Optimal reservoir operation system based on artificial intelligence and metaheuristics algorithms

• Main Author: Karim, Sherif Mostafa Hassan Ibrahim
• Format: Final Year Project / Dissertation / Thesis
• Published: 2023
• Subjects: T Technology (General); TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utar.edu.my/6240/1/KARIM_SHERIF_MOSTAFA_HASSAN_IBRAHIM.pdf; http://eprints.utar.edu.my/6240/
• Institution: Universiti Tunku Abdul Rahman

Over the immediate past decades, global warming across the world and in Malaysia has caused extreme changes to the climate by disturbing its hydrological cycle, causing drought to some areas and floods to another. As a result of this, it has become more important to meet the rising need for water in order to facilitate increased agricultural, residential, and industrial production. Therefore, the process of developing a prediction for water inflow and optimising the release operation in a reservoir has become an increasingly crucial undertaking for the management of a reservoir. Conventionally, the simulation and forecasting scenarios were calculated using traditional linear models but however, traditional linear models lack the capability to grasp the dynamic and non-linear aspects inherent in hydrological applications. Therefore, if inflow projections were more accurate, there would be a greater need for monitoring of the water quality, and the management of the reservoir would be more effective. Following this, the difficulties caused by flash flooding and the water crisis in Malaysia and the rest of the world may be mitigated with the aid of machine learning. In this study, four different machine learning models were proposed to forecast the reservoir inflow namely, Support Vector regression (SVR), Multi-layer perceptron neural network (MLPNN), Adaptive Neuro Fuzzy Inference System (ANFIS) and Extreme Gradient Boosting (XG-Boost). All the four models were given historical data for training and testing that were collected over 19 years (2000-2019) at Klang Gate Dam which is located in the Gombak District. The data was divided into monthly and daily timeframe while the daily is further time lagged and subdivided into 7 main scenarios starting with scenario-1 to scenario-7 with one, three, and five-days lag for water level and inflow. Four optimisation algorithms were used to simulate reservoir operation over a 12-month period and generate a release curve. For the sake of validation four statistical analysis namely coefficient of determination (R²), Mean Square Error (MSE), Median Absolute Error (Mead) and root mean squared error were adapted to validate and test the machine learning models. On the other hand, risk analysis test was carried out to test the optimisation algorithm output. The risk analysis consisted of reliability, resiliency to recover from failure, vulnerability degree and finally sustainability. Results revealed that: (1) For the monthly period, all four models were capable of predicting effective monthly reservoir inflows by achieving at least an R² of 0.5; the XG�Boost model was rated as the best model, followed by the MLPNN, SVR, and finally ANFIS. (2) For forecasting daily inflow, the XG-Boost still surpasses all other models, but with diminished efficiency. The models were still placed in the same order, with the ANFIS doing very poorly in scenarios 2, 3, and 4. (3) The best scenarios for daily inflows are scenarios 5, 6, and 7, since the models were developed using 1, 3, and 5 days of anticipated inflow, and XG�Boost consistently beats all other models. Moreover, for the optimisation phase all four models were able to produce release curves with a certain degree of validity. Firefly algorithm (FA) was ranked as the best optimisation algorithm after obtaining a vulnerability percentage of 39% as the finest optimisation algorithm model; followed by the particle swarm (PSO) at 59.99%, then the genetic algorithm (GA) with a vulnerability of 68%, and lastly, the nuclear reaction (NRO) with 72%. The present study proposes a novel approach for developing a one-size-fits-all system for forecasting water reservoir inflow and optimizing reservoir operation. The proposed system is based on a combination of machine learning and optimisation techniques, and it has the potential to eliminate the need for tedious and time-consuming manual data collection and interpretation. This could lead to significant improvements in the efficiency and effectiveness of water resources management, and it could also facilitate the development of more informed and timely water resources policies at the national and international levels.