知识与财富的链接
将柑橘幼苗作为试验对象,利用传感器采集空气相对湿度和温度,以基质相对湿度、温度和EC值作为环境因子,采用称重法实时采集作物的质量变化量作为作物的蒸发量;以环境因子为模型输入,作物蒸发量为模型输出,构建长短期记忆神经网络(LSTM)预测模型,优化后的最优模型结构以及训练参数包括LSTM模型的隐藏层1层,隐藏层节点数为120,迭代样本数为128,训练迭代次数为175,网络的激活函数选择tanh函数,学习率为0.001,时间步长为72。LSTM预测模型的决定系数(R2)、均方根误差(RMSE)、平均绝对误差(MAE)分别为0.993 9、0.015 5 g、0.011 3 g;与循环神经网络(RNN)、门控循环单元(GRU)的预测效果进行对比,LSTM预测模型的预测蒸发值更接近真实蒸发值,预测结果相对误差范围波动最小,RMSE、MAE最小,R2最大,说明这3种模型中LSTM预测模型的预测效果最佳。
In this study, citrus seedlings were selected to estimate the predictions of evaporation. The air relative humidity and temperature were collected by sensors and mass method was used to collect the mass change of crops in real time as crop evaporation. The substrate relative humidity, temperature and EC value were used as environmental factors. With environmental factors as model input and crop evaporation as model output, a long short-term memory neural network(LSTM) prediction model was constructed. The optimized model structure and training parameters included 1 hidden layer of the LSTM model, 120 hidden layer nodes, 128 iteration samples, and 175 training iterations. The activation function of the network is tanh function, the learning rate was 0.001, and the time step was 72. The coefficient of determination(R2), root mean square error(RMSE) and mean absolute error(MAE) of LSTM prediction model were 0.993 9, 0.015 5 g and 0.011 3 g, respectively. Compared with the prediction effect of recurrent neural network(RNN) and gated cycle unit(GRU), the predicted evaporation value from LSTM prediction model was closer to the real evaporation value, and the relative error range of prediction results had the smallest fluctuation, RMSE and MAE were the smallest, and R2 was the largest, indicating that the prediction effect of LSTM prediction model was the best among these three models.