Value of process understanding in the era of machine learning: A case for recession flow prediction

Abstract

Streamflow is one of the key variables that fulfils various anthropogenic as well as natural functions. Accurate prediction of streamflow poses significant challenges, especially during recession periods when the streamflow continuously decreases with time. Consequently, it becomes imperative to develop specialized models specifically tailored to the accurate prediction of recession flow. Continuous advancements in machine learning (ML) algorithms led to their incorporation in hydrological modeling with an aim to improve hydrologic prediction. Few recent studies have demonstrated the superior performance of ML models over conceptual hydrological models, reigniting the old debate: do we need to understand hydrological processes? In this study, we attempted to answer the question if process understanding is important for ML models. The other objective was to investigate the usefulness of process information for different ML algorithm-based model structures. To answer these questions, we performed a model inter-comparison study in 252 basins in the United States using a conceptual power-law regression model (PLR), two ML models, viz. Artificial Neural Networks (ANN) and Long Short-Term Memory (LSTM) Networks, and their process-informed versions. Each model was allowed to exploit the information from the past 120 days of daily streamflow data to predict recession flow. Results show that the performance of PLR (median NSElog = 0.89) is better than both the ML models, ANN (median NSElog = 0.33) and LSTM (median NSElog = 0.80), during the low flows. Upon incorporating process-informed inputs, both the artificial neural network (ANN) and long short-term memory (LSTM) models exhibited noteworthy enhancements in their performance. Specifically, the NSElog for the ANN model increased from 0.33 to 0.66, while the LSTM model witnessed an improvement from 0.79 to 0.83. These findings underscore the efficacy of leveraging process understanding to bolster the predictive capabilities of both ANN and LSTM. In conclusion, our study emphasizes the significance of process understanding in the machine learning era.