Robust Imitation Learning for IIoT-Enabled Power Plant Control Under Packet Loss and Transmission Delay

Abstract

Deploying learning-based controllers over Industrial Internet of Things (IIoT) networks exposes the control loop to packet loss and stochastic transmission delays that corrupt the spatiotemporal sensor observations on which modern imitation learning policies critically depend. We propose RCE-GAIL (Robust Communication-Enhanced Generative Adversarial Imitation Learning), a framework that natively integrates IIoT channel awareness into the policy learning pipeline for combined cycle power plant control. The core contribution is a communication-aware spatiotemporal encoding mechanism comprising two coordinated components: an availabilityweighted convolutional encoder that normalizes spatial feature extraction by instantaneous channel availability density, preventing feature collapse under high packet loss; and a faultaware recurrent state propagation module whose cell update interpolates between new observations and prior memory as a continuous function of window-level availability, gracefully degrading to memory retention under near-total channel loss. A delay-aware temporal alignment module further compensates for gateway-induced inter-stream latency before window construction, restoring cross-channel causal structure. Validated through a dual-dataset protocol combining PSID power plant simulation with empirical fault profiles extracted from the TON_IoT IIoT benchmark, RCE-GAIL sustains tracking error below 5% at packet loss rates up to 30%-a 58.5% reduction over the strongest baseline-while incurring no performance penalty under fault-free operation.