The data base corresponds to 2312 pre-contingency operating states.
154 candidate attributes were defined and their values computed at
after the loss of 600MW of local generation within the
Brittany region. Among the candidate attributes, 68 are HV or EHV
voltage magnitudes and the remaining are active and reactive regional
load level and power flows, reactive generation reserves and
topological indicators. The number of candidate attributes is
representative of real large scale applications of the computer based
learning framework, while the number of learning states remains
moderate.
A 320-bus model is used representing the 225kV and 400kV system in the relevant part of the EDF system. The HV and MV subsystems are represented by an equivalent model at each EHV bus, composed of a two transformer cascade with shunt compensation at the HV and MV terminals. To simulate the behavior of the system a fast time domain simulation is used, which models the effect of OLTCs, overexcitation limiters and secondary voltage control, thereby reproducing only the relevant phenomena for our problem while filtering faster dynamics [27]. A JAD state is classified as non-critical if the subsequent mid-term dynamic behavior is deemed acceptable and leads to an equilibrium state whose load-power margin is larger than a given threshold. Accordingly, in addition to a discrete ``critical vs non-critical'' classification, the JAD states are also characterized by their load-power margin. About 24% of the states are classified as critical.