Sensitivity Analysis

Although the used models are simple there are about 30 parameters that can be varied. It is therefore important to know which parameters have a high influence on the results. With the implemented sensitivity analysis it is relatively easy to determine which parameters are important for a specific task. But it should be emphasised that the results of the sensitivity analysis are only valid for the specific project data, that means sensitivity analysis should be repeated for each scenario because it depends on the chosen mean parameter values. REBEKA2 performs a local sensitivity analysis: each parameter is varied within its range set by the user while all other parameters are kept constant at their mean values. The number of calculation points for each parameter can be set by the user. It ranges from 3 to 11. Default number is 5. If e.g. the lower limit of a parameter is 9.0 and the upper limit is 11.0, the following 5 calculation points are chosen: 9.0, 9.5, 10.0, 10.5, and 11.0. If p parameters are varied and n number of calculation points are chosen p times n simulation runs are executed for the sensitivity analysis.

The program creates a matrix of scatter charts. Each row represents a result and each column a parameter. The  following screenshot shows a part of this chart matrix, for two results (total bedload events and bedload events caused by drainage discharges) and three parameters (Manning-Strickler coefficient , average grain size and 90% grain size of the river bed). It can be seen that the average grain size is the most important parameter for the number of beload events.


To make it easier to navigate through the sensitvity analysis matrix a navigation window shows a small copy of the matrix.



Each pixel in the grid represents a chart with a varied parameter on the x-axis and a result on the y-axis. The colors indicate how sensitive a result reacts on the parameter change. Red are very important parameters, green less important and blue of minor importants. White indicates parameter with nearly no influence on the results. The navigation can be changed to another view where a chart can be selected by its x- and y-axis attributes.

Random Sampling

The choice of parameter values randomly taken from the parameter range is called random sampling. REBEKA2 allows choosing different sampling techniques: Random, Latin hypercube and quasi-random sampling. Latin hypercube and quasi-random distributions are generated by an external program RANDSAMP from the statistical library UNCSIM (Reichert, 2001). They create random values which - compared to random sampling - optimally fill the multi-dimensional para-meter space.

REFERENCES

Reichert P. (2001). UNCSIM - A program package for uncertainty analysis and Bayesian inference: User Manual. Dübendorf, Switzerland, EAWAG: 36.