High　impedance　faults　(HIFs)　are　difficult　to　detect　because　of　low　amplitude　of　the　current　signal.　The　interference　from　switching　cases　also　jeopardizes　the　reliability　of　HIF　detection　(HIFD).　Moreover,　a　long　detection　time　is　more　likely　to　cause　accidents.　To　diagnose　HIFs　promptly　and　precisely,　a　time-adaptive　(TA)　HIFD　model　is　proposed.　Firstly,　the　zero-sequence　current　data　of　the　faulty　feeder　are　processed　into　a　variable　length　training　set　to　train　gated　recurrent　unit　(GRU)　networks.　Secondly,　a　cost-sensitive　method　employed　to　train　two　biased　GRU　models　with　contrary　preference.　Then　the　two　models　are　combined　into　high　reliability　evaluation　model.　The　predicted　reliability　depends　on　the　consistency　of　predicted　results　of　the　two　models.　Reliable　results　are　output,　while　the　unreliable　results　are　set　aside.　To　prevent　untimely　detection,　an　equal　GRU-based　model　is　activated　when　reaching　the　time　threshold.　Delayed　judgment　improves　accuracy　of　HIFD　and　reduces　probability　of　harm　caused　by　untimely　detection.　The　performance　of　proposed　method　validated　by　simulated　data,　and　tested　in　a　realistic　10　kV　distribution　network　experimental　system.　In　the　true　type　10kV　system,　the　TA　HIFD　model　can　achieve　an　accuracy　of　96.73%　with　average　detection　time　of　4.315ms.