Multi-objective generation control and optimization for large-scale power system dispatch

xv, 179 leaves : ill. (some col.) ; 30 cm.PolyU Library Call No.: [THS] LG51 .H577P EE 2013 ZhouWith the renew interest in energy-saving generation dispatch and the growing environmental concern for the power industry, developing a modernized dispatch infrastructure, associated with its optimized generation scheduling and dispatch strategies, have become a global priority to contribute towards the formation and development of the future smart grid. Under the smart and green grid paradigm, automatic generation control (AGC) and electric power dispatch will play an important role in improving the long-term control performance and sustainability of energy management system (EMS) for the efficient operation of large-scale interconnected power systems. Therefore, based on the foundation laid by the pioneering research studies already presented, this thesis strives to make further investigations for the design of the AGC strategies and power dispatch algorithm with multi-objective operation in mind. The AGC performance in interconnected power system operation is used to measure against the control performance standards (CPS) released by North American Electric Reliability Council (NERC) in 1997. Since the introduction of this new NERC CPS, there are fundamental changes in the conventional AGC control philosophies. So far, researches on AGC strategies under CPS were mostly based on the classical proportional-integral (PI) control structure. Even with the wide adoption of the CPS nowadays, the existing AGC systems have commonly not yet been optimized to fully explore the potential of CPS standards. Dynamic response studies showed that the AGC system under CPS in fact can better be formulated as an uncertain stochastic system from the statistical and probabilistic point of view. Consequently, this thesis harnesses reinforcement learning (RL) and Markov decision process (MDP) techniques to develop a robust and adaptable AGC optimized for the NERC's CPS standards and optimal relaxed operation. Over the years, in order to determine the optimal steady-state operation of dispatchable generators, economic dispatch (ED) is a standard function in the EMS so that the total generation cost is minimized while satisfying a set of operational and physical constraints. However, nowadays more operating objectives, such as energy conservation and emission reduction, should also be considered to establish a multiobjective power dispatch (MOPD) optimization. Consequently, this thesis would also propose a new algorithm based on the Pareto optimality to solve this highly constrained large-scale MOPD problem with multiple contradictory and noncommensurable objectives.First and foremost, in this thesis, a new concept referred as optimal relaxed AGC control is proposed to allow the AGC plants to maneuver in less costly manner in finding the optimal economic dispatch policy on the premise of complying with the CPS1 and CPS2 metrics. A Q-learning based AGC controller, in which the CPS control and relaxed control objectives are formulated as multi-criteria reward function via linear weighted aggregate method, is presented for interactive self-learning control rules to maximize the long-term discounted reward. In addition, for the thermal-dominated power systems, to overcome the long time-delay problem caused by the steam turbine of AGC thermal units in the secondary frequency control loop, a multi-step Q(λ) learning based AGC is proposed to regulate the degree of CPS compliance and relaxation for the desirable relaxed control. The effectiveness and validity of the proposed RL based AGC strategies have been successfully verified on a two-area load frequency control (LFC) model and the practical-sized China Southern Power Grid (CSG) with four control areas. The goal of average reward RL is to maximize the long-term average rewards of a generic system. This coincides with the design objective of the CPS which was established to improve the long-term performance of an AGC used for real-time control of interconnected power systems. Therefore, a novel R(λ) imitation learning (R(λ)IL) method based on average reward optimality criterion is proposed to develop an optimal AGC under the CPS. This R(λ)IL based AGC is capable of operating online in real-time with high CPS compliances and fast convergence rate in the imitation pre-learning process. Its capability to learn the control behaviors of the existing AGC by observing the system variations enable it to overcome the main difficulty in practically applying the conventional RL controllers, in which an accurate power system model is required for the offline pre-learning process, and significantly enhance the learning efficiency and control performance for power generation control in various power system operation scenarios. On the other hand, an equilibrium-inspired multiple group search optimizer (MGSO) is developed to solve the highly constrained MOPD problem. In this algorithm, a stochastic learning automata based synergistic learning is employed to achieve the information interaction and credit assignment among groups for cooperative search, and an average linkage-based hierarchical clustering is used to provide the power dispatcher with a manageable and representative Pareto-optimal front (PF). Furthermore, the Nash equilibrium point is first introduced to identify the best compromise solution from the PF. An alternative constraint handling, which separates constraints and objectives with different search strategies, is presented to orient towards a well-distributed PF. In addition, two special implementations, space reduction strategy and chaotic sequence dispersion, have been incorporated in convergence process to facilitate the local exploitation and global exploration of PF solutions respectively. Simulation tests on the benchmark power systems, including the IEEE 30-bus system with 6 units and the IEEE 118-bus system with 54 units, have demonstrated the enhancement and superiority of the proposed MGSO algorithm, and confirmed its potential to cope with this type of large-scale multiobjective optimization problems with high-dimensional and more objective functions.Department of Electrical EngineeringPh.D., Dept. of Electrical Engineering, The Hong Kong Polytechnic University, 201