基于小波神经网络的麦蚜发生程度预测模型

Forecasting model for the occurrence degree of wheat aphids based on wavelet neural network

【目的】建立基于小波神经网络病虫害预测预报模型,对提前采取防病防虫措施、减少农作物病虫害损失、提高农作物产量与质量具有重要意义。【方法】本研究以山西省运城市芮城县1980-2014年麦蚜发生程度和气象因子数据为基础,采用主成分分析法从40个基础气象因子中整合形成9个新的自变量输入模型,采用试凑法筛选隐含层节点数,用1980-2009年的数据进行网络训练,对2010-2014年麦蚜发生程度进行回测,建立了以Morlet小波函数为传递函数的小波神经网络模型,并与以Sigmoid函数为传递函数的BP神经网络模型进行了比较。【结果】小波和BP神经网络两种模型对训练样本的平均拟合精度均有10年以上超过80%,两者MAPE值分别为89.83%和83.07%,MSE值分别为0.0578和0.6192。【结论】两个模型都能较好地描述麦蚜发生程度;从预测精度和模型的稳定性来看,小波神经网络好于BP神经网络。

[Aim]This study aims to build up a pest and disease forecast model based on wavelet neural network（ WNN）,so as to provide a basis for taking measures to prevent pests and diseases,reducing crop damage by pests and diseases and improving quantity and quality of crop yields. [Methods]Based on the occurrence degree of wheat aphids from 1980 to 2014 and the meteorological factors in Ruicheng County,Yuncheng City,Shanxi Province,we integrated and created 9 new independent variable input models from 40 fundamental meteorological factors through Principal Component Analysis（ PCA） and screened hidden layer nodes by trial and error method,conducted training with data from 1980 to 2009 and retested the occurrence degree of wheat aphids from 2010 to 2014. Finally,the study built up a WNN model by taking wavelet function as transfer function and contrasted itself with BP neural network（ BPNN） model which takes Sigmoid function as transfer function. [Results]The average fitting accuracy of both models,namely,WNN and BPNN models,were above 80% in at least 10 years. Their MAPE values were 89. 83% and 83. 07%,and their MSE values were 0. 0578 and 0. 6192,respectively.[Conclusion]Both models can well illustrate the occurrence degree of wheat aphids. As for the forecast accuracy and model stability,however,WNN is better than BPNN.