Large-scale analysis of machine learning based flow forecasting models for Southern Ontario

The effectiveness of machine learning models for flow forecasting has been widely featured in academic research. These studies typically aim to demonstrate the applicability of various machine learning approaches to forecast flow in a small number of watersheds. This has resulted in a gap in knowledge in the efficacy of some machine learning techniques across multiple watersheds with distinct characteristics. This gap in knowledge, combined with improvements in computational power, there are increasing calls for increasing the scale of these studies that investigate machine learning applications in hydrology. We present a large-scale comparison of daily flow forecasting models for several catchments across Southern Ontario. Localised models are automatically parameterised on a catchment basis using openly available hydrometeorological timeseries data. The number of monitoring stations and length of time series is highly variable between catchments. To better understand data requirements of these models, we evaluate model performance across varying amounts of exogenous data, training observations, and temporal trends in test performance. Next, we compare model performance across spatiotemporal watershed characteristics, such as flow seasonality, landcover, and total annual rainfall, in order to relate machine learning outcomes with hydrological processes. Early results indicate two clusters in performance, one with strong performance (Nash-Sutcliffe Efficiency > 0.80) and another, with fair performance (0.80 > NSE > 0.5). The cluster performances are closely correlated with flow seasonality; highly seasonal catchments tend to have stronger performance, but an overreliance on autoregressive input features and poor performance relative to the naïve model.