Task 1 Developing the Datasets used to Train the Machine Learning Model

The machine learning model will be used to predict the zonal adjustment factors using observed meteorologic and hydrologic conditions in the watershed. In order to train the machine learning model, a dataset that included HEC-HMS model adjustment factors for a number of events needed to be developed as well as the meteorologic and hydrologic conditions preceding the event. 

Flood Event Identification 

Flood events were identified as those with more than 3 inches of precipitation over a 5-day period for the entire Russian River watershed. This criteria resulted in 48 events. Some of the events had suspect precipitation or flow information and were removed from consideration. The following table contains events for training and testing. Most major precipitation events occurred between November and March. The watershed state in November and December are typically drier than in winter and spring months. An effort was made to include a distribution of event dates and precipitation magnitudes in the test period that were similar to the train period. 

HEC-HMS Model Adjustment Factors and Results for the Training Events

A separate HEC-HMS forecast run was created for each training event and the parameters reflecting initial states were modified using zonal adjustments. Again, the zonal factor was multiplied by the base values within each subbasin. The following tables contain the adjustment factors for the training events. 

The following figures show the simulated and observed hydrographs at all four locations for the Feb 2000 training event (occurred on a wet watershed), and for the November 1997 training event (occurred on a dry watershed). Both events had a similar amount of precipitation but different watershed states. Overall, performance metrics, Nash Sutcliffe Efficiency and Percent Bias in runoff volume demonstrate good model performance for the 32 training events.

Predictor Variables for the Training Events

The precipitation, temperature, and flow preceding these events are contained in the following table. These values were used as predictors for initial state adjustment factors in the training events. The values in the table were gathered from the day preceding the start date of the simulations. For example, the 3-day watershed total precipitation is the cumulate 3 day precipitation before the start date of the event. Both precipitation and temperature values represent the total precipitation and average temperature for the entire Russian River watershed.