Runoff Identification and Delimitation in Discharge Series Using Forward and Backward Moving Averages

Evaluation of single-event hydrologic forecast models requires identification of independent runoff events. Event identification commonly uses peak over threshold (POT), annual maximum flood (AMF) methods, and recursive baseflow separation algorithms, all of which have advantages and disadvantages. For example, in the AMF method only one event is selected per year confirming independent events. However, there are other significant events during the year that are completely missed by AMF. On the other hand the POT method detects multiple storms within the same year but the number of events is highly dependent upon the selected threshold. Recursive baseflow separation algorithms require estimation of uncertain catchment parameters.

We propose a new methodology to identify independent events within time intervals. The proposed method involves calculating N-point forward and backward moving averages and taking the difference between them (delta). We identified key parameters including the size of the moving average windows, the minimum delta value required to identify an independent event (peak-threshold), the regression length used to a local slope, as well as the end-of-event threshold. Subtracting the forward moving average from the backward moving average provided a key indicator that proved useful for bounding the beginning and end of unique hydrographs.

Sensitivity analysis results suggest that the parameters determined as optimal by comparing against hydrograph extents vary depending on geographical location. Evaluation of the method used data from the USGS's Hydro-Climatic Data Network (HCDN) and employed a cluster analysis to identify stations with similar parameters in different regions of the U.S. The identified events in the HCDN dataset were compared with those produced by the retrospective runs of the National Water Model (NWM) versions 1.2 and 2.0. Accuracy of the events produced by the NWM were evaluated using error metrics to confirm reliability, as well as magnitude and duration of the storm events. Additional results indicate that the parameters needed to run the developed event identification method depend on the annual runoff ratio. The method is quite sensitive to changes in the peak and end-of-event thresholds and less impacted by the size of the moving average window.