Scopus Indexed Publications

Paper Details


Title
Physics-informed temporal attention network for river temperature and dissolved oxygen forecasting near Rooppur Nuclear Power Plant

Author
Nafis Wadud Ansari, Huzifa Khatun Hera, Md. Selim Reza, Shifat Ullah,

Email

Abstract

This study presents a Physics-Informed Temporal Attention Network (PITAN) for forecasting river temperature (T) and dissolved oxygen (DO) in the Padma River surrounding the Rooppur Nuclear Power Plant (RNPP). PITAN combines a hybrid CNN–BiLSTM backbone, multi-head scaled dot-product temporal self-attention, and physics-informed regularisation to capture nonlinear interactions among hydrological, atmospheric, and physicochemical variables while discouraging physically infeasible forecasts. Multi-source environmental records from the Bangladesh Water Development Board were quality-controlled and temporally harmonised, yielding 82 monthly observations from July 2016 to September 2023. One-step-ahead forecasts were generated using a seven-step historical window and evaluated using rolling-origin temporal validation with multiple random initialisations. PITAN achieved pooled RMSEs of 2.33°C for temperature and 1.78 mg L⁻¹ for DO, compared with 2.88°C and 2.02 mg L⁻¹, respectively, for persistence forecasts. Component-wise ablation showed that physics-informed output anchoring was the principal contributor to temperature forecasting skill, with its removal reducing R² from 0.484 to −0.036 (p < 0.001), whereas the contributions of the convolutional, recurrent, attention, and penalty components were not statistically distinguishable at the available sample size. Predictive uncertainty was characterised using Monte Carlo dropout and benchmarked against split-conformal intervals, while SHAP analysis highlighted the annual cycle, antecedent dissolved oxygen, and river discharge as the predictors most strongly associated with model output. Overall, the findings indicate that physically anchored deep learning can improve forecasting of coupled thermal and oxygen dynamics in a data-scarce monsoonal river while providing uncertainty estimates, interpretable predictors, and explicit evidence regarding which architectural components are supported by the available data.


Keywords

Journal or Conference Name
Results in Engineering

Publication Year
2026

Indexing
scopus