Advancement of Data-Driven Short-Term Flood Predictions on an Urbanized Watershed Using Preprocessing Techniques
MetadataShow full item record
Supervised classification can be applied for short-term predictions of hydrological events in cases where the label of the event rather than its magnitude is crucial, as in the case of early flood warning systems. To be effective, these warning systems must be able to forecast floods accurately and to provide estimates early enough. Following the approach of transforming hydrological sensor data into a phase space using time-delay embedding, an attempt was made to improve the performance of the models and to increase the lead-time of reliable predictions. For this, the available set of attributes supplied by stream and rain gauges was extended by derivatives. In addition, imbalanced data techniques were applied at the data preprocessing step. The computational experiments were conducted on various data sets, lead-times, and years with different hydrological characteristics. The results show that especially derivatives of water level data improve model performance, increasingly when added for only one or two hours before the prediction time. In addition to that, the imbalanced data techniques allowed for overall improved prediction of floods at the cost of slight increase of misclassification of low-flow events.