雄先型核桃雄花疏除的二次回归与BP神经网络模型研究

雄先型核桃雄花疏除（去雄）是提高产量的重要管理措施，为提高核桃去雄的效率，建立二次回归与BP神经网络模型。分别以乙烯利、赤霉素和甲哌鎓为自变量和核桃雄花脱落率为响应指标，进行田间建模试验，建立了二次多项式回归方程和BP神经网络模型，并于翌年进行BP模型田间确认试验。试验数据分为训练集、确认集和试验集，中心组合（二次旋转回归试验设计）田间建模试验得到的20组数据随机划为训练集(17)和确认集(3)数据，试验集为翌年田间确认试验得到的数据，BP神经网络的拓扑结构为3-5-1。（1）BP神经网络对确认集样本的预测值误差分别为1.3550%、0.4291%、0.3538%；（2）BP神经网络的预测值与田间确认试验结果相差为2.04%，回归预测值与田间确认试验结果相差为3.12%；（3）BP神经网络预测比回归预测提高预测精度1.0%以上。将二次多项式逐步回归分析和BP神经网络方法有效的结合使用，既可明确各因子的作用效应亦可得到相对准确的预测结果。

Study on the Model of Excessive Staminate Catkin Thinning of Proterandrous Walnut based on Quadratic Polynomial Regression Equation and BP Artificial Neural Network

It’s an important management measure for raising yield of proterandrous walnut that excessive staminate catkin of walnut (SCW) was thinned. The model of quadratic polynomial regression equation and BP artificial neural network was developed to apply SCW thinning measure. The effect of three independent parameters, namely ethrel, gibberellin and mepiquat on the average thinned percentage of SCW were considered with field experiment in which central composite design (CCD) and quadratic polynomial regression equation were employed. BP artificial neural network was developed to predict result of field experiment on the base of field model experiment and BP model of validation testing was carried out in the next year. Total 20 sets data of field model experiment CCD quadratic rotation regression design method were divided randomly into training group (17 sets data) and validating group (3 sets data) respectively, the data of validation testing was employed as testing group and the topological structure of BP was 3-5-1. The experiment result indicated that: (1) The error of predicted value of obtained BP artificial neural network for validating group was 1.3550%, 0.4291%, 0.3538%, respectively. (2) Difference between BP predicted value and the value of field experiment validation test was 2.04% , while the difference between regression predicted value and the value of field experiment validation test was 3.12%. (3) Predictive accuracy of BP was 1.0% more than predictive accuracy of regression predicted. Based on the practical strategy of integrated quadratic polynomial stepwise regression equation with BP artificial neural network could distinguish the effect of independent parameters and improve the predictive accuracy.