Behavior cloning (BC) for expert controllers in multiscale dynamical systems is difficult because slow states, fast (stiff) states, and predictable exogenous inputs evolve on different timescales, creating highly nonlinear couplings that a shared latent space cannot capture well, harming closed‑loop generalization. We introduce Structured Representation Learning for Behavior Cloning (SRLBC), an architecture that exploits physical knowledge of multiscale systems by mapping slow variables, fast variables, and inputs into separate embeddings.
On a pressurized water reactor (PWR) load‑following benchmark—high‑dimensional, stiff, with strict safety constraints—we run 500 long‑horizon roll‑outs across 20 random seeds. Compared with shared‑embedding baselines, the separated embeddings consistently achieve lower cumulative cost, higher imitation accuracy, and better constraint satisfaction. Stand‑alone policies reduce computational cost by about 15× relative to NMPC while approaching its closed‑loop performance.
Because stand‑alone policies still leave a few percent of trajectories infeasible, we embed them in a safety‑augmented approximate NMPC framework that calls the optimizer only when the predicted input sequence is not certified safe. This hybrid controller remains on average 6.8× faster than NMPC, with feasibility at least 99.8% for all architectures; SRLBC retains the lowest closed‑loop cost.
Under distribution shift, pure BC policies degrade sharply; SRLBC degrades the least and, when safety‑augmented, keeps feasibility above 98% with the lowest closed‑loop cost. Jacobian‑based sensitivity analysis shows the resulting controller relies on a sparse, less redundant set of latent features, each tied to a single group of physical variables and aligned with the system’s timescale decomposition.
Review