An　accurate　and　reliable　multi-step　prediction　model　for　dam　displacement　prediction　can　provide　decision-　makers　with　crucial　forecast　information,　mitigating　safety　risks　associated　with　abnormal　displacements.　However,　conventional　deep　learning　models　overlook　the　variability　exhibited　by　the　lag　effect　of　dam　displacement　in　continuous　time　domains.　The　detrimental　effect　of　this　limitation　on　model　performance　is　further　amplified　in　the　dam　displacement　multi-step　prediction　scenario.　To　address　this　issue,　this　paper　proposes　a　sequence-to-sequence　Chrono-initialized　long　short-term　memory　(S2S-CLSTM),　which　can　take　into　account　the　dynamic　nature　of　dam　displacement　lag　effect　in　explaining　the　nonlinear　relationship　between　displacement　and　complex　external　features.　Specifically,　we　redefine　LSTM　initialization　using　the　Chrono　initializer　and　construct　a　sequence-to-sequence　model　(S2S-CLSTM)　based　on　the　Chrono-initialized　Long　shortterm　memory　network　(CLSTM).　CLSTM　adapts　to　dynamic　displacement　lag　effect　by　incorporating　temporal　dependency　information　in　implicit　parameters,　enhancing　model　generalization.　The　S2S　paradigm　aids　in　maintaining　robustness　while　interpreting　dam　displacements　across　different　time　scales.　Secondly,　the　study　introduce　a　special　Re-optimization　strategy　tailored　for　S2S-CLSTM　to　mitigate　performance　degradation　caused　by　ambiguous　definitions　of　displacement　lag　extents.　Using　monitoring　data　from　a　real　arch　dam,　the　effectiveness　of　model　and　method　is　verified.　Within　the　prediction　step　range　of　3　to　7,　the　S2S-CLSTM　achieves　an　impressive　average　R2　of　0.764,　with　MAE　and　RMSE　values　of　only　0.623　mm　and　0.797　mm,　respectively.　In　addition,　this　work　also　explores　the　influence　of　the　dam　section　location　on　the　displacement　lag　effect　model　to　emphasize　the　importance　of　the　proposed　methods　in　dam　displacement　multi-step　prediction.