Modeling Reservoir Surface Water Temperature With Neural Networks And Anfis: Insights From Sensitivity Analysis
Auteurs : Belouz Khaled . Zereg Salah . Matallah Salim .
Résumé
Accurate prediction of reservoir surface water temperature (Tw) is vital for effective water resource planning, environmental management, and maintaining water quality, as Tw impacts numerous physical, chemical, and biological processes. However, in many developing countries, Tw data and related meteorological variables are often scarce. This study assesses the predictive performance of generalized artificial neural networks (ANN) and adaptive neuro-fuzzy inference systems (ANFIS) in predicting Tw using a minimal set of climatic inputs and a site-specific cross-validation approach. The dataset was split into two parts: a pooled dataset from six reservoirs in different climatic regions for training, and data from the Beni Bahdel reservoir for testing. Eight input combinations, including air temperatures (mean, maximum, and minimum), relative humidity, and day of the year (DOY), were explored. Both ANN and ANFIS outperformed traditional multi-linear regression (MLR). Among the tested models, those using air temperatures and DOY as inputs—ANN-M2 and ANFIS-M7—showed the best performance, with ANN-M2 achieving an R value of 0.97, RMSE of 1.52°C, and MAPE of 7.43%, and ANFIS-M7 achieving an R value of 0.97, RMSE of 1.54°C, and MAPE of 7.68% during testing. The results highlight that incorporating DOY improves prediction accuracy, and sensitivity analysis identified DOY and mean air temperature as the most important predictors. Overall, ANN and ANFIS offer reliable, user-friendly methods for accurately predicting reservoir Tw, supporting efforts to manage water quality and ecological health in reservoir ecosystems.
Mots clés
Surface water temperature ; Artificial neural networks ; ANFIS ; Multi linear regression ; Reservoirs
