基于自注意力机制的双向门控循环单元和卷积神经网络的芒果产量预测

Prediction of mango yield based on self-attention CNN and bidirectional GRU

针对影响芒果产量的相关气象要素繁多,它们与产量之间的关联关系复杂、难以用数学函数准确地描述的问题,提出一种基于自注意力机制具有长短期记忆功能的双向门控循环单元和卷积神经网络组合(Self-attention CBiGRU)模型。首先,利用CNN卷积层(1D CNN)提取局部特征;其次将Self-attention机制用于进一步提取依赖特征,然后双向门控循环单元(BiGRU)会充分考虑年份之间的关联性,学习长期依赖特征;最后,利用广西某地3个气象站所收集到的24个芒果生产周期年份(从前一年第22旬到当年第21旬)每旬9个气象要素及芒果产量数据进行分析建模,建立了芒果产量预测Self-attention C-BiGRU模型。实验结果表明,Self-attention C-BiGRU模型预测的产量与实际产量的均方根误差为10.67,比支持向量回归(SVR)、误差后向传播神经网络(BPNN)、门控循环单元(GRU)、基于注意力机制的双向门控循环单元(BiGRU-Attention)、门控循环单元和卷积神经网络组合模型(GRU-CNN)、双向门控循环单元和卷积神经网络组合模型(C-BiGRU)分别平均降低了37.7%、42.1%、17.6%、4.1%、5.3%和5.9%。Selfattention C-BiGRU模型具有较高的预测准确性,对提升广西芒果产业发展、推进农业信息化有重要意义。

There are many related meteorological factors affecting mango yield and the relationship between them and the yield is complicated,which is difficult to describe accurately with mathematical functions.A Bidirectional Gated Recurrent Unit and Convolution Neural Network based Self-attention(Self-attention C-BiGRU)combined model with long short-term memory function was constructed to solve the problems.Firstly,the CNN convolutional layer(1D CNN)was used to extract local features.Secondly,the Self-attention mechanism was used to further extract the dependent features.Then BiGRU was used to fully consider the correlation between the years and learn the long-term dependent features.Finally,the Self-attention C-BiGRU model for mango yield prediction was established by using the data of meteorological factors and mango yield collected from three meteorological stations in somewhere of Guangxi.The data include 24 years of mango production cycle(from the 22nd ten days of the previous year to the 21st ten days of that year),and there are nine meteorological factors in per ten days.The experimental results show that the root mean square error between the actual yield and the predicted yield by the Self-attention C-BiGRU model is 10.67,which is 37.7%,42.1%,17.6%,4.1%,5.3%,5.9%lower than that of the Support Vector Regression(SVR),the Back Propagation Neural Network(BPNN),the Gated Recurrent Unit(GRU),the Bidirectional Gated Recurrent Unit based attention combined model(BiGRU-Attention),the Gated Recurrent Unit and Convolution Neural Network based combined model(GRU-CNN),the Bidirectional Gated Recurrent Unit and Convolution Neural Network combined model(C-BiGRU).The predicted value of Self-attention CBiGRU has high prediction accuracy and is of great significance to improve the development of Guangxi mango industry and promote agricultural informationization.