An application of a machine learning algorithm to determine and describe error patterns within wave model output

This study uses a machine learning algorithm, the bagged regression tree, to detect error patterns within 24-h forecasts of significant wave height time series. The input to the machine learning algorithm were bulk parameter outputs of the numerical wave model (WaveWatch III) and wind information from the Global Forecast System at buoy locations along the California-Oregon border in the United States. The output of the algorithm were predictions of hourly deviations between numerical model output and buoy observations of significant wave height. When these deviations were applied as corrections to the forecasts, error metrics root-mean-squared-error, bias, percent error, and scatter index were reduced in several different experiments, confirming that the error pattern was successfully detected by the machine learning algorithm. Furthermore, the detected error pattern was consistent between buoys at different locations, as presented in a geo-spatial application of the machine learning algorithm. As a descriptive tool, the algorithm delineated regions of similar error within the context of model phase space (significant wave height and mean wave period (Tm01)). Specifically, the algorithm detected significant wave height overestimations for significant wave heights greater than 3.4 m, wave period greater than 9.1 s, and waves coming from the W-NW quadrant.Also, for significant wave heights greater than the 95th percentile value (5.4 m), the algorithm detected differences in model phase space associated with mean error patterns.