Research Publications

Artificial intelligence (AI) and deep learning (DL) models are effective tools for simulating reservoir operations by representing nonlinear relationships among reservoir states, hydrological and climatic conditions, and human-controlled release decisions. However, the value of systematic hyperparameter optimization remains insufficiently evaluated across heterogeneous reservoir systems. This study applies Bayesian optimization using a Tree-structured Parzen estimator (BO-TPE) to reservoir-specific long short-term memory (LSTM) tuning for the simulation of daily releases from 441 reservoirs across the contiguous United States (CONUS). The BO-TPE-tuned LSTM (LSTM-TPE) was evaluated against two fixed-hyperparameter LSTM configurations and conventional machine-learning (ML) models, including artificial neural network (ANN), support vector regression (SVR), and random forest (RF), under both grid-search and fixed-budget TPE tuning. All models used six input variables: reservoir storage, inflow, precipitation, air temperature, the Palmer Drought Severity Index (PDSI), and snow water equivalent (SWE), representing key reservoir-state, hydrological, meteorological, and climatic controls on human-managed releases. The results showed that LSTM-TPE generally outperformed the conventional ML models and fixed-hyperparameter LSTM configurations, achieving median NSE and KGE values of 0.78 and 0.81, respectively. Using satisfactory-performance thresholds of NSE > 0.6 and KGE > 0.6, LSTM-TPE achieved satisfactory performance for 345 reservoirs (78.2%) and 356 reservoirs (80.7%), respectively. Bootstrap analysis was used to derive heuristic 95% confidence intervals (CIs) summarizing performance variability across sampled LSTM hyperparameter configurations. Narrower CIs were generally associated with better-performing reservoirs, whereas wider CIs were more common in southern and Intermountain regions and may reflect greater operational complexity and hydroclimatic variability.