A large sample study of the effects of upstream hydrometeorological input features for LSTM-based daily flow forecasting in Canadian catchments

Recent years have seen an increase of deep learning applications for flow forecasting. Large-sample hydrological (LSH) studies typically try to predict the runoff of a catchment using some selection of hydrometeorological features from the respective catchment. One aspect of these models that has received little attention in LSH is the effect that data from upstream catchments has on model performance. The number of available and stations and distance between stations is highly variable between catchments, which creates a unique modelling challenge. Existing LSH studies either use some form of linear aggregation of upstream flows as input features or omit them altogether. The potential of upstream data to improve the performance of real-time flow forecasts has not yet been systematically evaluated on a large scale. The objective of our study is to evaluate methods for integrating upstream features for real-time, data-driven flow forecasting models. Our study uses a subset of Canadian catchments (n>150) from the HYSETS database. For each catchment, long-short term memory networks (LSTMs) are used to generate flow forecasts for lead times of 1 to 3 days. We evaluate methods for identifying, selecting, and integrating relevant upstream input features within a deep-learning modelling framework, which include using neighbouring upstream stations, using all upstream stations, and using all stations with embedded dimensionality reduction. Early results indicate that while the inclusion of upstream data often yields improvements in model performance, including too much upstream information can easily have detrimental effects.