Enhanced Imitation Learning of Model Predictive Control Through Importance Weighting

The approximation of model predictive control (MPC) algorithms with neural network models has received significant attention in the power electronics research community, as a valuable tool to enable the real-time implementation of computationally expensive methods. Training machine learning models to imitate MPC algorithms typically follows a traditional supervised learning flow, with a static training data set generated through either uniform sampling or simulation. Training with uniformly distributed data allows models to maintain consistent performance throughout the entire operating space, but very small models may fail to achieve satisfactory results in the regions corresponding to expected operation. Conversely, simulation data can be used in training to obtain models that can accurately track certain trajectories but fail to achieve adequate performance in underrepresented regions. The present article proposes a method to combine the advantages of both methods, by training models on sampled data with importance weights obtained through the kernel density estimation of experimental data. This allows the training process to target specific regions of the operating space, and results in measurable performance improvements. The method is demonstrated through the imitation of a moving discretized control set MPC algorithm for a dual active bridge converter, and is validated through experimental results.