Cluster-based hyperparameter optimisation for LSTM-based flow forecasting in Canadian catchments

Improvements in large dataset availability and computing power have led to an increase in large-sample hydrological (LSH) studies. While these studies bring a breadth of new knowledge, they also introduce new challenges. One such challenge is the optimisation of model hyperparameters, which can be prohibitively computationally expensive on a large scale. Machine learning (ML)-based flow forecasting models have been steadily rising in popularity due to their high accuracy and ease of development. While traditional physics-based models have hyperparameters rooted in hydrological concepts (e.g., the number of hydraulic response units is determined based on spatial heterogeneity), ML-based models do not typically use a physical basis for selecting hyperparameters (e.g., neural network topology). Instead, ML model hyperparameters are typically determined using heuristic or exhaustive search methods.  Clustering has been previously applied to watersheds for identifying homogenous regions for flood frequency analyses. In these cases, unsupervised clustering, based on static watershed characteristics and flow statistics, is used to identify homogenous regions on which to conduct frequency analyses. We propose an application of clustering to optimise ML model hyperparameters on a large scale. The objective of this study is to determine whether grid-search optimisations are transferrable to similar catchments, identified through unsupervised clustering. Our study is conducted using a subset of Canadian catchments (n>500) from the HYSETS database. For each catchment, an LSTM is trained to forecast flow at a daily resolution using hydrometeorological input features (flow, precipitation, temperature, SWE). Grid-search hyperparameter optimisation is conducted on model architecture (number of hidden states and layers), learning rate, dropout rate, and input sequence length. We evaluate the effectiveness of cluster-based hyperparameter optimisation based on a comparison against a non-optimised baseline, for an increasing number of clusters. The impacts of this work have the potential to improve the effectiveness of ML-based flow forecasting models in cases where exhaustive hyperparameter searches are not possible. The results will also allow us to make recommendations for typical hyperparameter values based on watershed characteristics.