摘要
大豆是重要的经济作物,同时也是我国市场化和国际化程度最高的大宗农产品,对其价格进行预测具有重要意义。采用Q-RBF神经网络模型对国产大豆价格进行预测,该模型具有如下两个特点:(1)通过分位数回归功能来描述大豆在不同价格水平下的分布特征;(2)通过RBF神经网络结构来刻画大豆价格的非线性关系。在模型参数优化时,由于遗传算法是一种全局搜索优化方法,但是搜索速度慢、对初始值具有一定依赖性;而梯度下降法具有收敛快,对初始值没有特定要求等优点,所以本文提出遗传算法与梯度下降法相结合的混合改进算法,其基本思想是利用梯度下降法的局部寻优能力加快遗传算法的收敛速度。采用2010年1月-2015年12月的国产大豆月度价格数据进行预测研究,结果表明,算法收敛速度较快,模型预测精度较高,是可以泛化应用的预测模型。
Soybean is an important cash crop, and it is also the most important agricultural product with the highest degree of marketization and internationalization character in China. So it is essential to forecast the soybean price. A Quantile-RBF （ Q- RBF） neural network model is proposed to predict Chinese domestic soybean price in this paper. The model has two characteristics as follows ： （ 1 ） Quantile regression models describe the distribution over the range of the soybean price ; （2） RBF neural networks approximate the nonlinear part of soybean price. The parameters of Q-RBF neural network model can be optimized through the genetic algorithm （GA） and the gradient descent method. GA is a global optimization method, however, it might be slow in convergence. On the contrary, the gradient descent method quickly converges to an optimal solution, but may converge to a local minimum or maximum and is not efficient in discontinuous problems. Therefore, an improved algorithm combi- ning GA with gradient descent method is proposed in this paper. In the improved algorithm, the gradient descent method is used to improve the convergence efficiency of GA. The data of monthly soybean price from Jan. 2010 to Dec. 2015 were analyzed. Experimental resuhs demonstrated that the Q-RBF neural network model and improved algorithm were accurate and effective.
出处
《大豆科学》
CAS
CSCD
北大核心
2017年第1期143-149,共7页
Soybean Science
基金
国家自然科学基金(71301077)
南京农业大学中央高校基本科研业务费人文社会科学基金(SK2014011)
关键词
预测
Q—RBF神经网络
梯度下降法
遗传算法
概率密度函数
Forecasting
Quantile regression-radial basis function （Q-RBF） neural network
Gradient descent method
Genetic algorithm
Probability density function
