The　measurement　of　soil　thermal　conductivity　plays　a　key　role　in　the　soil　heat　source　pump　air　conditioning　system　in　the　design　area.　In　order　to　solve　the　problem　that　the　measurement　period　of　traditional　soil　thermal　conductivity　measurement　is　too　long,　a　prediction　model　of　soil　thermal　conductivity　based　on　improved　BP　neural　network　learning　algorithm　is　proposed.　By　collecting　the　average　temperature　in　the　U-pipe　after　50　hours　in　different　parts　of　the　country,　TRAINSCG　(quantized　conjugate　gradient　method),　TRAINGD　(gradient　descent　training　function)　and　TRAINRP　(elastic　gradient　descent　method)　were　used　to　optimize　and　improve　the　prediction　model　to　obtain　the　final　prediction　model　of　soil　thermal　conductivity,　and　the　accuracy　of　the　three　improved　algorithms　was　calculated　by　testing　samples.　The　results　show　that　the　BP　neural　prediction　model　using　TRAINSCG　learning　algorithm　has　the　best　prediction　effect　on　soil　thermal　conductivity,　the　BP　neural　prediction　model　using　TRAINSCG　learning　algorithm　has　a　better　prediction　effect　on　soil　thermal　conductivity,　and　the　BP　neural　network　prediction　model　using　TRAINGD　learning　algorithm　has　a　poor　prediction　effect　on　soil　thermal　conductivity.　©　2021　IEEE.