Bad data detection

[petersalemink95]

[jaccoheres]

For SEAMLESS it would be very valuable to filter out unreasonable high or low sensor-values, so that convergance issues of the PGM state-estimator are prevented. It would be even better if also wrong values for e.g. the impedances and other asset properties could be detected that could lead to convergence issues.

[scud-soptim]

Another aspect would be to completely suppress grossly incorrect measurements in the estimation by adjusting the weight. Bad data can already be detected after the third iteration by evaluating the magnitude of ( \frac{r}{\sigma} ), where ( r ) is the residual ( (z - h(x)) ). If the magnitude exceeds a threshold, bad data is detected. If this measurement is assigned a new weight $\sigma_{\text{new}}$, bad data can be completely suppressed. The new weight is given by:

$w_i^{\text{new}} = \frac{w_i}{(1 + \alpha \cdot \frac{r_i^2}{\sigma_i^2})} $