Research On Improvement Of Differential Evolution Algorithm And Its Application In Load Distribution For Aluminium Tandem Hot Rolling

Differential evolution(DE) algorithm is a random and parallel search algorithm based on individual differences. DE algorithm has been applied to many fields because of its characteristics, such as simple structure, less control parameters, strong capability of global search, and so on. However, there are some drawbacks with DE, such as falling into local optimum easily, evolutionary stagnation, disability on multi-objective optimization problem, which have restricted the performances and applications of DE severely. Therefore, the research on improvement of the differential evolution algorithm has important theoretical significance and practical value.On the basis of the deep study on DE algorithm,three improved algorithms are presented in this thesis. Moreover, the proposed algorithms are applied in rolling force prediction and load distribution for aluminium tandem hot rolling of Mingtai Aluminium Industrial. The main contents of this thesis are summarized as follows:(1) Considering the problem that the mutation strategy and control parameters of the standard DE algorithm are fixed in evolutionary process, a self-adaptive DE algorithm based on exponential smoothing and chaotic mapping(ECADE) has been proposed. According to the success rate of producing better individuals by each mutation strategy in current generation, the exponential smoothing is employed to predict the success probability of each mutation strategy in candidate pool for next generation by ECADE. Then the roulette wheel selection is used to select mutation strategy from candidate pool for each individual in next generation according to the success probability. In addition, the function that can balance the capabilities between exploration and exploitation and the Logistic mapping are used to generate the control parameters. It has been demonstrated that ECADE has many benefits, such as faster convergence rate, higher convergence precision and the ability of balancing exploration and exploitation by testing by the benchmark functions.(2) In view of how to increase the diversity of the initial population, a self-adaptive DE algorithm based on symmetric Latin hypercube design(SLADE) has been designed. In SLADE, the initial population is initialized by the symmetric Latin hypercube design(SLHD). According to a larger probability, the mutation strategy assigned to each individual is randomly selected from the strategy list which consists of some mutation strategies that product better individuals, or selected from the candidate pool. Moreover, SLADE employs Cauchy distribution and normal distribution to generate the control parameters, and update adaptively according to the control parameters that product better individuals. Experimental results show that the optimizing ability of SLADE is better than other DE algorithms, and SLHD is effective for improving the performance of SLADE.(3) Aiming at solving the multi-objective optimization problem by DE algorithm, a multi-objective DE algorithm based on angle neighborhood(ANMODE) is proposed. The weak domination is introduced to obtain the capacity of solving the multi-objective optimization problem. To ensure the evolutionary direction of individual, the mutation operation is executed in angle neighborhood. Additionally, the distributivity of the approximate set optimized by ANMODE has been greatly improved with the maintenance mechanism of external archive. Experimental results show that the approximate set of the Pareto front deduced by the ANMODE algorithm has better convergence and distribution. Moreover, the performances of the proposed algorithm are the best among all testing algorithms.(4) In view of how to reduce the error existing in the traditional rolling force model, a prediction model of rolling force based on BP neural network optimized by SLADE(S-BP) is presented to improve the prediction precision. To improve the resistance to interference, the self-learning model is introduced into S-BP, and the stability of the rolling force prediction has been enhanced. Experimental results indicate that the prediction precision of S-BP is much better than the traditional rolling force model and BP neural network. Moreover, the robustness of S-BP also has been improved by the self-learning model.(5) Considering the problems that unreasonable load distribution and the slipping for finishing mills, the ECADE, SLADE and ANMODE algorithms are employed to optimize the load distribution for the aluminium tandem hot rolling. The load distribution has been considered as single objective optimization problem and multi-objective optimization problem, respectively. The above proposed algorithms are used to optimize the load distribution according to the different objective functions. Experimental results indicate the proposed method is more effective in dealing with the slipping phenomenon, load balance and shape improvement. These optimized load distributions provides a useful base for the production practice.