基于长短时记忆的农作物生长环境数据预测

Prediction of crop growth environmental data using LSTM

针对传统温室农作物生长监控系统控制灵活性差且精确度低等问题,设计了一个面向智慧农业的农作物生长闭环监控系统.引入单变量长短时记忆(long short-term memory,LSTM)网络模型,对土壤含水率、土壤温度和土壤电导率3个农作物生长环境数据进行预测研究.在优化时间步长参数的基础上,分析不同预测步长对单变量LSTM模型预测准确性的影响,采用不同时间段的测试集数据对模型的预测性能和稳定性进行验证.分别采用单变量LSTM模型、最小绝对值收敛和选择算法、随机森林回归、双向LSTM模型和编解码LSTM模型进行预测对比,结果表明,单变量LSTM模型预测的平均绝对误差值和均方根误差值均为最小,模型具有更好的准确性和稳定性.本研究设计的农作物生长闭环监控系统能有效预测农作物的生长环境数据,为农作物监控系统的智能控制提供有效数据支撑.

Traditional greenhouse crop growth monitoring system has the disadvantages of poor control flexibility and low accuracy.To address these issues,this paper focuses on developing a closed-loop crop growth monitoring system for purpose of smart agriculture and employing the univariate long short-term memory(LSTM)prediction model for the crop growth environmental data prediction.Based on optimizing the time-step parameter,the prediction accuracy of the univariate LSTM prediction model conducted with different prediction steps was discussed.Then the model stability was verified by using the testing datasets of different time periods.Finally,the prediction results obtained from the proposed method were compared with those obtained from least alosolute shrinkage and selection operator(LASSO),random forest regression,bidirectional LSTM,and encoder-decoder LSTM.The experimental results show that the univariate LSTM has better prediction accuracy and stability in comparison with other prediction models.The designed closed-loop crop growth monitoring system can effectively predict the environmental data,which can provide effective data support for the intelligent control of crop monitoring system.