土壤温度时间序列预测的LSTM神经网络研究

LSTM neural network for edaphic temperature time series prediction

适宜的土壤温度是作物生长发育的重要环境因素,在作物全生育期都起着至关重要的作用,研究土壤温度的预测预报模型对农业生产具有重要意义。本试验以河北省石家庄市藁城区某试验田作为研究对象,对试验田的样本点10~30 cm深度处土壤温度的长期监测数据进行了拟合,分别建立基于LSTM神经网络的日均土壤温度预测模型和时均温度预测模型。结果表明,采用2-60-80-1网络结构的LSTM神经网络对日均土壤温度时间序列数据预测效果最优,其均方根误差(RMSE)达最小值0.603;采用2-60-80-1网络结构的LSTM神经网络对时均土壤温度时间序列预测效果最优。在对日均和时均土壤温度预测时,LSTM神经网络模型的平均的均方误差(RMSE)仅为0.665,较之BP神经网络模型降低了0.053,说明了LSTM神经网络模型用于土壤温度时间序列预测的优势,可满足土壤温度日常预报的需要。

Suitable soil temperature,as an important environmental factor of crop growth and development,played a key role in the whole growth period of crops.Therefore,it was of great significance to study the prediction model of edaphic temperature for agricultural production.Taking the experimental field in Gaocheng District,Shijiazhuang City,Hebei Province as the research object,the long-term monitoring data of soil temperature at the depth of sample points in the experimental field 10~30 cm were fitted.The prediction models of average daily soil temperature and time average soil temperature based on LSTM neural network were established,respectively.The results showed that the LSTM neural network model with the structure of 2-60-80-1 had the best effect when applied to the training simulation of the time series of daily mean edaphic temperature,and its root mean square error(RMSE)reached the minimum value of 0.603. The LSTM neural network model with the structure of 2-60-80-1 could achieve thebest effect when applied to the prediction of time series of average edaphic temperature. The average mean squareerror (RMSE) of the LSTM neural network model was only 0.665 when predicting the daily and hourly edaphictemperature, which was 0.053 lower than that of the BP neural network model. It showed that the LSTM neuralnetwork model was superior to the edaphic temperature time series prediction and could meet the needs of dailyedaphic temperature prediction.