Demand　responsive　transit　(DRT)　is　expected　to　offer　enormous　possibilities　for　fulfilling　the　ever-growing　diversified　demand　while　promoting　the　urban　sustainability.　Nevertheless,　it　seems　to　remain　critical　but　intractable　to　make　the　routing　decision　in　response　to　the　time-dependent　travel　speeds.　Considering　the　risk　of　speed　uncertainty,　deep　reinforcement　learning　(DRL)　algorithm　is　presented　to　address　the　time-dependent　electric　DRT　problem　in　this　study.　Long　short-term　memory　(LSTM)　is　integrated　into　the　attention　mechanism　at　the　decoding　process.　To　evaluate　the　sustainability,　the　reward　function　can　be　subdivided　into　the　number　of　served　passengers,　regret　utility,　and　carbon　emissions.　To　testify　the　effectiveness,　we　made　a　comparison　between　the　CPLEX　solver,　NSGA-II,　and　the　proposed　algorithm　in　a　realistic　transportation　network　of　Beijing.　The　computational　results　demonstrate　that　DRL　algorithm　with　shorter　computation　time　and　better　solutions　is　dramatically　superior　to　the　other　approaches.　Therefore,　the　DRL　algorithm　provides　a　more　efficient　framework　for　addressing　the　time-dependent　electric　DRT　problem　while　improving　the　sustainability　of　the　environment　and　society.