Hydro-Lstm: A Hydrological Approach To Lstm Machine Learning Based Modeling
Presenter: Luis De la Fuente1
Co-Author(s): Reza Ehsani, Hoshin V. Gupta
Advisor(s): Dr. Laura Condon
1Department of Hydrology and Atmospheric Sciences, University of Arizona
Panapto Presentation Video
Oral Session 2: Data-Driven and Physically-Based Modeling
Recent applications of the Long-Short Term Memory (LSTM) network-based machine-learning approach for streamflow prediction have demonstrated their ability to outperform traditional spatially-lumped process-based models. However, difficulties can arise when interpreting the internal processes and variables of the LSTM model, mainly due to the structural complexity of the network, which includes gating operations and internal sequential processing of the data. Utilizing isomorphic relationships between the LSTM model structure and the water budget rules for updating the state variables in a Dynamic Environmental System (DES), we propose and test a modification called Hydro-LSTM. This structure mimics many behaviors inherent in a DES such as its bottleneck information storage in a state variable and its sequential tracking of the state. Moreover, we have modified how data is fed to the new representation, away from the sequential method used in standard LSTM to allow simultaneous access to past input features in Hydro-LSTM, which thereby explicitly acknowledges the importance of the past data. We compare the Hydro-LSTM and LSTM architectures using data from ten different catchments from varied hydroclimatic conditions. Preliminary results indicate that Hydro-LSTM requires a smaller number of nodes (cell states or neurons) to obtain similar performance to that of a traditional LSTM model. Additionally, the weights/parameters associated with the input variables have a more direct interpretation when using the new structure. This study indicates the possibility of being able to add interpretability and extract useful information from the poorly named “black box” of Machine Learning.