Towards secure and decentralized energy dispatch: fuel optimization using distributed learning for energy demand and storage strategies

For complex nonlinear systems that involve constrained resources distributed across multiple collaborating subsystems, achieving optimal resource allocation while preserving data privacy and maintaining operational constraints is a challenge. Conventional centralized optimization techniques require aggregation of data, relying heavily on centralized computation and full data visibility across subsystems. However, these approaches face limitations when scaled to distributed environments due to privacy concerns, communication overhead, and computational complexity. This dissertation addresses these limitations by proposing a novel framework that integrates distributed learning methods and decentralized nonlinear optimization strategies. Distributed learning allows individual subsystems to generate accurate predictive models locally without exposing raw data, providing data privacy and reducing communication burden. Leveraging these distributed predictive models, a decentralized optimization algorithm employing consensus-based coordination is developed, allowing subsystems to independently optimize resource allocations through iterative local computations while still converging toward a globally optimal solution. This framework is demonstrated through the application of fuel consumption minimization in power generation systems, a domain characterized by nonlinear efficiency curves, dynamic load demands, and complex operational constraints. Simulation results confirm the framework’s capability to achieve optimal, privacy-preserving, and scalable resource management, highlighting its applicability to broader classes of constrained resource optimization problems in decentralized environments.