Data-driven stable neural imitation of economic MPC via an optimization-informed loss function

Economic model predictive control (EMPC) enables efficient system operation based on economic criteria. However, its real-time implementation faces significant computational challenges. Neural networks, particularly recurrent neural networks with long short-term memory (RNN-LSTM), have emerged as a promising alternative for emulating EMPC behavior. This article proposes a neural controller (L J -EMPC) based on an RNN- LSTM that emulates the behavior of an EMPC for energy management in a university microgrid. The methodology relies on a custom loss function that incorporates the EMPC problem’s structure and guarantees the network’s incremental input-to-state stability ( δ ISS). It also uses a closed loop that feeds back the actual plant state and a projection layer to ensure the feasibility of actions. Ablation studies confirm that the proposed loss consistently reduces prediction errors compared to a standard loss. Experiments show that the L J -EMPC reproduces the optimal policy with average errors below 1%. Furthermore, its cumulative economic cost closely matches that of the EMPC, demonstrating that the learned policy preserves the original economic performance. The controller effectively dissipates any disturbance in its internal states within the first few control steps, as formally guaranteed by the δ ISS property. Inference runs in approximately 100 ms with less than 5 MB of memory, enabling its implementation on embedded devices.