Hydraulic modeling and model predictive control using Neural Network

Student thesis: Master thesis (including HD thesis)

  • Michael Mølskov Rasmussen
  • Kristoffer Stenkær Schneidelbach
4. term, Water and Environment, Master (Master Programme)
The purpose of this project is to investigate how neural networks can be used for modelling of different hydrological problems in Vejle in relation to the utilization of real-time measurements. By this means, how well the neural network is sufficient in solving and finding the necessary interpretation and interaction between the included hydrological processes.

This study examines the scenario of modelling the run-off to the stream Grejs Å by the use of neural network as the primary modelling tool, based on the input of precipitation from a nearby-located rain gauge. Furthermore, examination of a potential lead-time for the forecast of the water flow in Grejs Å due to the delay from precipitation to actual run-off. In addition, examination of modelling and forecast of the water level in Vejle harbor, based on the water level from other harbors correlated to the same conditions as Vejle. This mainly presupposes that the other harbors have similar locations in which the water level will be obtained, where forecast of the water level in Vejle harbor requires that a certain lead-time exists between the different harbors. Lastly, an examination of how well a neural network is suitable of controlling the weir that divide the water from Grejs Å into Mølleåen and Omløbsåen in order to reduce potential flooding caused in the Omløbsåen. The input to the neural network controlling the weir level is the previous modelled run-off in Grejs Å and water level in Vejle harbor.

This project considers the type and processing of the input to the neural networks, in which more favourable results from the networks will be achieved. This is in order to examine what input that is preferred in relation to the focus of the specific case and type of system process.

To examine this, different types of models and methods are to be included. Due to the uncertainty of the approach and processing, secondary models is incorporated in order to explore different effect to the affected systems. This includes the implementation of a linear reservoir model that has shown how different included processes affects the outcome of the model setup. This results in the design of a drought-index made of precipitation and evaporation. The achieved understanding of the influence of different processes is used in the creation of a better neural network for the modelling of run-off to Grejs Å. The network for forecast of water level in Vejle harbor is enhanced by the try of a self-designed wind-factor included as an input to the neural network. In addition, due to the lack of the position for the level of the weir at Grejs Å, a conceptual model (MIKE URBAN-model) has been created to produce suitable data as input to a neural network. This makes possible of examination and handling of the control and forecast for the weir level in which the ability of the neural network can be studied.

The results show that neural networks is applicable as the primary modelling tool of run-off to the stream of Grejs Å, where the rain signal and self-designed drought-index is transformed into flow to the stream. The neural network shows a potential as a useful tool when combining similar signal types to estimate each other. This enables forecast of the water level in Vejle harbor by the use of secondary water level measurements from other harbors in combination with the self-designed wind factor. Finally, the neural networks shows that controlling of the weir at Grejs Å is possible as long as the requested input is available, including measurements or estimates of water level and flow and the corresponding position for the weir level.

The production of these results presuppose that a thorough understanding and interpretation of the natural system exists. Additionally that the structure, configuration and processing of the input to the neural network is made appropriate to the specific case and furthermore that a certain amount of data is available for calibration and validation of the neural network.
Publication date1 Jul 2016
Number of pages192
ID: 234928392