Model predictive control is an emerging embedded control scheme that is increasing relevant to modern power systems. It is known to be capable of improving the stability and regulation performance of the power grid with high penetration of power electronic converters. However, this model-based dynamical control method is known to suffer from model parameters mismatch, which could be caused by device aging, temperature fluctuation, magnetic saturation, etc., subsequently affecting the prediction accuracy and control performance. To mitigate the negative impact of parameter mismatch, the work summarizes the design and assessment of neural networks to enhance the predictive grid current control scheme against the inherent problem of parameter mismatches. Neural network approach is selected over other tools for its versatility and scalability in other higher-order converter/filter topologies and applications. Two neural networks, one to detect the level of mismatch, and another to adjust the parameter in parallel to the predictive control loop, are developed. Different network configurations are assessed, and optimal designs are recommended.