基于NARX神经网络的基质草莓腾发量超前预测OA

To address the limitations of traditional models in achieving lead-time prediction of crop evapotranspiration(ET),the lead-time prediction capability and inherent limitations of the conventional NARX neural network for ETof substrate-grown strawberries were systematically investiga-ted.An enhanced NARX-based architecture incorporating a delay elimination technique for single-step-ahead prediction was proposed.The automated greenhouse system continuously monitored microclimate parameters,including greenhouse temperature,humidity,and photosynthetically active radiation(PAR),as well as gravimetric data.ETwas quantified using the water balance method.Data from the fruit expansion to maturity stage were selected,and the original time series was segmented into six time intervals(1,2,3,4,6,24 hours)for modeling and analysis.The results show that the NARX open-loop model only establishes a dynamic baseline mapping between microclimate parameters,ETat time t and ETat the same time step t,and therefore lacks lead-time prediction capability.The NARX closed-loop model exhibits pronounced initial-state sensitivity and error accumulation effects during multi-step forecasting.Furthermore,its dependence on future environmental inputs substantially amplifies predic-tion uncertainty.The prediction error exhibits a supralinear growth with extended time intervals,and the 1-hour step size(RMSE=3.31 g)best matches the dynamic drip irrigation demand for strawberries.The NARX single-step prediction model achieves robust forecasting with only 6 historical observations.During peak transpiration periods,it maintains stable errors of 5.2%±4.5%,outperfor-ming BP and TDNN neural network models of comparable parameter scale.Therefore,the NARX sin-gle-step-ahead prediction model can provide reliable technical support for intelligent irrigation of sub-strate-grown strawberries.