Short-term forecasting in strip thickness of hotrolling is critical to rolling technology, so that dynamic controlcan be accomplished to increase production and improveproduct quality. Predicting strip thickness behavior has beenalways a challenging task due to its complex and non-linearnature. Autoregressive integrated moving average(ARIMA)model has been verified with a better short-term forecastingperformance, for the problem of low multi-step predictionaccuracy, the rolling strategy is proposed to update modelparameters, which develops rolling ARIMA(RARIMA) model.In addition to improve the overall forecasting accuracy of stripthickness, hybrid forecasting of time series data is considered.Hybrid forecasting typically consists of an ARIMA predictionmodel for the linear component of time series and a nonlinearprediction model for the nonlinear component. In this paper,back propagation neural network(BPNN) is further introduced to forecast the residual of RARIMA model, and rollingARIMABPNN(RARIMABPNN) continuous forecasting modelwill be developed, in which rolling forecasting mechanism isused. To the effectiveness of the comprehensive evaluationmethod, a stability evaluation index is presented, in addition,the proposed method is examined on the two groups of stripthickness data from 620mm strip finishing mill group of hotrolling and the results are compared with some of the basicforecast methods. The results show that the proposed hybridmethod could provide a considerable improvement for theforecasting accuracy and stability.
展开▼
