基于时间序列与横截面数据的吉林省水稻产量预测对比分析

Comparative analysis of rice yield forecasting based on time series analysis and cross-sectional prediction in Jilin Province of China

关键词:  水稻产量  ARIMA  LSTM  SVR  MLP  时间序列

Key words:  rice yield  ARIMA  LSTM  SVR  MLP  time series

摘要:

目的 对比不同模型预测效果，分析各模型预测水稻产量的特点、不足及适用条件，为粮食产量预测问题模型选择提供依据。方法 从时间序列预测和横截面数据预测两种角度，利用ARIMA、LSTM、SVR、MLP这4种模型，通过吉林省水稻产量、病虫害及其他特征历史数据对吉林省水稻产量进行预测，并对不同模型的预测结果进行了对比分析。结果 基于ARIMA模型和LSTM模型的时间序列预测结果较好，横截面数据预测中，原始数据经主成分分析PCA降维处理后，可提高模型预测性能。结论 对于水稻产量预测，应根据掌握的影响产量因素的数据以及趋势延续性情况合理选择预测模型，以达到较理想的预测效果。

Abstract:

Purpose Rice is one of the main food crops in China. Rice yield greatly contributes to China’s food security. Rice yield forcasting is of great significance to China’s agricultural development. Traditional approaches use time series or cross-sectional data， and apply statistical methods or machine learning methods to predict rice yield. However，the comparative analyses between multiple methods are still lacking. This paper aims to compare the differences between the model’s forecasting performance，and to analyze the usability of models in specific circumstances，in order to better provide implications for national food security issues.Methods In this paper，we use multiple data sources，including historical rice yield，disease and pest outbreak，and other characteristics data，to forecast rice yield in Jilin province，China. Four models are used to forecast rice yield. ARIMA and LSTM are utilized for time series forecasting. Cross-sectional prediction is also conducted using SVR and MLP. The forecasting results of different models are analyzed and compared.Results Both ARIMA and LSTM have resulted in a good forecasting performance using time series data. The forecasting performance would be further improved if a principle component analysis is applied to reduce the dimensionality of the original data.Conclusion Comparing to the cross-sectional data forecasting，time series data forecasting has achieved a better performance. This indicates that a rational selection of data and model would improve the performance in rice yield forecasting.