A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Water Engineering and Management

Thesis (M. Eng.) - Asian Institute of Technology, 2025

Hydrological models increasingly adopt deep learning architectures such as Long Short-Term Memory (LSTM) networks for streamflow forecasting. Despite their high predictive accuracy, basin-specific implementations often exhibit limited generalization and seldom incorporate physical catchment characteristics. Addressing this limitation, this study explores multiple LSTM-based configurations{u2014}both per basin and regional, and with and without catchment attributes. A key emphasis is placed on the Entity-Aware LSTM (EA-LSTM), an extension of the standard LSTM that integrates static descriptors through the input gate mechanism, enabling physically informed generalization across basins. Seven configurations were evaluated across six sub-basins of the Chao Phraya River (Chao, Pasak, Nan, Ping, Wang, and Yom), representing diverse hydrological and physiographic conditions. The regional EA-LSTM, conditioned on attributes such as slope, potential evapotranspiration, urban area, and erosion, generally outperformed per-basin models, particularly in structurally complex or data-sparse basins. However, performance improvements varied across basins, underscoring the interaction between model architecture and catchment behavior. To enhance interpretability, SHAP-based feature attribution, embedding analysis, and input gate bias evaluations were applied. These analyses revealed that the model captures hydrologically meaningful relationships{u2014}such as the roles of evapotranspiration, soil texture, erosion, urbanization, etc. in regulating flow. The learned patterns correspond closely to established physical processes, including runoff delay and flow regulation, offering insights into the model{u2019}s basin-specific responses. This work highlights the potential of interpretable regional modeling in hydrology when catchment descriptors are appropriately integrated. Beyond performance improvements, the study provides a structured modeling framework and a CAMELS style dataset for Thailand, contributing to the integration of data-driven forecasting with hydrological realism.