Click Pictures
New Automated Calibration of Watershed Model
Inflow Vista has several forecast engines available including a rainfall-runoff model based on the Sacramento Soil Moisture model and the SNO-17, which are also used in the National Weather Service Model. The model has over 30 parameters per watershed element. These parameters cannot be determined directly from physical catchment characteristics and they have to be calibrated and/or selected for each deployment of the model. The traditional manual calibration approach requires considerable training and experience and can be a very time consuming and laborious. Furthermore, because of the subjectivity involved in the ‘trial-and-error’ approach, the selected parameters may not necessarily be the best set. Therefore, the decision was made to embark on the development of a parameter search facility for Inflow Vista.
An Automatic Parameter Facility has been added and it is essentially a wrapper around existing data classes and simulation engine. The parameter search techniques employed is referred to as the Shuffled Complex Evolution Metropolis Algorithm (SCEMA) and it is a population-evolution based optimization algorithm. Conventional gradient search techniques do not work well for hydrologic forecast model for several reasons. The solution space is usually irregular (non-convex) and includes multiple local optima. Hydrologic models include a large number of parameters, many of which are interdependent. The data available is sparse and estimates of mean areal precipitation and temperature may be very inaccurate at times. This is why algorithms such as SCEMA have been shown to work well for this type of problems. SCEMA begins with a random sample sets distributed throughout the feasible parameter space, and then continuously evolves the population toward better solutions within the search space.
A new Parameter Search Screen has been added to Inflow Vista for controlling the parameter search, it has six tabs; Set-up, Snowmelt, Loss, Evapotranspiration and Unit Hydrograph (ET UH), Initial state and Results, as shown in Figure 1. The user has control over which groups of parameters to include in the search, as well as the initial values and the maximum and minimum range for each parameter. In addition, several options for objective function are available, including
• Root mean square error (RMSE)
• Weekly volume RMSE
• Flow above threshold RMSE
• Flow below threshold RMSE
• Goodness of fit statistic (1 – Rm2)
As part of the development, a number of code ‘efficiency’ changes were made to the simulation engine to facilitate fast execution and these have reduced the run time for a typical simulation to less than a second; allowing the search engine to evaluate thousands of parameter sets in an hour.
This new facility has been of great value in the calibration of the model for several recent applications and not only reduced the time and cost of setting up the model, but also gives some assurance that the model has been calibrated as well as it can be. Results from a parameters search for one of the watersheds is shown in Figure 2.