POSSIBILITIES FOR DAY-STEP FLOOD FORECASTING IN SMALLER CATCHMENTS USING MACHINE LEARNING METHODS

Today, it is possible to work with a wide range of freely available data, often on a daily basis. These data can be used to create an early warning system for estimating the approximate hazard. It can also be used to develop models of long-term catchment behaviour. Most applications are generally carried out on catchments larger than 200 km2. For this reason, areas between 20 and 350 km2 were selected to test the hypothesis using models based on machine learning methods. How good results can be achieved when using daily data to predict increased flows caused by previous precipitation (summer) or a combination of snowmelt and precipitation (winter). The one day step was chosen for availability data (free data) and for testing area size limit for this step. The results showed that floods caused by a combination of rain and snowmelt were significantly better than those caused by rain alone. Two methods were compared. The neural networks ANN and fuzzy model. For both methods were founded the best architecture in training period. The results of the experiment showed that the limit of applicability of the data is above (around) 130 km2 in the case of pure rainfall. In the case of floods caused by a combination of rain and snow, the daily step can be used even for catchments of about 20 km2 with a one-day time shift.