The Application Of LSSVM Based On Parameter Optimization On LDG System Prediction

Energy cost has been the main part of the total cost in steel enterprises, in which the effective recycling and dispatching energy are the key issue to balance the generation and consumption of energy and reduce the cost greatly. The scientific dispatch relies on accurate prediction for the generation-consumption of gas system and gas holder level. At present, the byproduct gas is the significant part of energy optimal scheduling problem. LD converter gas (LDG) is being widely used as the secondary fuel for a large number of equipments in steel production. Since there are a great number of gas users and a large fluctuation in LDG system, the accurate prediction for the LDG system would be a big challenge for the current enterprises. Therefore, precisely and efficiently predicting LDG system, for the balance of the whole energy system in steel industry, becomes a significant research goal.In this paper, on the basis of the background of Shanghai Baosteel, the users of gas system are divided into three types by manufacturing technique and generation-consumption's characteristics, which consists of production users, consuming users and adjustable users. As for the production users, a class of fitting method based on signal trigger is developed to predict the generation flow; as for the consuming users, LSSVM based on time series is proposed; and an average value method is used to predict the adjustable users. Furthermore, a relational model is built to predict the gas holder level of LDG based on multi-users prediction. Then, an intensive study is performed for the optimization of model's parameters. With respect to the time delay and embedding dimension of phase-space reconstruction in time series, the mutual information method and Cao method are used to optimize the parameters respectively in the study. For the super-parameters of LSSVM model, the particle swarm search based on fast leave-one-out is proposed, and we build the dynamic prediction model for these users who have a big fluctuation. Compared to other methods, such method obtains a better prediction performance.Combined with the requirement of prediction model, the application software for the LDG system prediction is developed, which consists of server and client terminal. The server terminal predicts the LDG system (including all users and gas holder level) at regular intervals and the client one presents the predict data through the user interfaces. The prediction software has been applied in the Energy Center of Baosteel and the running results demonstrate the effectiveness of the proposed methods.