Physics-Informed Neural Network Method for Parabolic Differential Equations with Sharply Perturbed Initial Conditions