Time-series rolling balanced optimization based on machine learning prediction power distribution control for modular power architecture

Abstract

Modular power architecture has gained significant research attention for its module deployment and control strategies. This study presents a learning-based balanced optimization control strategy for a modular system using multiple hydraulic free piston engines. By integrating learning-based energy management with online sequential rolling control, the approach achieves efficient power system operation while balancing power unit runtime and reducing computational costs. First, prediction models were trained using optimal power distribution data with Bayesian optimization applied to tune hyperparameters, significantly improving prediction accuracy. An online time-series rolling control strategy was then introduced to mitigate uneven power unit operation. Combined with the machine learning prediction model, this balanced power distribution strategy reduced real-time computation time. Finally, Hardware-in-the-loop tests validated the strategy’s real-time performance and effectiveness. Results show the Bayesian-optimized artificial neural network achieved over 95% prediction variance at real and normal condition. The online strategy reduced root mean square error and mean absolute error by 61.2% and 77.8%, respectively, compared to unbalanced optimization. The learning-based power distribution improved system efficiency by over 5.5% versus rule-based and heuristic methods, while cutting average real-time computation time by over 60% compared to prior optimization-based strategies, effectively freeing computational resources. Overall, the proposed strategy maintains significant comprehensive advantages over existing methods.