基于长短期记忆的柑橘园蒸散量预测模型

Modeling on Prediction of Evapotranspiration of Citrus Orchard Based on LSTM

传统的柑橘灌溉方式主要依赖人工经验,一方面有可能导致灌溉时机不准确,另一方面有可能造成灌溉量过高或者过低,对果实的生长都会产生负面影响。柑橘果园水分蒸散量是表征耗水量的重要指标。为了实现对大面积柑橘果园蒸散量(Evapotranspiration,ET)的准确估算,制定更加科学精细化的灌溉策略,基于气象数据集,应用长短期记忆(Long short-term memory,LSTM)、极限学习机(Extreme learning machine,ELM)和广义回归神经网络(General regression neural network,GRNN)方法对蒸散量建立预测模型并验证其准确性。结果表明,LSTM模型的平均绝对误差(Mean absolute error,MAE)和均方根误差(Root mean square error,RMSE)是3种模型中最优的,ELM和GRNN模型的性能接近。为了估算3种模型结果的可信度,在训练时加入了蒙特卡洛不确定性分析方法。结果表明,LSTM模型在不同输入特征数量下具有较高的精度,而ELM模型存在预测值偏高的现象,GRNN模型则偏低。

Citrus is an important fruit and it’s strongly relevant between quality and irrigation.Traditional irrigation strategies relying on human experience caused two problems,i.e.inaccurate irrigation timing and quantity.Both of the two problems have negative influence on citrus.The evapotranspiration of citrus orchard is an important index of water consumption.In order to evaluate citrus orchard evapotranspiration(ET)to make more scientific and precise irrigation strategies,the long short-term memory(LSTM),extreme learning machine(ELM)and general regression neural network(GRNN)methods were applied to model ET and test its performance based on climatic data.The result showed that LSTM performed the best in mean absolute error(MAE)and root mean square error(RMSE)than the other two models.And ELM model performed closely to GRNN.In order to evaluate the certainty of three models,the Monte Carlo analysis method was added to the process of training.The result indicated that LSTM had good accuracy in different input features while ELM tended to overestimate ET and GRNN tended to underestimate ET.It’s practical to applicate the proposed method to make precise irrigation strategies.