Generalize the effect of the discount factor on protection security

Determine whether increasing the discount factor b3 in offline conservative Q-learning for line-selective protection improves security, particularly by reducing false trips in non-fault distribution-grid episodes.

Background

The study compares a default offline conservative Q-learning policy with discount factor b3 = 0.95 against one exploratory post-hoc policy using b3 = 0.99. The higher-discount policy reduces non-fault nuisance trips from eight to six episodes and slightly increases the number of correct first trips on fault episodes, but the evaluation uses only 11 non-fault episodes and one training seed.

Because the higher discount factor is tested only once and outside the prespecified experimental matrix, the observed security improvement cannot establish whether larger discount factors generally improve terminal protection behavior. Establishing this effect requires broader experiments across discount factors, datasets, seeds, and non-fault operating conditions.