An LSTM rainfall-runoff model is a recurrent neural network trained on past weather and flow records to predict river discharge.
An LSTM rainfall-runoff model uses a long short-term memory network, a type of recurrent neural network, to predict streamflow from a sequence of weather inputs such as rainfall, temperature and evaporation. The network has internal memory cells with gates that decide what to keep, add and forget at each time step. This lets it learn slow processes such as snow storage, soil moisture and groundwater release, which simpler neural networks handle poorly. The model reads, for example, the past 365 days of daily weather for a catchment and outputs today's discharge. It is trained by comparing its output with observed flow over many years and adjusting its weights to reduce the error. Static catchment attributes such as area, slope, soil type and aridity can be added as inputs, so that one model learns across many catchments. Work by Kratzert and colleagues in 2018 and 2019 on hundreds of US catchments in the CAMELS dataset showed that a single LSTM trained across many basins could match or beat well-calibrated conceptual models on standard scores such as Nash-Sutcliffe efficiency, including in basins left out of training. This helped make LSTMs a common benchmark in hydrology. What it needs: long, good-quality records of weather and flow, ideally many years across many catchments. A short record from a single gauge gives a weaker model. Limits: a standard LSTM has no built-in mass balance, so its flows need not add up to the water that came in. It can extrapolate badly to conditions outside its training data, such as a record flood, a new dam upstream or a changed climate. It does not explain why it predicts what it does, and its internal states are not the same as soil moisture or storage, though studies have found links. Test it on held-out years and on extremes before use.