基于IAO算法的LSTM改进策略及在葡萄产业时序预测中的应用

打开文本图片集

中图分类号：S9；S24 文献标识码：A 文章编号：2095-5553（2026）03-0148-09

LSTM improvement strategy based on IAO and application to time series prediction in grape industry

Feng Jianying，Li Zihan，He Miao，Wang Siwen，Tian Dong （School of Information and Electrical Engineering， China Agricultural University，Beijing，looo83，China）

Abstract：TraditionalLSTMmodelsoftenfacechalengessuchascomplexparametertuning，susceptibilitytolocaloptima， anddiicultyinfullcapturingthecharacteristicsofcomplexdata.Toaddresstheselimitations，thisstudycariedouttefollowing work：First，theAquila Optimization（AO）algorithm，which ispronetogeting trappdinlocaloptima，wasimproved by introducingawhale-inspiredspiralsearchstrategyandanadaptivet-distributionmutationstrategy.Theseenhancements significantlystrengthenedtheglobalsearchcapabilityanditerationspeedoftheimprovedAOalgorithm（IAO）.Second，toimprove theinterpretabilityoftheLSmodelandmetthenedforxtractingultipledatafeatures（seasonality，periodity，olieaity etc.），aC—ALSTMunivariatetimeseriesforecastingmodelwasproposedFolowinga“decompositiointegration”strategy，the timeseries datawere first decomposed intomultiplesub-seriesusing the CEEMDAN method.Anattention mechanism was thenintroducedtoincreasetheweightofimportantinformation，andtheIAOalgorithmwasusedtooptimizemodelparameters forsubsequenceprediction.Finall，themodel’seffectivenesswasvalidatedusingfourpubliclyavailabledatasetsand two self-constructed datasetsongrape pricesandlogisticsenvironmentalfactors.ExperimentalresultsshowedthattheC一 AILSTM model achieved R2 values of 0.899 5，0.9620，0.9533，and O.958Oon the publicly available datasets.For the self-constructed dataset， the R2 values reached 0.940 1，0.977 9，and 0.978 3. The prediction accuracy and error metrics aresignificantlyimprovedcomparedwiththe standardLSTMmodel，aresultfurtherconfirmed throughcomparisonswith variousLSTMvariants.These findings demonstrate thatthe proposed modelcanaccurately forecastgrapeindustrytime series data， providing meaningful support for price regulation and production decision-making within the grape industry.

Keywords：grape industry；timeseries；timeseries prediction;long short-term memory networks （LSTM）；atention mechanisms

0 引言

时间序列数据普遍存在于自然和社会中，时间序列预测对辅助决策、资源优化及止损措施具有基础且重要的价值。（剩余11730字）