Scholarlink

<p>Accurate reservoir water balance modelling is essential given the increasing frequency of hydrological extremes and competing water demands. This study evaluates and compares the performance of Physics-Informed Neural Networks (PINNs) and Long Short-Term Memory (LSTM) models in simulating daily storage changes at the Samanalawewa Reservoir in Sri Lanka, based on 18 years of hydrological data. The LSTM was designed as a stacked two-layer recurrent network optimised for short-term temporal dependencies, while the PINN incorporated the governing water balance equation into a residual-based deep neural architecture. Both models were trained on peak-removed regimes and tested on peak-inclusive regimes. Results show that the PINN achieved high predictive accuracy under both scenarios, with a Nash-Sutcliffe Efficiency (NSE) value of 0.87 and Root Mean Square Error (RMSE) of 0.34&#xa0;Million Cubic Meters (MCM) for normal flows and maintained robust performance under extremes with NSE of 0.80 and RMSE of 0.63 MCM. The LSTM performed comparably under normal flows (NSE: 0.89, RMSE: 0.32 MCM) but deteriorated substantially under extremes (NSE: 0.67, RMSE: 0.82 MCM), failing to capture the timing and magnitude of sharp storage transitions. Storage-change residual analysis showed that the PINN residuals were narrowly centred around zero (mean: −0.03 MCM), while LSTM exhibited positive bias and heavy-tailed errors (mean: 0.35 MCM, max: 12.36 MCM). Sensitivity analysis further suggested that PINNs captured nonlinear hydrological feedback and maintained mass balance, whereas the LSTM responses remained monotonic and less interpretable. These findings demonstrate that the PINN improved the robustness of reservoirs under hydrological extremes, supporting operational decision-making.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Evaluation of Physics-Informed and Data-Driven Neural Networks for Reservoir Water Balance Simulation During Hydrological Extremes

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • S. Tharuka,
  • L. Gunawardhana,
  • B.C. Dissanayake,
  • L. Rajapakse

摘要

Accurate reservoir water balance modelling is essential given the increasing frequency of hydrological extremes and competing water demands. This study evaluates and compares the performance of Physics-Informed Neural Networks (PINNs) and Long Short-Term Memory (LSTM) models in simulating daily storage changes at the Samanalawewa Reservoir in Sri Lanka, based on 18 years of hydrological data. The LSTM was designed as a stacked two-layer recurrent network optimised for short-term temporal dependencies, while the PINN incorporated the governing water balance equation into a residual-based deep neural architecture. Both models were trained on peak-removed regimes and tested on peak-inclusive regimes. Results show that the PINN achieved high predictive accuracy under both scenarios, with a Nash-Sutcliffe Efficiency (NSE) value of 0.87 and Root Mean Square Error (RMSE) of 0.34 Million Cubic Meters (MCM) for normal flows and maintained robust performance under extremes with NSE of 0.80 and RMSE of 0.63 MCM. The LSTM performed comparably under normal flows (NSE: 0.89, RMSE: 0.32 MCM) but deteriorated substantially under extremes (NSE: 0.67, RMSE: 0.82 MCM), failing to capture the timing and magnitude of sharp storage transitions. Storage-change residual analysis showed that the PINN residuals were narrowly centred around zero (mean: −0.03 MCM), while LSTM exhibited positive bias and heavy-tailed errors (mean: 0.35 MCM, max: 12.36 MCM). Sensitivity analysis further suggested that PINNs captured nonlinear hydrological feedback and maintained mass balance, whereas the LSTM responses remained monotonic and less interpretable. These findings demonstrate that the PINN improved the robustness of reservoirs under hydrological extremes, supporting operational decision-making.

Graphical Abstract