基于有限标量量化与分阶段跨模态注意力融合的光伏功率预测

打开文本图片集

中图分类号：TP391.4；TM615 文献标志码：A 文章编号：1001-3695（2025）12-009-3594-08

doi：10.19734/j. issn.1001-3695.2025.05.0159

Photovoltaic power prediction based on finite scalar quantization and staged cross-modal attention fusion

Zhang Haiqing 1，2 ，Liu Jialing1，²，Tang Dan1†，Xiang Xiaoming³，Yang Dong4，Guo Benjun1，2 （1.ColegeffenduUesifu;cnee anceDesignandDetectionEngineringResearchCenter，Chengdu625，China；3.SichuanMeteoroogicalObseration&DataCte， Chengdu 610o72，China；4.SichuanMeteorological ServiceCenter，Chengdu 610072，China）

Abstract：Toaddress thelimitationsof multi-modal fusioninphotovoltaicpowerprediction，specificallythechalenges posed bymodalityheterogeneityindatarepresenationandinsuffcientcross-modalcorelationodeling，tispaperproposdamultimodal meteorologicaldatafusionmodel（MMDF）.The modelfirstlyappiedfinite scalarquantization（FSQ）toachieveunifiedrepresentationofheterogeneousdatasources，effectivelyovercomingthebotleneckofcross-modalinformationalignment andreducing computational complexityduring fusion.Then，itdesigneda hybrid encoding featureextraction module byintegratingtheglobalspatial modelingcapabityof the VisionTransformerwiththetemporaldynamicscaptureof theGRU-Linear architecture，whichsignificantlyenhancedthediscriminabilityofmulti-modal features.Furthermore，itconstructedastaged cros-modal fusion strategybasedonagateddynamicfusion mechanism，and employed CrossTransformer todeeply integrate time-seriesandcloud image features，therebycapturingcomplex inter-modalrelationships.Experimentalresultsshowthat， compared with the FusionSF algorithm，MMDF model improves MAE，RMSE，and R2 by 5.69% ，12. 44% ，and 4. 52% ，respectively，onaulti-modalsolarpowerdataset，providing bothatheoreticalbreakthroughandanewparadigmforengineering applications in photovoltaic power prediction under complex weather conditions.

KeyWords：photovoltaic power prediction；multi-modal fusion；information alignment；finite scalarquantization

0 引言

光伏功率预测的间歇性主要受云层动态等气象因素影响，需融合卫星云图与数值气象数据以提升预测精度[1]。（剩余19533字）