Multi-objective optimization of high-quality lithium extraction from lepidolite roasting based on neural network coupled modeling
The rotary kiln roasting of lepidolite for lithium extraction faces challenges of unstable lithium conversion rates and high energy consumption. To address this, a multi-objective optimization method coupling improved neural network simulation with a multi-objective genetic algorithm was proposed, targeting the synergistic optimization of lithium conversion rate (TRLi) and natural gas consumption intensity (EIng). Using long-term industrial time-series data of batching parameters and kiln operating variables, back-propagation (BP) neural network and its particle swarm optimization (PSO) improved variant were developed to model TRLi and EIng. The PSO-BP model demonstrated superior accuracy in capturing the complex nonlinear relationships, reducing mean absolute percentage errors (MAPE) to 0.278 and 0.284 for TRLi and EIng, respectively. Subsequently, the non-dominated sorting genetic algorithm II (NSGA-II) was employed to construct a multi-objective optimization model, yielding a Pareto-optimal set of process parameters that maximize TRLi and minimize EIng. The results revealed that under NSGA-II optimized conditions, TRLi could be stabilized between 82.45% and 87.96%, an average increase of 3.61 percentage points over baseline operations, while EIng could be reduced to 53.7 m3 per ton of clinker. For an annual processing capacity of 3.2×105 tons of lepidolite concentrate and sulfate mixture, this corresponds to an additional 127.1 tons of lithium metal recovery, a reduction of 1,964,912 m3 in natural gas consumption, and a decrease of 3,763.84 tons in CO2 emissions annually. This study provides theoretical and technical support for the green, high-quality, and low-carbon supply of critical raw materials for the lithium battery new energy industry.