Prediction Model of Blast Furnace Molten Iron Temperature Based on Time Series Data

摘要

The thermal controlling for a blast furnace is very important. But it is hard to measure the thermal state inside the blast furnace. The molten iron temperature can be measuring and corresponds with the heat of the blast furnace. From when the ores are feed in to the molten iron, the process needs 6~8 hours. The molten iron temperature is a lagging indicator and difficult to predict by physical model. This experiment proposes an AI model training by two time series models, long short-term memory (LSTM) / Gated Recurrent Unit (GRU), to predict the molten iron temperature with in two hours. Before training the AI model, there are 66 features selected from a total 2600 features by dynamic time warping (DTW). The 66 features have different response times that were analyzed to find their critical response time within 8 hours of every feature via the feature importance method of the xgboost (XGB). The trained AI models with the 66 features and the critical response times can predict the molten iron temperature in 2 hours. The mean absolute errors of the two models of the LSTM and GRU are 25 and 20.5. The trained model of the GRU is better than the LSTM.