Artificial neural networks applied to streamflow simulation in the Itapemirim River basin
Name: REGIANE SOUZA VILANOVA
Publication date: 23/02/2017
Advisor:
Name |
Role |
|---|---|
| ROBERTO AVELINO CECÍLIO | Co-advisor * |
| SIDNEY SARA ZANETTI | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| ROBERTO AVELINO CECÍLIO | Co advisor * |
| SIDNEY SARA ZANETTI | Advisor * |
Summary: For some time man has sought the adequate knowledge of the hydrological processes to obtain the best use of them. Flow simulations are widely used and suggested for the sustainable management of water resources. Artificial neural networks (ANNs) are empirical models widely used to model the rain-flow process. The present study aims to apply and test the feasibility of using ANNs as an option to simulate the flow in the Itapemirim River basin (BHRI), ES. This study evaluated the capacity of the neural network to model the rainfall-flow process on a daily basis using 34 years of rainfall and fluviometric data in 12 sub-basins. Three types of flows were simulated: total daily flow (q), daily outflow (qSup) and daily outflow (qSub). In the network training process, several combinations of input data, including precipitation and potential evapotranspiration data, were tested in three sub-basins: Paineiras (larger area); Rive (intermediate area) and Fortaleza (smaller area). The networks trained in these sub-basins were also tested in the other sub-basins. The results show that the ANNs have higher efficiency in the basins WHERE they were trained. (Pt, Pt-1, Pt-2, Pt-3, Pt-4, Pt-5, P30) best simulated the total daily flow rate in all sub- Basins in which it was trained, with USE of 0.861, 0.837 and 0.711 to Paineiras, Rive and Fortaleza Plant, respectively. In order to train the network to Paineiras and then extrapolate to the sub-basins of smaller areas (Lajinha, Iuna, Ibitirama and Usina Fortaleza), the results proved to be unsatisfactory. Due to these unsatisfactory results, tests were performed for sub- With the purpose of verifying that the extrapolation of a smaller sub-basin to a smaller one would present better results. The networks trained for Rive (intermediate area) presented better results when tested in the other sub-basins, indicating the probable influence of basin scale in this type of behavior. In relation to the separation of the surface and underground flow, the simulation of the surface run presented better results. Comparing the values obtained with the entrance of the total flow in the network and the separation of the same in surface and underground flow, the values were similar for Paineiras, presenting NSE of 0.861 and 0.902, respectively, indicating that
there is no significant improvement when simulating the Flow rates separately. From the tests performed, it can be concluded that it is possible to estimate the daily flow in the BHRI, in a satisfactory way, using RNA.
Keywords: artificial neural network; flow simulation; rain-flow process; modeling
