This report is the product of the work done by two graduates in the 3rd and 4th semesters of the master in Water and Environmental Engineering at Aalborg University under the Department of Civil Engineering. The report surrounds the danish city of Vejle, where flooding has been a major issue for many years. Because of the flooding issues, the municipality of Vejle has built a sluice and a distribution control structure, which can control the water level in the stream of \textit{Omløbsåen}, which flows through the city centre.

The main objective of this Master's Thesis is to develop data-driven forecast-methods to predict the water levels in Omløbsåen, and in the stream of \textit{"Grejs Å"}, which covers the runoff from most of the catchment area. The forecast-methods are developed at different lead times, where the use of Neural Networks has proven to be very effective for establishing complex relations - for example between measurements of rainfall and water levels, which historically have been described by deterministic modelling of the freshwater system through the use of governing equations.

Predictions of water levels at different lead-times are valuable for a number of reasons. Short-term water level predictions in Omløbsåen have an influence on decision making and real-time control of the sluice and the distribution structure, while long-term water level predictions can be used to warn and prepare the citizens of Vejle against increased runoff from the catchment of Grejs Å.

Analyses show that the best data-driven models for short-term predictions are based on the actual water level and the tendency in the measurements. The municipality desires to keep the sluice open as much as possible, because closure of the sluice will prevent fish from travelling upstream. The short-term predictions can act as a tool to not only determine the closure of the sluice in critical situations, but also to indicate whether it is safe to open the sluice again.

The best models for long-term predictions are also based on the actual water level, but they include the rainfall, as this is an indication to whether the water level is rising or falling. It is concluded that models for long-term predictions, which includes rainfall, performs best, when the catchment consists of mostly a rural environment, where the runoff is linked to natural hydrological processes.