Physics-informed LSTMs improve streamflow prediction when storage–discharge dynamics are temporally resolved

Alberto Ardid , David Dempsey, Celine Cattoën, markus pahlow

Flood prediction at sub-daily resolution requires models that are accurate and physically plausible beyond observed conditions. We augment a per-basin Long Short-Term Memory network with a linear reservoir constraint (Q = KS), embedding storage-discharge dynamics as a soft physics loss, and evaluate it in hourly and daily streamflow-prediction experiments across New Zealand and Great Britain. At hourly resolution, the physics-informed LSTM improves performance in 99% and 91% of catchments, respectively, increasing median NSE from 0.30 to 0.44 in New Zealand and from 0.61 to 0.70 in Great Britain while learning physically plausible recession timescales. The largest gains occur where test-period flood recessions differ most from training conditions, indicating that the constraint is particularly useful for hydrological behaviour that is weakly represented during training. However, these benefits largely disappear at daily resolution, with improvements observed in only 14% and 21% of catchments in New Zealand and Great Britain, respectively. At hourly resolution, recession limbs span multiple timesteps and directly inform the storage–discharge constraint, which learns physically plausible effective recession timescales (median ~11.5 hours in New Zealand) that vary among catchments independently of catchment size; daily aggregation compresses these dynamics and substantially weakens that information. These results show that the value of physics-informed learning depends on whether the constrained physical process is resolved at the modelling timestep, and suggest that temporal resolution should be considered explicitly when designing and evaluating physics-informed hydrological models.