RiverFormer: Attention Based Operational Streamflow Forecasting

Operational streamflow forecasting requires models to translate historical meteorological conditions and future weather forecasts into discharge predictions. An important distinction from conventional time-series forecasting is that observed discharge is unavailable as an initial condition at inference time. Models must instead infer the evolving catchment state from historical meteorological forcings and static attributes by learning rainfall--runoff dynamics rather than extrapolating the observed target series. We introduce RiverFormer, an end-to-end, fully parallel Transformer architecture for this task. RiverFormer uses temporal patch tokenization to encode 365 days of historical meteorological forcings and conditions on seven days of numerical weather forecasts. It predicts the initial and complete seven-day discharge trajectory in a single forward pass, without recurrent state updates or autoregressive feedback between forecast leads. We train RiverFormer using a global set of gauges and compare it against published global operational baselines. RiverFormer matches or outperforms the recurrent baseline across forecast leads and geographic regions, demonstrating that fully parallel attention is a practical alternative for large-scale operational forecasting. By reducing temporal sequence length through patching and eliminating sequential decoding, RiverFormer provides a scalable foundation for future spatially distributed models operating directly on full meteorological grids.