A federated learning-enabled distributed optimal control mechanism for energy supply chains

Abstract

This paper addresses the privacy barriers of underlying time-series data not leaving the domain and the limitations of macro-network interaction latency faced by the energy supply chain in the distributed interconnection of multi-source heterogeneous devices. A distributed collaborative optimization control strategy mechanism for microgrid clusters based on a federated learning architecture is designed. By coupling a proximal regularization dimensionality reduction operator with an adaptive attention aggregation alignment algorithm based on the high-dimensional space cosine angle, the correlation data dependency between the microgrid's local real load time-series characteristics and the cloud-based global scheduling command update closed loop is decoupled. Test data from a digital twin-on -a-loop platform containing an ARM heterogeneous computing cluster shows that, while ensuring local data physical isolation, this strategy reduces the time consumption of a single closed-loop control inference link to 38.5ms, and suppresses the number of iterations for the global network strategy algebraic optimization error gradient to complete convergence to 42. This application mechanism reduces the average daily comprehensive economic operating cost ratio of the system's microgrid test cluster by 7.6% to the $1985.6 mark. Under extreme combined disturbance stress test conditions with a communication packet loss rate of 30% and a bottom-level sensor acquisition offset ratio of 20%, the maximum extreme value of the upper limit of the physical jitter variance increment of the control command quantity is limited to within the range of 14.2%, providing a standard benchmark for covert security control under a massive high-dynamic node network architecture.