Research On Forecasting Method Of Capacity Planning Of Steel Enterprises Based On Deep Learning

Demands of overall planning and optimization of intelligent manufacturing and capacity resources are increasing for steel enterprises under the background of industry 4.0.China’s iron and steel enterprises urgently need to shift from the traditional idea,sales determined by production,to the direction of production and sales balance.Therefore,the plan of capacity resource is going to assume greater responsibility and play a greater role in smart manufacturing with no doubt,which not only provides new development space,but also brings new challenges for capacity resource planning business in the future.In recent years,a series of theories and methods represented by deep learning have been more and more applied to the real industrial manufacturing field with the rapid development of artificial intelligence technology,providing an unneglectable driving force for the development of intelligent manufacturing.Capacity planning is generally considered to be an optimization issue in the field of operations research.It is difficult to establish accurate mathematical models in the practical application of steel companies because of fuzzy rules,complex and changeable constraints and other factors.The capacity plan of the steel enterprise is the plan amount setting of each flow for a certain period in the future,which can be attributed to time series forecasting through the analysis of historical data of capacity planning.The deep learning method has achieved good results in the application of time series prediction problems in the past few years.Therefore,this thesis studies the prediction of capacity planning based on the deep learning method.The main research contents are as follows:(1)Field investigation as well as data collection and processing.Based on the capacity demand planning of a large steel company in China,this thesis analyzes the business logic and status quo of the production capacity planning,and attributes this kind of problem to a time series prediction problem.The factors that have a major impact on the capacity planning can be identified by analyzing the actual data and combining the actual situation of the business,so as to find out the available data items related to the results.Then Normalize the data and other preprocessing operations to complete the data preparation and cleaning work.(2)A capacity planning prediction model based on Long Short-Term Memory(LSTM)structure is proposed.Considering the essence of the task is to predict a series of time series,LSTM is taken as the basic model based on the cyclic neural network,and the Bi-directional Long Short-Term Memory(Bi-LSTM)structure is compared with that of the ordinary LSTM as a result of that the later period data will have an impact on the previous data in the actual business.The feasibility of using LSTM model to solve this problem is verified by experiments,and the conclusion that Bi-LSTM model has better performance in this thesis is obtained.(3)Through the improvement of Bi-LSTM model,a capacity planning model based on LSTM with attention mechanism and two-dimensional convolution(AM2DC-LSTM)is proposed.Firstly,features are considered to be strengthened and enriched on the basis of BiLSTM model.Two dimensions of time and space are added into the convolutional neural network structure.The incomplete bidirectional temporal convolutional network designed in this thesis plays a significant role in improving the accuracy of the model.Secondly,the model is modified by adding attention mechanism,which improves the training speed and accuracy of the model.The experimental results show that the improved model achieves 7.7%accuracy improvement and nearly 10%acceptance improvement compared with the basic LSTM model,which verifies the effectiveness of the improved scheme.(4)Using the AM2DC-LSTM model as the base learner,a set of ensemble learning solutions based on results convergence was proposed.From the perspective of enriching the input feature content,the iterative calculation is performed using the prediction results of the previous model to obtain the inventory situation within the planning period.Then combine the predicted inventory situation with the original data to train a new phase model.This process is iterated to obtain multiple stage models based on the AM2DC-LSTM model architecture.For multiple stage models,linear regression is used to find the appropriate weight coefficients.The input results of the multiple stage models are fused according to the weight to improve the accuracy of the prediction results.It is proved by experiments that the AM2DC-LSTM model can improve the accuracy of 4.14%and the acceptability of about 5%.Finally,the experiment verifies the generalization of the model to such models.