The increasing frequency and intensity of weather extremes due to climate change highlight the need for reliable forecasting of river discharges in Slovenia, particularly because of the risk of flash floods. The aim of this thesis was to investigate the applicability of machine learning methods for short-term, 18-hour discharge forecasts at hydrological stations across Slovenia. The analysis included discharge and precipitation measurements from the Slovenian Environment Agency (ARSO) and weather forecasts from the ALADIN model. A deep learning model, TOK, was developed, which, in addition to discharge history and local measurements, also incorporates the spatial characteristics of precipitation over a wider catchment area. Results from 119 hydrological stations on 77 rivers showed that TOK outperforms linear regression and ARSO’s existing Hydrological Forecasting System (HPS), both in standard hydrological metrics and in early flash flood warning. Using the SHAP values, the main factors influencing the predictions of the TOK model were explained. The work provides an improved approach to early flood warning, and contributes to the development of advanced hydrological forecasting systems in Slovenia.
|
