Learning-Based Control of Hybrid Fuel Cell Power Plant

摘要

Direct fired Solid Oxide Fuel Cell (SOFC) Turbine hybrid plants have the potential to dramatically increase power plant efficiency, decrease emissions, and provide fast response to transient loads. The US Department of Energy's Hybrid Performance Project is an experimental hybrid SOFC plant, built at the National Energy Technology Laboratory. One of the most significant challenges in the development and commercialization of this plant is control. Traditional control techniques are inadequate for this plant due to poor system models, high nonlinearities, and extreme coupling between state variables. Learning-based control techniques do not require system models, and are well suited for controlling nonlinear and highly coupled systems. In this work, we use neuroevo-lutionary control algorithms to develop a controller for this plant, and demonstrate the controller can accurately track a desired turbine setpoint profile within 50 RPM, even in the presence of 10% sensor noise. In order to ensure the neuroevolutionary algorithm is computationally tractable, we develop a computationally efficient neural network simulator of the plant, using data collected from actual plant operation.