Namangan Institute of Engineering and Technology
nammti.uz
10.25.2023
Pg.292
result, it may be used to refine physical models by identifying and quantifying unknown effects in
historical operational data. Our machine learning model, however, simply recognises connections
in the data, which may or may not be indicative of causal linkages. Consequently, in order to debate
and confirm our data-driven results, we require domain knowledge, such as that which comes from
physical models (cf. ref. [12]).
We illustrate the broad applicability of our machine learning approach with our case studies,
which can also be applied to improve power system control and operation. Risks to frequency
stability were highlighted by the significance of prediction errors in the Nordic grid and quick
generation ramps in Continental Europe. These risks can be reduced with more optimisation or new
system rules. Model predictive control or preventative control measures can be implemented using
day-ahead estimates of frequency variations. Lastly, SHAP provides additional tools like interaction
analysis and monitoring plots that may be used in the future to address other elements of the
energy system in addition to frequency stability.
