This study investigates the impact of dynamical representational error (RE) on the analysis of an ocean ensemble Kalman filter-based data assimilation system, LETKF-ROMS (Local Ensemble Transform Kalman Filter - Regional Ocean Modeling system) configured for the Indian Ocean and assimilating in-situ temperature and salinity observations from Argo. Three different approaches to account for the RE are studied and inter-compared: (i) static RE (varies in horizontal and vertical direction), (ii) dynamic RE (varies in space and time) estimated from concurrent observations, and (iii) dynamic RE estimated using concurrent high resolution model outputs. RE estimated from the model outputs exhibits rich spatial and temporal variability with an estimated temporal mean RE for temperature below 0.5 °C and 0.2 °C in the surface and deep layers, respectively, and reaching up to 1°C in the thermocline layers. The region encompassing the Great Whirl displays a large seasonal variability reaching up to 0.8°C, and the South Equatorial Current (SEC)a large inter-annual variability reaching up to 0.4°C. Neglecting such spatio-temporal variations of RE and assimilating with a static RE limited the benefits of assimilation by entertaining over-fitting issues that caused degradations in the Bay of Bengal, the western parts of the Arabian Sea, and the equatorial Indian ocean. Assimilating with the observations-based dynamic RE improved the results in these regions, but the best performances were obtained with the configuration using the model-based dynamic RE, which yielded further improvements (e.g. reduction of sea surface height root-mean-square-errors reaches 30% with respect to the observations-based dynamic RE). The latter also better handled the rich spatial variability regions and areas not well sampled by the observations. Improved estimates of the spatial and temporal variations of RE helped to better exploit the assimilated observations and provided enhanced analyses less prone to assimilation shocks.