基于改进的BAS-BP神经网络的参考作物腾发量预测

Reference crop evapotranspiration prediction based on the improved BAS-BP neural network

针对求解参考作物腾发量时较多关联性强的气象因素以及BP神经网络自身的局限性,通过平均影响值法(mean impact value,MIV)和SPSS软件对相关参数进行降维筛选,并通过改进后使用非线性递减步长的天牛须搜索(nonlinear decreasing beetle antennae search,NDBAS)算法优化BP神经网络来进行预测,同时建立基于BP神经网络、BAS-BP神经网络的预测模型进行对比分析。结果表明,NDBAS-BP模型的决定系数R^(2)为0.8858,优于另外两个对比模型;且其平均绝对误差M_(AE)为0.3587 mm/d,低于BAS-BP和BP模型的0.3981和0.3797 mm/d。3种模型中,NDBAS-BP模型的R^(2)值最大,M_(AE)最小,证明NDBAS-BP模型的预测精度更加接近真实数据。

Aiming at the meteorological factors with strong correlation and the limitations of BP neural network itself when computing the reference crop evapotranspiration,the mean impact value(MIV)method and SPSS software were used to reduce the dimension of relevant parameters.The BP network was optimized by using the improved nonlinear decreasing beetle antennae search(nonlinear decreasing beetle antennae search,NDBAS)algorithm to make prediction.And establish the prediction model based on BP neural network and BAS-BP neural network for comparative analysis.The results show that the coefficient of determination R^(2)of the NDBAS-BP model is 0.8858,which is better than the other two prediction models;and its mean absolute error M_(AE)is 0.3587 mm/d,which is lower than the 0.3981 mm/d,0.3797 mm/d of the BAS-BP and BP models.Among the three models,the NDBAS-BP model has the largest R^(2)value and the smallest M_(AE),suggesting that the prediction accuracy of the NDBAS-BP model is closer to the real data.