Data-Driven Modeling And Intelligent Optimization For Blast Furnace Ironmaking Process

With the rapid development of intelligent manufacturing and industrial Internet,it is possible to further improve the production level of iron making process and achieve the digital transformation.Industrial Internet platforms and algorithms of industrial data analysis,modeling and intelligent optimization carried on the platforms,as the core of smart manufacturing and industrial Internet,are receiving more and more attention.In order to realize the intelligent empowerment of blast furnace iron-making process and achieve the purpose of energy saving,emission reduction and quality improvement in the production process,this paper researched the data modeling and intelligent optimization of blast furnace iron-making process from the perspective of industrial data analysis.The prediction of the key production indexes of the ironmaking process was realized by data modeling,which provided guidelines for the timely adjustment and optimization of the production process.The constrained optimization problem of the ironmaking process is researched on the basis of the well established models.In order to realize the technology application,sharing and reuse,as well as to promote the digital transformation of the iron-making industry,an industrial Internet cloud platform is built.And the algorithms are encapsulated and deployed on the platform.The main research content of the full paper specifically contains the following parts.(1)To address the problem that most existing modeling methods focus on solving the dynamic and nonlinear problems of the blast furnace process,but fail to adequately consider the impact of different correlations between process variables and quality variables,as well as deal with the mismatch problem of input and output variables due to the large time lag of the blast furnace production process,a context-aware enhanced GRU model based on feature-temporal attention is proposed.The model based on feature-temporal dimensional attention mechanism can effectively express the nonlinearity and dynamics of the process,and solve the input-output variables time mismatch problem caused by large lag of the production process.The proposed method can automatically assign attention weights to different input variables through data learning,so that the key variable information can be fully exploited.In addition,differing from the traditional point wise attention computation methods,the paper proposes a causal convolution-based self attention computation method,which extracts local context information to the attention computation and achieves a more reasonable attention allocation in the temporal dimension.The results of the experiments proved that the method effectively improved the index prediction accuracy.(2)An optimization algorithm based on generative adversarial network distribution mapping is proposed for the problem with non-analytic constraints that the analytic formulas of the optimization constraints of the blast furnace ironmaking process are difficult to obtain and the degree of the feasibility of the optimization solution is difficult to quantify.The algorithm transforms the constrained optimization problem into an unconstrained optimization problem by establishing a mapping from multidimensional independent space to the non-analytic constrained multidimensional coupled physical space.For the generation data boundary overflow problem,a two-stage generation model is proposed to establish the mapping from the unconstrained independent space to the constrained physical space,which effectively ensures the consistency between the generated data distribution and the practical data distribution.The infeasible solutions in the physical space are eliminated by means of the space mapping,which reduces the search space and improves the optimization capability.A high-performance optimization method under the condition of known solution space density distribution,the density amended gray wolf algorithm,is proposed.The search step of this algorithm can be adaptively adjusted with the variation of solution density of the search space,achieving a balance of search speed and accuracy.The method was verified by numerical and practical data,and the results proved the effectiveness of the proposed algorithm.(3)Consistent optimization framework based on controllability guaranteed modeling is proposed to address the problem of unavailability or high uncertainty of optimization solutions due to the uncontrollability of some auxiliary variables for the predictive modeling of datadriven heuristic optimization methods.The optimization method can achieve the information extraction of uncontrollable variables in the form of process supervision to achieve the improvement of prediction accuracy of quality index distribution while ensuring that uncontrollable variables are not used as inputs in the prediction stage of the model.A semi-supervised autoencoder regression method is proposed,which combines the autoencoder and regression model into an end-to-end integrated model,with the regression guiding the feature construction in the encoding phase to further improve the model prediction accuracy.A consistent optimization framework is proposed in combination with the multi-objective gray wolf algorithm.With this method,reliable optimal solutions and confidence estimates of the solutions can be obtained.(4)To address the problems existing in the blast furnace ironmaking process,such as scattered production elements,isolated information islands,weak innovative application development capability,difficult accumulating and reuse of knowledge and methods,an industrial Internet platform is constructed to facilitate information fusion,microservice packaging and sharing,and the development of innovative applications.Virtualization technology is used to divide the hardware resources.Then the container runtime environment is deployed inside the virtual machine.Next,a computing cluster is built through Kubernetes to achieve container orchestration management.After the deployment of Rancher for the management of multiple clusters,the platform is completed.A process parameter optimization method based on the digital twin technology of industrial Internet is proposed.The development and deployment of the optimization APPs are completed using the microservice component encapsulated on the platform.And the application of the optimization algorithms for the blast furnace ironmaking process are realized.