A model that turns rainfall and other weather inputs into a prediction of streamflow at a catchment outlet over time.
A rainfall-runoff model takes rainfall, and usually evaporation or temperature, as input and predicts the streamflow that results at a catchment outlet. These models are used for flood forecasting, filling gaps in flow records, design flood estimation, reservoir operation and studies of land use and climate change. There are four broad types. Empirical models link rainfall and flow directly, such as the rational method or the unit hydrograph. Conceptual models represent the catchment as a set of connected stores, or buckets, for snow, soil, groundwater and channel, with simple rules for how water moves between them; HBV, GR4J and the Sacramento model are well known examples. Physically based models solve equations for flow through soil and over land, and need much more data. Data-driven models, such as LSTM neural networks, learn the link from long records. Most conceptual models rest on a water balance for each store: change in storage equals inputs minus outputs. Take one soil bucket that holds 50 mm, where outflow each day is 10% of storage. On a day with 20 mm of rain and 3 mm of evaporation, outflow is 0.1 × 50 = 5 mm, and storage at the end of the day is 50 + 20 - 3 - 5 = 62 mm. Real models chain several such stores and add thresholds and routing. Models can run event by event, for a single storm, or continuously, keeping track of soil wetness between storms. Most need calibration against observed flow, then testing on a separate period. Common scores are the Nash-Sutcliffe and Kling-Gupta efficiencies. Limits come from the inputs as much as the model. Rain gauge networks often miss the true catchment rainfall, flow records carry rating curve errors, and parameters fitted to past conditions may not hold under a changed climate or land use.