Enhancing reservoir simulation models with genetic algorithm optimized neural networks across diverse climatic zones / Saad Mawlood Saab

Saad Mawlood , Saab (2025) Enhancing reservoir simulation models with genetic algorithm optimized neural networks across diverse climatic zones / Saad Mawlood Saab. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (120Kb)
	PDF (Thesis PhD) Download (3315Kb)

Abstract

Dams and reservoir systems help improve livelihoods, agriculture productivity, and farmers' drought resilience by regulating and increasing water supply reliability. In fact, the reservoir simulation depends on several hydrological parameters. Since hydrologic parameters exhibit a high degree of stochasticity, developing an accurate forecasting model that reproduces such a complex pattern is becoming increasingly challenging. A well-designed and reliable forecasting model is key to the successful reservoir simulation so as to maximize the use of water resources. Since the hydrological parameters are difficult to handle mathematically, existing prediction models are burdened with several drawbacks. The aims of the study are to develop robust models to predict two different parameters of hydrology in the dam reservoir and examine their performance under different climate conditions. Also, the study introduces a new procedure for reservoir simulation. The current research presents three different AI approaches: i) Multi-Layer Perceptron Neural Network (MLP-NN), ii) Radial Basis Function Neural Network (RBF-NN), and iii) Deep Learning Neural Network (DLNN). The proposed models were utilized to predict two key hydrological parameters related to reservoir simulation: inflow and evaporation. The research improved the predictive models by integrating them with the Genetic Algorithm (GA). The optimizer algorithm (i.e., GA) determines the optimal input variables and internal parameters in the prediction models. To illustrate the models' efficacy, predictive models were applied to predict reservoir inflow and evaporation in two different case studies representing different environmental conditions, semi-arid and tropical case studies. The first case study is Dukan Dam, located in Iraq (semi-arid region), and the second is Timah Tasoh Dam (TTD), located in Malaysia (tropical region). Comparative analysis was performed between predictive models based on several statistical indicators. The prediction outcomes demonstrated that the GA-DLNN performs better than other proposed models. The GA-DLNN achieved well results in forecasting inflow values where it attained low (RMSE (23.49 m3/sec at Dukan, 2.92 MCM month-1 at TTD) MAE (15.55 m3/sec at Dukan, 2.06 MCM month-1 at TTD) and high correlation coefficient (R2 = 0.967 at Dukan, R2 = 0.969 at TTD). Also, the results indicated that the GA-DLNN achieved high level accuracy in prediction reservoir evaporation values, where it attained a low (RMSE (0.73 mm day-1 at Dukan, 6.77 mm month-1 at TTD) MAE (0.30 mm day-1 at Dukan, 3.87 mm month-1 at TTD) and high correlation coefficient (R2 = 0.976 at Dukan, R2 = 0937 at TTD) . On the other hand, the current study introduced a new procedure for simulating reservoirs under realistic conditions. This procedure was performed by including the prediction results obtained by the best and worst models in the balance equation. Reservoir condition assessment under new and conventional procedures was performed by calculating the percentage error during the simulation period. It was observed that the reservoir condition changed significantly with the inclusion of the predicted flow and evaporation data within simulation session. This research found that GA-DLNN method is better than alternative models put forth in predicting reservoir inflow and evaporation data. The predicted data should be adopted while performing the reservoir simulation.

Actions (For repository staff only : Login required)