Towards an AI-based Early Warning System for Bridge Scour

The maximum error in scour trough and filling peak forecasts are provided in Table 3 and graphically shown in Figure 22. The maximum error based on the mean of predictions varies between 0.5m to 0.7m for scour troughs and 0.4m to 1.7m for filling peaks. The lower bound (LB) and upper bound (UB) errors show a reasonable degree of variability in the LSTM predictions, varying between 0.2m to 0.9m for scour, and 0m to 1.4m for filling. Impact of Flow Velocity (Discharge) In order to explore whether velocity is a critical feature in presence of stage timeseries, we incorporated the discharge measurements (discharge), obtained from the USGS website, into the LSTM models for bridge 742 as an input feature and compared the performance among three different feature combinations: ssd:[sonar, stage, discharge], sd:[sonar, discharge], and ss:[sonar, stage]. Discharge is computed based on gage-height records (flow velocity) multiplied the river cross-section area. Gage-height records are obtained by systematic observation of a non-recording gage, or with automatic water level sensors relayed by remote gagging stations (Sauer and Turnipseed 2010). Figure 23 provides histograms of the discharge time-series for bridge 742 and its cross-correlation with sonar and stage. Stage and discharge show a large positive correlation as observed both in Figure 23 and Figure 24.