A New Method to Predict Endpoint Phosphorus Content During Converter Steelmaking Process via Industrial Data and Mechanism Analysis

The control of endpoint phosphorus content is a crucial task in converter steelmaking, and accurately predicting the endpoint phosphorus content is essential for precise control. Due to the complex converter steelmaking process, which involves high-temperature physicochemical changes and a multi-variable, nonlinear, strongly coupled, and random interference system, neither mechanism modeling nor data modeling can accurately predict the endpoint phosphorus content. The paper proposes a multi-level integration method integrating metallurgical reaction mechanisms with industrial data to predict endpoint phosphorus content. The integration of mechanism and data occurs during data preprocessing and model construction to enhance accuracy and hit ratio. In the data preprocessing stage, the dephosphorization mechanism guides the selection of characteristic variables with strong correlation to target variables using a feature network structure model based on local linear embedding algorithm (LLE) to reduce the data dimension. A Back Propagation neural network (BPNN) algorithm is employed for building a data-driven model while genetic algorithms optimize its hyperparameters. The oxidation of phosphorus and its distribution between slag and metal are integrated into the fitness function of the genetic algorithm. During model testing, a composite loss function evaluates predictive ability by incorporating both data loss function and mechanism loss function. Test results conforming to the loss function ensure overall performance while adhering to metallurgical laws. Utilizing production data from 852 heats over one month yields a prediction accuracy for endpoint phosphorus content within +/- 0.005 pct at 93.2 pct, indicating excellent predictive capability.