Accurate　multi-energy　load　forecasting　is　prerequisite　for　achieving　balance　between　supply　and　demand　in　building　energy　system.　The　continuous　development　of　building　flexibility　control　techniques　has　led　to　increased　complexity　and　variability　in　characteristics　of　multi-energy　flexibility　loads　under　application　of　various　building　flexibility　control　strategies.　Existing　load　forecasting　methods　lack　the　ability　to　recognize　the　flexibility　features　and　face　challenges　in　considering　complex　coupling　relationships　and　variable　characteristics　of　multi-energy　flexibility　loads.　To　address　this　gap,　we　propose　a　multi-energy　flexibility　load　forecasting　method,　denoted　as　(BFFR-MTL-LSTM),　which　incorporates　building　flexibility　feature　recognition　(BFFR)　and　multi-task　learning　(MTL)　model　utilizing　Long　Short-Term　Memory　(LSTM)　neural　networks　as　the　shared　layer.　A　novel　ToU-K-means　method　combining　ToU　with　K-means　algorithm　is　proposed　firstly　as　BFFR　technique　to　accurately　recognize　the　flexibility　features　under　various　energy　consumption　patterns.　The　forecasting　accuracy　of　proposed　method　is　validated　using　actual　operational　data　under　various　energy　consumption　patterns　with　average　R2　of　0.973.　Additionally,　through　comparisons　with　numerous　existing　models,　the　proposed　method　demonstrates　a　forecasting　accuracy　improvement　ranging　from　33　%　to　47　%.　This　validation　supports　that　the　proposed　method　adeptly　recognizes　energy　consumption　patterns,　resulting　in　more　accurate　forecasts　of　multi-energy　flexibility　loads.