Data-based Prediction Methods For Secondary Energy System In Steel Industry And Their Applications

Steel industry contains high energy consumption and emission, and its related energy resource conservation and pollution reduction are the key issues of national industrial development. Researches on prediction of secondary energy system in steel industry are beneficial not only to energy scheduling and optimization, but also to energy-saving and emission reduction. Data-driven prediction methods for the secondary energy system are studied in this dissertation, due to the fact that it is rather difficult to model such a system by using the process mechanisms because of their complexity.First, point-oriented prediction for the generation or consumption flow of the secondary energy system units can be viewed as noisy time series prediction problems, giving the fact that the industrial data contain noises. An improved echo state network (ESN) model with noise addition is proposed, in which the additive noises describe the uncertainties of the internal states and the output. Considering the update uncertainties of the internal states and parameters caused by data noises, a CKF/KF based dual estimation is proposed to perform the supervised learning.Second, relational data based point-oriented prediction model is developed here for the level of the reservation units of the secondary energy system. Considering the manual interference coming from the adjustable gas users, a conditional fuzzy clustering is adopted to partition the input and output space. With the introduction of fuzzy concept, the proposed model is adaptive for industrial noises. Then, a Bayesian linear regression is designed to determine the parameters of the consequent part of the proposed model, which can effectively avoid the ill-conditioned phenomenon.Third, a bootstrap ESN ensemble consisting of a set of ESN individuals is proposed to construct the prediction intervals (PIs) for the secondary energy generation and consumption, and the0.632bootstrap cross-validation is adopted here to determine the number of the ESN individuals and the reservoir dimensionality. To estimate the parameters of the proposed model, a simultaneous training method based on Bayesian linear regression is also developed, which learns the parameters of ESN individuals simultaneously with information exchanges.Forth, a model based on Bayesian ESN with input uncertainties is proposed in this dissertation to construct the PIs for the secondary energy storage levels. The prediction distributions with consideration of the input data uncertainties, the feedback data uncertainties and the total uncertainties are derived, respectively. Finally, the PIs for the energy storage level are constructed with consideration of the total uncertainties.Finally, the proposed prediction methods are then verified based on the real world data coming from the industrial real-time database, and the results show that the proposed methods present the better performance for the prediction of secondary energy system. The developed software system based on the above researches is applied in the energy center of a steel plant, whose operational state shows that the prediction results exhibit a great significance for the optimization and scheduling of the secondary energy system in iron and steel enterprises. Furthermore, the prediction system plays an important role in achieving energy conversion and production cost reduction.