Physics-Informed Neural Networks (PINNs) gains attentions as a promising approach for applying deep neural networks to the numerical solution of nonlinear partial differential equations (PDEs). However, due to the challenging regions within the solutions of 'stiff' PDEs, e.g., shock front of CO₂ immiscible flooding, adaptive methods are essential to ensure the neural network accurately addresses these issues. In this work, we introduce a novel method for adaptively training PINNs, named Self-Adaptive PINNs (SA-PINNs). This approach employs fully trainable adaptation weights that are applied individually to each training point. Consequently, the neural network autonomously identifies challenging regions of the solution space and focuses its learning efforts on these areas. This method is hereby used to simulate a two-phase immiscible flooding in a low-permeability oil reservoir, with considering gas dissolution and the threshold pressure gradient of oil phase in low-permeability oil reservoirs, i.e., modified Buckley-Leverett (B-L) problem. The model is capable of generating a precise physical solution, accurately capturing both shock and rarefaction waves under the specified initial and boundary conditions, though the introduction of complicated physics increases the nonlinearity of the governing PDEs. The self-adaptive mechanism modifies the behavior of the deep neural network by simultaneously minimizing the losses and maximizing the weights. It, thus, can effectively capture the non-linear characteristics of the solution, thereby overcoming the existing limitations of PINNs. In these numerical experiments, the SA-PINNs demonstrated superior performance compared to other state-of-the-art PINN algorithms in terms of L2 error. Moreover, it was also achieved with a reduced number of training epochs. SA-PINNs can effectively model the dynamics of complex physical systems by optimizing network parameters to minimize the residuals of the PDEs.