基于Kolmogorov-Arnold网络的日光温室温湿度预测及最小建模数据量

张疏桐; 张传颂; 冯雄; 冯克鹏

文章摘要

张疏桐,张传颂,冯雄,冯克鹏.基于Kolmogorov-Arnold网络的日光温室温湿度预测及最小建模数据量[J].农业环境科学学报,2025,44(11):2835-2851.

基于Kolmogorov-Arnold网络的日光温室温湿度预测及最小建模数据量

Temperature and humidity prediction in solar greenhouses using Kolmogorov-Arnold networks and minimum modeling data volume

投稿时间：2025-05-20

DOI：10.11654/jaes.2025-0471

中文关键词: 日光温室温湿度预测随机森林（RF）循环神经网络（RNN）长短期记忆网络（LSTM） Kolmogorov-Arnold网络（KAN） SHAP

英文关键词: solar greenhouse temperature and humidity prediction RF RNN LSTM KAN SHAP

基金项目:清华大学-宁夏银川水联网数字治水联合研究院专项统筹重点项目（SKL-IOW-2023TC2301）

作者	单位	E-mail
张疏桐	宁夏大学土木与水利工程学院, 银川 750021
张传颂	宁夏大学土木与水利工程学院, 银川 750021
冯雄	宁夏大学土木与水利工程学院, 银川 750021
冯克鹏	宁夏大学土木与水利工程学院, 银川 750021 旱区现代农业水资源高效利用教育部工程研究中心, 银川 750021 宁夏节水灌溉与水资源调控工程技术研究中心, 银川 750021 宁夏大学干旱灌区水文与智慧水利野外科学观测研究站, 银川 750021	fengkp@nxu.edu.cn

摘要点击次数: 996

全文下载次数: 785

中文摘要:

针对日光温室在不同时间步长下温湿度预测精度低，不同温度条件下最小建模数据量不明确的问题，本研究引入Kolmogorov-Arnold（KAN）网络，与随机森林（RF）、循环神经网络（RNN）和长短期记忆网络（LSTM）3种机器学习算法比较，以确定不同时间步长下温湿度预测性能最优模型，然后通过对比不同建模数据量模型预测效果确定不同温度条件下的最小建模数据量，最后利用SHapley Additive exPlanations（SHAP）分析解释了各特征对4种模型预测结果的影响。结果表明：RF、LSTM和KAN模型均在15 min~1 h时间步长的温度预测中表现较好，R²均大于0.9，KAN模型在15 min~1 h时间步长的湿度预测中表现优秀，15min时的R²为0.82，均方根误差（RMSE）为0.14 kPa；1~3 d时间步长下RF、LSTM和KAN模型温度预测性能良好，R²均大于0.8，KAN模型在1 d时间步长下湿度预测效果良好，R²为0.62，7 d时间步长下各模型均无法准确预测温室温湿度。不同训练数据量的结果表明，仅需10 d数据即可构建精准的温度预测模型，湿度预测模型则至少需要20 d的数据。SHAP分析结果揭示，室外空气温度（ATO）是温度预测最重要的特征，室外空气湿度（VPO）则为湿度预测最重要的特征。研究表明，KAN网络在温室环境预测领域有广泛应用前景，尤其适用于湿度预测，在建模时仅需要10 d数据即可构建温室温度预测模型，20 d数据即可构建湿度预测模型，ATO和VPO分别对温室温度和湿度预测影响最大。

英文摘要:

To address the challenges of low prediction accuracy of temperature and humidity in solar greenhouses at different time steps and the unclear minimum data requirements under varying thermal conditions, this study introduces the Kolmogorov-Arnold network （KAN）and compares its performance with three machine learning algorithms: Random forest（RF）, Recurrent neural network（RNN）, and Long short-term memory（LSTM）. The objective is to identify the optimal model for temperature and humidity prediction across various time intervals. Subsequently, the minimum amount of training data required under different temperature conditions was determined by comparing the prediction performance of models trained on different data volumes. Finally, SHAP（SHapley Additive exPlanations） analysis was applied to interpret the influence of each input feature on the model outputs.The results show that: the RF, LSTM, and KAN models all exhibited favorable performance in temperature prediction at time steps ranging from 15 minutes to 1 hour, with R² values exceeding 0.9. The KAN model demonstrated superior performance in humidity prediction within the same interval, achieving an R² of 0.82 and an root mean square error（RMSE）of 0.14 kPa at the 15-minute time step. At time steps of 1 d to 3 d, all three models maintained good temperature prediction accuracy, with R² values above 0.8. The KAN model also performed well in humidity prediction at the 1 d time step, with an R² of 0.62. However, at the 7 d time step, none of the models were able to accurately predict greenhouse temperature and humidity. The minimum data requirement analysis revealed that only 10 days of data are sufficient for accurate temperature prediction, whereas humidity prediction requires at least 20 days of data. SHAP analysis indicated that outdoor air temperature（ATO）was the most important feature for temperature prediction, while outdoor vapor pressure（VPO）played the dominant role in humidity prediction. This study demonstrates that the KAN model holds significant potential for greenhouse climate prediction, particularly for humidity forecasting. Moreover, accurate models can be established with only 10 to 20 days of training data. Among all input features, ATO and VPO are identified as the most influential variables for predicting temperature and humidity, respectively.

HTML 查看全文查看/发表评论下载PDF阅读器