基于PINN的温室黄瓜蒸腾量估算方法研究

PINN-based Estimation Method for Transpiration Rates of Greenhouse Cucumbers

摘要:

植物蒸腾量（ETc）的准确估算对于温室精准灌溉和水分管理至关重要。传统彭曼-蒙特斯方程（Penman-Monteith）虽已包含净辐射（Rn）、温度（T）、湿度（RH）、风速等关键驱动因子，但在温室特殊环境下，方程结果存在误差，作物系数（Kc）难以确定；而纯数据驱动模型缺乏物理机制约束，泛化能力有限。为解决这一问题，本文提出一种基于物理信息神经网络（PINN）的温室黄瓜ETc估算新方法。以改进的彭曼公式作为物理硬约束，构建包含物理机理分支和数据驱动修正分支的并行混合网络结构，通过物理一致性损失函数作为软约束同时优化数据拟合精度和物理一致性。在玻璃温室开展试验，采集净辐射、空气温度、相对湿度及蒸腾量实测数据。结果表明，与全连接神经网络（FCNN）基准模型相比，PINN模型决定系数R2从0.7847提升至0.8377，均方根误差（RMSE）为0.4596 mm/d，降低13.17%，平均绝对误差为0.1923 mm/d，降低35.12%。研究结果表明，将物理先验知识融入神经网络，不仅提高了ETc估算精度，也使结果更符合物理机制，增强了模型可解释性与可靠性，从而为智慧农业精准灌溉决策提供了新的技术方案。

Abstract:

Accurate estimation of plant transpiration (ETc) is crucial for precise irrigation and water management in greenhouses. While the traditional Penman-Monteith equation incorporated key drivers such as net radiation(Rn), temperature(T), humidity(RH), and wind speed, its results exhibited errors under greenhouse conditions, making crop coefficients (Kc) difficult to determine. Conversely, purely data-driven models lacked physical mechanism constraints and exhibited limited generalization capabilities. To address this challenge, a novel method for estimating greenhouse cucumber transpiration was proposed by using a physics-informed neural network (PINN). This approach employed an enhanced Penman formula as a hard physical constraint, constructing a parallel hybrid network architecture with both a physics-based branch and a data-driven correction branch. A physics consistency loss function served as a soft constraint, simultaneously optimizing data fitting accuracy and physical consistency. Field experiments were conducted in glass greenhouses in Zhenjiang, Jiangsu Province, collecting measured data on net radiation, air temperature, relative humidity, and transpiration. Results indicated that compared with the fully connected neural network (FCNN) baseline model, the PINN model improved the coefficient of determination R2 from 0.7847 to 0.8377, reduced the root mean square error (RMSE) by 13.17% to 0.4596 mm/d, and decreased the mean absolute error (MAE) by 35.12% to 0.1923 mm/d. It was demonstrated that integrating physical prior knowledge into neural networks not only enhanced the accuracy of transpiration estimation but also made the results more aligned with physical mechanisms. This improved the model's interpretability and reliability, thereby providing a technical solution for precision irrigation decision-making in smart agriculture.