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

打开文本图片集

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中图分类号：TP18;S625文献标志码：A文章编号：1672-2043（2025）11-2835-17doi：10.11654/jaes.2025-0471

Abstract：Toaddress thechallngesoflowpredictionaccuracyof tmperatureandumidityinsolar greehousesatdiferenttiesteps andtheunclearminimumdatarequrementsudervaryingtheralcondions，thisstudyintroducestheKolmogorov-Aroldnetwork （KAN）andcomparesitspformanewithtreemachnelaingalgorithsRandomforest（RF），Recurrnteuraletork，and Longshort-termmemory（LST）.Theojectiveistoidentifythoptialmodelfortemperatureandhumiditypredictioacrosvarious timeintervals.Subsequently，theminimumamountoftrainingdatarequiredunderdiferenttemperatureconditionswaseterinedby comparingthepredictionperformanceofmodelstrainedondiferentdatavolumes.Finally，SHAP（SHapleyAditiveexPlanations） analysiswasappliedtointerprettheifluenceofachinputfeatureonthemodeloutputs.Theresultsshowthat：theRF，LST，andKAN modelsall exhibited favorableperformanceintemperaturepredictionattime stepsranging from15minutes to1hour，with R2 values exceeding 0.9.TheKANmodeldemonstratedsuperiorperformanceinhumiditypredictionwithinthesameinterval，achevingan R2 of 0.82 （2号 andanrootmeansquareerRME）fO.14Paate15utetestep.Attiestepsof1dtod，alltheemodelsmataidgood temperature prediction accuracy， with R2 values above O.8.The KAN model also performed well in humidity prediction at the 1 d time step， with an R2 of 0.62.However，atthe7dtimestep，noneofthe models wreable toaccuratelypredict greenhouse temperatureand humidity. The minimumdatarequirementanalysisrevealedthatonly1Odaysofdataaresuffcientforaccuratetemperatureprediction，whereas humiditypredictionrequiresatleast2Odaysofdata.SHAPanalysisindicatedthatoutdoorairemperature（ATO）wasthemostiportant feature fortemperatureprediction，whileoutdoorvaporpressure（VPO）playedthedominantroleinhumiityprediction.Thisstudy demonstratesthattheKANmodelholdssignificantpotentialforgrenhouseclimateprediction，particularlyforhumidityforecasting. Moreover，acurate modelscanbeestablishedwithonly10to2Odaysoftrainingdata.Amongallinputfeatures，ATOandVPOare identified as the most influential variables for predicting temperature and humidity，respectively.

KeyWords：solar greenhouse; temperature and humidity prediction; RF; RNN; LSTM; KAN; SHAP

温室种植因具有环境稳定、产量高的特点近年来快速发展，日光温室是其中的一种重要形式。（剩余22798字）