Conceptual hydrological models are widely employed for streamflow simulation due to their simplified representation of physical processes and computational efficiency. While effective in operational and planning contexts, these models often fail to capture event-scale runoff dynamics, particularly during complex storm events. Such limitations typically result in systematic errors in peak magnitude, timing, and hydrograph shape. This study introduces an event-based hybrid modelling framework that integrates the Unified River Basin Simulator (URBS), a well-established conceptual model in Australia, with a Long Short-Term Memory (LSTM) neural network. The LSTM was implemented as a residual post-processor, trained to predict and correct systematic errors between observed and URBS-simulated discharge while retaining the model’s physical foundation. The framework was applied to the Bulimba Creek catchment in Brisbane using 15-min rainfall and discharge data. Ten historical storm events were employed for cross-validation, with a separate multi-peak event reserved for independent testing. Across the development events, the hybrid model improved Nash-Sutcliffe Efficiency (NSE) from 0.34–0.87 (URBS) to 0.79–0.97, increased Kling–Gupta Efficiency (KGE) by 0.20–0.40, and reduced Root Mean Squared Error (RMSE) by 40–70%. For the independent event, NSE increased from 0.88 to 0.97 and volume bias improved from −28% to −5%. The framework also produced calibrated Monte Carlo dropout uncertainty bands, providing reliable confidence intervals around flood forecasts. The results demonstrate that residual learning can substantially enhance hydrograph accuracy while maintaining interpretability. Importantly, this work shows, for the first time in an event-based, high-resolution setting, that hybrid residual learning can strengthen operational flood forecasting in Australian catchments.