SecureDec: A Decentralized Scheduling Pipeline With Federated Learning and Efficient Encryption for Electricity-Gas Coupled Systems

This paper proposes SecureDec, a decentralized and secure computational pipeline for the day-ahead scheduling of electricity-gas coupled systems. At the lower level, we propose SecureADMM for the decentralized optimization of the coupled subsystems, where all exchanged variables in decentralized ADMM are properly secured with ex-ante Diffie-Hellman (D-H) keys. The generation of D-H keys, being the most time-consuming step in encryption, can be independently performed ahead of optimization, allowing subsystems to collaboratively update variables without compromising optimization efficiency. As the convergence of ADMM is contingent on the quality of starting points, at the upper level, we introduce multi-entity SecureBoost, a decentralized federated learning framework, to predict the warm starts of SecureADMM in a similarly decentralized and privacy-preserving manner. A two-stage statistical packing algorithm is incorporated in the preceding SecureBoost training process to improve the efficiency in data protection, which reduces the frequency of time-consuming encryption/decryption by 2–6 times. Case studies demonstrate that the convergence of SecureDec is up by 50.0% higher than the conventional ADMM due to hyper-accurate warm starts provided by SecureBoost. Further, 63.9% and 96.7% efficiency boosts are observed in the federated learning and day-ahead optimization of an electricity-gas coupled system, respectively, with no compromises in information privacy and scheduling quality. The pipeline is further scalable to larger systems, and demonstrates potentials in addressing non-convergence in the conventional ADMM.