Modeling of the flow time series for a short-term hydrological forecast

Abstract

Aim of the study: Within this article an example of an effective approach to real-time, short term forecast of flood rates within Vistula river differential catchment was presented. This forecast is based on flow rates time series measured at the water gauge input and output cross sections of the river system with a daily delay without taking into account any precipitation data. Material and methods: In order to assess the quality of the forecast, four types of time series models were developed for the Smolice outlet gage station. The first type of model is the conventional linear autoregressive relationship (AR), the second one - three layer neural network feedforward (SSN), the third one – two layer recursive neural network and the fourth one- three layer special kind of recurrent neural network (RNN). All models were trained and tested based on historical flood events data. Results and conclusions: Among the all tested model types, the most accurate prediction of the instantaneous value of the flow rate in the outlet cross section of the Vistula catchment was obtained using the RNN model. This type of model also had the greatest ability to generalize results confirmed by three independent tests.

Keywords

• river system
• recurrent neural network
• flow rate time series
• short-term forecast
• peak flow rate