We consider the security-constrained optimal power flow (SCOPF) problem in a linearized form where thermal line limits are enforced as soft constraints to reflect operational flexibility. We propose a constraint-reduction method for the penalized SCOPF based on a nested hierarchy of non-redundant constraint subsets. Each hierarchical level defines a reduced SCOPF that is equivalent to the full problem while involving fewer constraints. We develop an extraction algorithm to construct this hierarchy from the unpenalized problem and introduce a cumulative encoding coupled with a supervised learning approach to predict the minimal hierarchy level corresponding to a load profile for the penalized SCOPF. Exactness is then guaranteed through an optional correction step. The method is illustrated in numerical simulations and hierarchical constraint subsets are presented. Using LightGBM predictors on the IEEE 39-bus system, we achieve a 94.28% hierarchy level prediction accuracy, resulting in an average computational time reduction of 19.72%, including the correction step.