Runoff forecasting using a Takagi–Sugeno neuro-fuzzy model with online learning

SummaryA study using local learning Neuro-Fuzzy System (NFS) was undertaken for a rainfall–runoff modeling application. The local learning model was first tested on three different catchments: an outdoor experimental catchment measuring 25 m2 (Catchment 1), a small urban catchment 5.6 km2 in size (Catchment 2), and a large rural watershed with area of 241.3 km2 (Catchment 3). The results obtained from the local learning model were comparable or better than results obtained from physically-based, i.e. Kinematic Wave Model (KWM), Storm Water Management Model (SWMM), and Hydrologiska Byråns Vattenbalansavdelning (HBV) model. The local learning algorithm also required a shorter training time compared to a global learning NFS model. The local learning model was next tested in real-time mode, where the model was continuously adapted when presented with current information in real time. The real-time implementation of the local learning model gave better results, without the need for retraining, when compared to a batch NFS model, where it was found that the batch model had to be retrained periodically in order to achieve similar results.

► We model rainfall–runoff process by dynamic evolving neural-fuzzy inference system.
► The study is carried out for three different scales of catchment sizes.
► DENFIS was comparable if not superior compared to the physically-based models used.
► DENFIS and ANFIS were almost identical but DENFIS was faster in training time.
► We propose real-time implementation of the local learning which needs no retraining.