Tin Smelting Process Parameter Optimization Based on Offline-to-Online Reinforcement Learning

Improving direct recovery rate while reducing energy consumption remains a fundamental challenge in tin smelting, owing to the strong nonlinearity and complex interactions among process variables. To address this issue, this study proposes an offline-to-online reinforcement learning framework for multi-objective optimization of tin smelting process parameters. Spearman correlation analysis is first employed to identify the dominant variables influencing tin direct recovery rate and energy consumption. An Ensemble-bootstrapped probabilistic neural network (Ensemble-BPNN) is then developed to model the nonlinear relationships between operating parameters and process objectives. Based on the predictive model, an Offline-to-Online Reinforcement Learning algorithm with Policy Dynamic Fusion (O2O-PDF) is proposed. It combines offline policy pre-training with online policy refinement to efficiently identify Pareto-optimal operating policies. A hybrid decision-making strategy is further developed to select a preferred solution from the Pareto front. The strategy integrates the competing solution distance method and the entropy weight method. Industrial top-blown furnace experiments validate the effectiveness of the proposed approach. Compared with empirical operating parameters, the optimized parameters increase the tin direct recovery rate by 3.02% and reduce energy consumption by 5.67%.