Comparing deep-learning and semi-distributed models for flow forecasting at fine spatial and temporal resolutions: a case study of 40 urbanized catchments in Toronto, Canada.

Flood early warning systems rely on accurate streamflow forecasts. Deep learning based approaches have been widely shown to outperform traditional, process-based approaches. While literature is rich with comparisons between these opposing modelling paradigms, most comparisons have been conducted at daily temporal resolutions and feature spatially coarse (i.e., lumped) process-based models. Flood forecasting applications, especially those in flashy urban catchments, rely on sub-daily forecasts. In this work, we compare the performance of a state-of-the-art regionally trained LSTM models with semi-distributed StormWater Management Models (SWMM) at temporal frequencies ranging from 15-minutes to 1-day, for roughly 40 highly urbanised catchments in Toronto, Canada. Results show that the LSTM approaches struggle at fine temporal resolution and when limited observed data is available. In contrast, SWMM models can be automatically parameterized and calibrated using comparatively much less data. While the amount of available historical data would be enough to train deep learning models at a daily resolution, it is insufficient to train hourly models, which we attribute to the comparatively more complex urban rainfall-runoff system. Potential solutions to this problem include model transfer between space and different temporal frequencies. Finally, another contribution of this work is LSTM hyperparameter optimization, which is not widely documented at a sub-hourly resolution. Results from this research reaffirm the need for multi-model approaches for flood forecasting, particularly in urbanised catchments.