Recent works have shown the predominance of Deep Learning models, particularly LSTM, over traditional rainfall-runoff models [1]. However, while operational hydrology requires accurate discharge forecasts, these models are predominantly designed for simulation. As a drawback, they are also limited of performing data assimilation or persistence analysis, which remain crucial for effective forecast and analysis. Therefore, we propose, using three DA techniques, to separately integrate the benchmark models LSTM [1] and SACSMA [2] into an orchestrator to provide forecasts. Using the CAMELS dataset [3], the DA technique's added-values are assessed comparing to original benchmarks on 3 levels of lead time. The results indicate significant improvement, and remain relevant to the classical conclusions on DA expectations.