ABSTRACT With climate change, the accurate prediction of river water levels has become increasingly critical, particularly in confluence areas where multiple tributaries merge, resulting in complex hydrodynamic interactions. This study evaluates direct prediction (DP) and recursive prediction (RP) using a virtual sensor approach, with a focus on the role of forecasted rainfall. An LSTM model was trained using upstream rainfall and water level data to predict downstream levels, and its performance was assessed across various lead times (LT) using MAE, RMSE, NSE, and QER. For a short event (Event #8), DP without forecasted rainfall achieved an NSE of 0.42 at LT = 12 h, while RP dropped to −11.69. With forecasted rainfall, RP improved, maintaining an NSE of 0.75 compared to DP's 0.51. For a long multi‐peak event (Event #9), RP with forecasted rainfall achieved NSE values of 0.98 at 1 h and 0.91 at 12 h, outperforming DP (0.97 at 1 h, 0.42 at 12 h). These results demonstrate that DP is more reliable when forecasted rainfall is unavailable, whereas RP becomes superior when such data are available. Overall, the study highlights the potential of virtual sensors to enhance flood forecasting and disaster preparedness in confluence zones lacking direct monitoring stations.