基于BP神经网络的油茶产量预测模型构建

Building Camellia oleifera yield prediction model based on BP neural network

【目的】改良传统农业生产预测方法,实现油茶产量的快速高效预测,为油茶实际生产提供参考。【方法】收集油茶栽种较为集中的湖南、江西、浙江、广西的油茶籽年度总产量、实有油茶林面积、气象等数据,选择17个气象指标作为油茶产量的影响因素,使用MATLAB软件,通过主成分分析提取出主成分,再将主成分作为BP神经网络的输入集,在传统神经网络模型基础上构建主成分分析与BP神经网络组合模型,对4个地区油茶籽单位面积年产量进行预测。BP神经网络的训练集和测试集选用1990—2018年的数据,采用2019年的数据对模型预测效果进行验证,最后应用模型对2025年油茶籽单位面积产量进行预测。【结果】对主成分有重要贡献的气象因子有日照时长、6—11月气温、3—5月降水量、平均最低气温、露点温度、平均风速、最大持续风速以及海平面气压。改进后模型迭代耗时更少,拟合度较高,对4个地区油茶产量的预测结果的平均相对误差均低于3%。应用模型预测得到2025年湖南、江西、浙江、广西的单位面积油茶干籽产量分别为0.831、0.583、0.449、0.512 t/hm^(2)。【结论】与传统预测模型相比,改进后的主成分分析与BP神经网络组合模型的预测效率和预测精度均有提高,在今后一段时期内4个地区的油茶产量有较好的发展趋势。

【Objective】The purpose of this study is to improve the original prediction method of agricultural production,realize the rapid and efficient prediction of Camellia oleifera yield, and provide theoretical basis for the actual production of C. oleifera.【Method】The total annual output of C. oleifera seed, actual C. oleifera forest area, meteorological data and other data were collected in Hunan, Jiangxi, Zhejiang and Guangxi which were the main cultivation areas of C. oleifera. 17 meteorological indicators were selected as the influencing factors of C. oleifera production. MATLAB software was used to extract the principal component through principal component analysis, and then the principal component was used as the input set of BP neural network. Based on the traditional neural network model, the combined model of principal component analysis and BP neural network was constructed to predict the annual output per unit area of C. oleifera seeds in four provinces. The training set and test set of BP neural network were selected from 1990to 2018, the prediction effect of the model was verified by 2019 data.【Result】The meteorological factors which had important contributions to the principal components were sunshine duration, temperature from June to November, amount of precipitation from March to May, minimum averaged temperature, dew point temperature, average wind velocity,maximum sustained wind speed, and sea-level pressure. The improved model took less iteration time and had higher fitting degree. The average relative error of the four provinces was less than 3%. By using the model, the yield of C.oleifera dry seed per unit area in Hunan, Jiangxi, Zhejiang and Guangxi in 2025 was predicted to be 0.831, 0.583, 0.449and 0.512 t/hm^(2) respectively.【Conclusion】Compared with the traditional prediction model, the prediction efficiency and accuracy of the improved principal component analysis and BP neural network combination model are improved, and there is a better development trend of C. oleifera production in the four regions in the future.