Multidimensional Time Series Modeling And Its Application On Forecasting Of Insect Pest Emergence Size

In the area of agricultural production and scientific research, a time series is usually defined as a sequence set which is combined by some measured values observed within a series of moments. The sequence which has the meaning of time is also known as dynamic data. Study on the dynamic time series data, people often considering the effect of multiple factors on the observed value at the same time, thus formed the multidimensional or multivariate time series. Due to the time series is often affected by the observation and independent variables in previous period, so it is necessary to add additional dependent variables and independent variables during the several moment of the past to the independent variables in the current moment, this process is named as the ordering process. On the one hand, this article in view of the disadvantages of existing ordering methods of multidimensional time series such as the ordering process is complex and the upper threshold cannot be given automatically, we made full use of the advantages of the half variation function model of geostatistics, the ordering process of dependent variable could be conducted rapidly and automatically by the delayed effect time interval. On the other hand, multidimensional time series are often characterized by nonlinear relationship between the variables, and redundant information often present within them. So it is necessary to perform independent variable selection and following modeling and prediction with nonlinear method. After the ordering of independent and dependent variables on the multidimensional time series, we took advantage of the peculiarities of nonlinear and avoid over fitting represented on the support vector machine (SVM), the redundant variables were eliminated nonlinearly by SVM. The following SVM model could be constructed by retained independent variables and the observed value, and then to predict the unknown observed values of the time series.China has vast territory, and is also a country whose pest occurrence frequency is very high. The huge outburst of the major pests influences the development of the national agricultural economy seriously. Only improve the predictive accuracy of pests happened, can we further control the occurrence tendency of pests, take effective controlling measures timely and ensure normal growth of crops. Insect pests emergence size is not only affected by a variety of external factors (such as weather, physiology, ecology, etc.), but also is related tightly with their own occurrence tendency in the past year. It is belongs to the typical complex non-linear multidimensional time series data. We applied the proposed method to the emergence size prediction of two kinds of pests, has obtained better predictive results than other reference models. The predictive Root Mean Square Error (RMSE) of our method is84.28for the peak day amount of second generation of Cnaphalocrocis medinalis (Guenee), and the RMSEs of other four reference models (SVR-Nonlinear Independent-variable Screening, Support Vector Regression, Multiple Linear Regression, Stepwise Linear Regression) are117.29,122.77,269.44and283.47. The predictive Root Mean Square Error (RMSE) of our method is293.66for the fifth generation of Nilaparvata lugens (Stal) on the late rice, and the RMSEs of other four reference models are423.30,472.29,543.49and583.64, respectively.This method we proposed has several advantages of the half variation function model of geostatistics and SVM, it is appropriate for the prediction and forecasting of nonlinear multidimensional time series which is affected by multiple factors.