关键词: 芒果, 产量预测, Self-attention机制, 双向门控循环单元, 卷积神经网络, 循环神经网络

Abstract: 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.

Key words: mango, yield prediction, Self-attention, Bidirectional Gated Recurrent Unit (BiGRU), Convolution Neural Network (CNN), Recurrent Neural Network (RNN)