Observation Definitions

The observations provide the primary (and only) measure that informs the optimization algorithm about the model-to-measurement misfit. For each observation FePEST requires:

• its type (e.g., hydraulic head, fluid rate budget),

•  its location (e.g. an observation coordinates, stored nodal selection),

• the time at that it was recorded (in case of a transient model) and

• the value that was observed in reality.

In many cases the observation definitions are created based on FEFLOW Observation Points. Supported observation types are all observations of system state (e.g., hydraulic head, saturation, mass concentration, temperature and age spe­cies) and rate and/or period budget values (fluid, mass and heat) depending on the models FEFLOW Problem Class. Reference values are constant or time-varying values depending on the problem state (steady or transient).

Advanced users may also use IFM plug-ins or the FEFLOW Python Interface to define a customized observation definition. See details in Customization section.

Observation Weights

The definition of the measurement objective function used by GLMA requires certain observation weights.

The weight of an observation controls how strong its residual (the deviation between computed and measured result) contributes to the measurement objective function. A reasonable choice of weights can positively influence the convergence behaviour and result of the .

Different weighting strategies can be applied (alone or in combination), some examples are given in the following.

Weighting by Measurement Noise

A common strategy of adjusting observation weights is applying the inverse of its expected measurement noise as a weight factor. The contribution of less trustworthy observation values to the measurement objective function is reduced, limiting the risk of inaccurate measurements having a negative impact on the optimization and leading to the estimation of parameter values which are thereby in error.

Weighting by Absolute Measurement Value

The absolute values of observations in a PEST optimization can encompass several orders of magnitudes, especially (but not limited to) if observations of different types are involved (e.g., Hydraulic head [m] and Mass concentra­tions [mg/l]).

Observations with small values are therefore under-represented in the meas­urement objective function. Normalizing the values by assigning a weight equal to the inverse of the absolute compensates for this effect and makes sure that the information contained in these values finds appropriate rep­resentation in the optimization.

Equalizing Observation-Group Contributions

Observations of different types are assigned to different observation groups. One may also decide to manually assign observations of the same type to dif­ferent observation groups. Using a spreadsheet or the PEST tool PWTADJ1 observation weights can be adjusted to equalize the total contribution of each observation group to the total objective function at the start of the optimization process. This helps to ensure that the information that is contained in each of these observation groups is used in estimation of model parameters, and not undervalues or overvalued because of too low or too high a contribution to the initial objective function.

De-clustering

Observations can be correlated. Water level measurements at observation wells close to each other are often not independent. It is likely that values and changes at these wells are similar, and the worth of information contributed by one well is diminished because it was already contributed by a different well in its vicinity. The worth of the information provided by each of the wells is therefore lower than a separate measurement at a larger distance. In this case, the weight of correlated observations should be reduced.

Time series are good examples of this principle as well. Daily measurements of the groundwater level might carry the same worth of information as a measurement taken on a monthly basis does. Because the daily measure­ment has more measurement points, it would be over-represented if the weights of each of these measurement points are not compensated for. This makes it advisable to normalize the weights of observations of time series by the sampling rate of the measurement.

 

Further reading: PEST Manual (5th Ed.), Chapter 2.1.2: Observation Weights.