| With the development of power systems, the tradeoff among objectives such as economy, security, stability and reliability has become more and more common, and the solving difficulty is also rising constantly. As a kind of stochastic method, Computational Intelligence (CI) offers an outstanding robustness and opens up a new way for solving the multi-objective optimization problem (MOP). A systemic CI-based method is proposed in this paper, which focuses on two major areas: artificial neural network (ANN) and multi-objective evolution algorithms (MOEA). The main work includes the following four aspects:Firstly, MOEAs such as NSGA-II, PESA, SPEA2 and NMPSO have different advantages in search ability and convergence speed, among which exists a strong complementarity. This paper introduces a coevolution mechanism and puts forward a multi-objective coevolutionary algorithm MOCEA. Our core idea is using various MOEAs to solve MOP in a complementary way, so as to enhance the overall search ability. Along this idea of thought, an ecosystem model is established, where multiple populations compete and exchange their dominant genes in addition to evolution under different mechanisms.Secondly, realistic MOPs in power systems are usually extremely data-intensive. To improve the optimization speed, a three layer parallel architecture called Master-Keeper-Slave is proposed in this paper, which devides involved processors into three categories:1) The Master taking on initialization and elite competition; 2) Keepers taking on task allocation, fitness assignment, dominant genes exchange and generating offspring.3) Slaves taking on objective functions calculation. The structure of the proposed coevolutionary algorithm is modified based on MPI and a parallel algorithm is brought into being.Thirdly, considering the fact that existing diversity strategy of MOEAs are highly dependent on objective functions calculation, which brings a high time consumption, a sparse Autoencoder clustering model based on multiple hidden layers ANN is proposed, which can be trained to study the intrinsic relevancy between genotype and phenotype using a deep learning method called wake-sleep algorithm which includes two steps of bottom-up unsupervised learning and top-down supervised learning. The proposed diversity strategy makes it possible to fast cluster populations directly based on genotype without objective functions calculation.Finally, the modeling and solving procedure of three realistic MOPs in power systems are narrated in detail, included:1) Multi-objective Optimal Power Flow (OPF):A multi-objective OPF method considering transient stability is proposed, which models stability as an objective function rather than an inequality constraint and consider classic transient stability constrained OPF as a tradeoff procedure using Pareto ideology. Case study results show MOCEA has a better diversity of solutions in OPF analysis and the proposed parallel algorithm has an excellent acceleration performance on data intensive problems.2) Current limiters configuration:Current limiters are allocated considering the total investment, the level of short-circuit current, as well as the transient stability using the proposed MOCEA combined with a search space reduction technique based on sensitivity. Case study results show the proposed sparse autoencoder clustering based on deep learning is a feasible method for MOEA diversity maintenance.3) Power supply system planning in microgrid:Discrete probability distributions are used to model the random factors in microgrids and a multi-objective planning model is established with objectives of minimization the investment, minimization the nodal voltages violating probability and electricity inadequacy probability. The proposed planning model is solved by MOCEA combined with an improved probabilistic load flow algorithm.This paper has been fully demonstrated that CI is a feasible method to solve the MOPs considering economy, security, stability and reliability in power systems. |
PDF Full Text Request