are that, provided there is sufficient historical information on pest abundance and meteorological data, it can provide predictions a long time in advance and in the absence of local historical meteorological data. Since there is often insufficient local data to create accurate models the results of this study have broad general applicability.