H. Ghezel-Ayagh

Development of a stack simulation model for control study on direct reforming molten carbonate fuel cell power plant

A nonlinear mathematical model of an internal reforming molten carbonate fuel cell stack is developed for control system applications to fuel cell power plants. The model is based on principles of energy and mass component balances and thermochemical properties. Physical data for this model is obtained from a 2 MW system design that is a precursor to a demonstration fuel cell power plant running on natural gas at the City of Santa Clara, CA. The model can be used to provide realistic evaluations of the responses to varying load demands on the fuel cell stack and to define transient limitations and control requirements. Simulation results are presented for a transient response to a power plant trip at full load.

Training neuro-fuzzy boiler identifier with genetic algorithm and error back-propagation

A multi-layer neuro-fuzzy system presents identification of a drum type boiler. This identification provides a rule-based approach to approximate the boiler dynamics. The interconnections of neuro-fuzzy layers furnish these fuzzy rules. A genetic algorithm (GA) trains the neuro-fuzzy identifier and extracts the linguistic fuzzy rules from measured boiler data. This GA training takes the advantages of nonbinary alphabet and compound chromosomes to train the neuro-fuzzy identifier. An error backpropagation training methodology is chosen to tune the membership function parameters. This neuro-fuzzy identifier obtains time response similar to boiler model while it avoids mathematical complexity of model dynamics.

Intelligent predictive control of a power plant with evolutionary programming optimizer and neuro-fuzzy identifier

An intelligent predictive controller is implemented to control a fossil fuel power unit. This controller is a non-model based system that uses a self-organized neuro-fuzzy identifier to predict the response of the plant in a future time interval. The control inputs are optimized in this prediction horizon by evolutionary programming (EP) to minimize the error of identifier outputs and reference set points. The identifier performs automatic rule generation and membership function tuning by genetic algorithm (GA) and error back-propagation methods, respectively. This intelligent system provides a predictive control of multi-input multi-output nonlinear systems with slow time variation.

Performance implications of rapid load changes in carbonate fuel cell systems

There are several key performance objectives of molten carbonate fuel cell systems operating under load perturbations. Among these include: regulation of stack temperature, regulation of differential pressure between anode and cathode, maintaining acceptable fuel utilization and steam-carbon ratio, and electrical load tracking. Utilities are interested in rapid load cycling of carbonate fuel cell systems while the military is interested in the response of carbonate fuel cell systems under sudden application and removal of electrical load. An integrated fuel cell system may endure these types of disturbances while satisfying performance goals, depending largely on the control system. In this paper, the authors examine two important performance variables: fuel utilization and steam-carbon ratio, under both ramping operation and sudden increase in load. Setpoint control laws are proposed for determining the proper steam and natural gas flows corresponding to steady state or transient conditions. In the case of a sudden load increase, they illustrate a trade-off between good load tracking and good utilization/steam-carbon ratio when considering rate constraints on valves.

Neural Network Supervisor for Hybrid Fuel Cell/Gas Turbine Power Plants

The neural network (NN) supervisor is developed for online estimation of optimal feedforward (FF) control inputs and setpoints for hybrid fuel cell/gas turbine power systems. The approach consists of determining a neural network structure suitable for predicting FF control inputs and setpoints based on optimal operating trajectories. The optimal trajectories were obtained in a previous study via nonlinear dynamic optimization based on a dynamic power plant model. Determination of the NN structure involves an a priori decision of the type of NN, the overall topology of input/output pairing, definition of a training epoch, as well as an identification of the number of hidden layer neurons and the number of iterations for the training epochs. This allows for straightforward training of the NN using the global training method, which includes all power profiles to define an epoch. In addition to training the NN with all available data, the networks prediction capabilities were tested by training it with all but one dataset and then determining the prediction results based on the untrained dataset. Results from eighteen case studies show that the developed NN supervisor is capable of predicting the optimized FF and setpoint trajectories satisfactorily.

Plant-wide simulation of direct reforming molten carbonate fuel cell systems

A plant-wide simulation model for a molten carbonate fuel cell power plant is outlined in this paper. The simulator is being developed for intelligent control applications to fuel cell systems as distributed generators. The dominant thermodynamics and chemical reactions are modeled for the cell stack and balance-of-plant, including heat recovery unit and anode exhaust oxidizer. The model is based on a 2 MW demonstration plant that had been running on natural gas at the City of Santa Clara, CA, USA. The fuel cells in this design utilize direct reforming of methane gas through placement of internal reforming catalysts within the cells.

Application of neuro-fuzzy identifier for a fossil fuel boiler system

A multi-layer neuro-fuzzy system presents identification of a drum type boiler. This identification provides a rule-based approach to approximate the boiler dynamics from the experimental boiler data. The interconnections of neuro-fuzzy layers furnish these fuzzy rules. A genetic algorithm (GA) trains the neuro-fuzzy identifier and extracts the linguistic rules from measured boiler data. GA training takes the advantages of nonbinary alphabet and compound chromosomes to train the multi-input multi-output (MIMO) neuro-fuzzy identifier. The fuzzy membership functions need to be adjusted during the training to minimize the identifier response error. Error back-propagation training methodology is chosen to tune the membership function parameters. The identifier response is investigated in several operating points. This neuro-fuzzy identifier is implemented within an object oriented programming tool that provides portability of the identification process. Therefore, it is a strong candidate to substitute model-based identifiers in applications such as model reference control system or predictive control problem to reduce the required design time.

APPLICATION OF NEURO-FUZZY IDENTIFIER IN THE PREDICTIVE CONTROL OF A POWER PLANT

Abstract An adaptive predictive control methodology is applied for a fossil fuel boiler control. The control algorithm takes advantage of a neuro-fuzzy identifier system for prediction of the boiler response in a future time window. An optimizer algorithm based on evolutionary programming technique (EP) uses the identifier-predicted outputs and determines input sequence in a time window. The present optimized input is applied to the plant, and the prediction time window shifts for another phase of plant output and input estimation. The neuro-fuzzy identifier is trained to provide a good estimation of boiler outputs. Neuro-fuzzy rules and membership parameters are trained based on the data log, applying genetic algorithm and back-propagation, respectively. The obtained intelligent control system is highly structural and applicable on different boiler systems.

Experimental transient validation of a Direct FuelCell(R) stack model

This paper compares the dynamic responses of a first principle model for internally reformed molten carbonate fuel cell stack with laboratory test results. ne dynamic model is currently being utilized in the design of a Ship Service Fuel Cell (SSFC) power plant. It is being employed to study the transient characteristics of the power plant and to aid in the design of the control system. The tests were performed on a nominally rated 20-kW, 30-cell laboratory test unit with full-scale cell area. Transient tests consisted of a set of step load changes it two different operating points. The predicted multi-time scale behavior is seen in the experimental results, which show a good agreement with the computer model.

An interactive distributed simulator for commercial power plant

This paper presents the details of a power plant distributed simulator with interactive graphical user interface. To enhance the performance of a large-scale simulation model for a commercial power plant, centralized simulation on a single processor is replaced by several concurrent simulations distributed over a network. These decentralized simulations are defined as the subsystems: boiler, turbine-generator, feedwater, and condensate. To maintain the fidelity of a central integration, the subsystems communicate interaction data among themselves using the TCP/IP protocol via a shared memory. The shared memory is utilized further in developing a graphical user interface to provide interactive simulation control while masking the internal complexities and management of multiple simulation programs over a network.