Hybrid algorithm-optimized neural network for predicting tar-rich coal pyrolysis products with pareto optimization

To improve the prediction accuracy of pyrolysis product distribution from tar-rich coal and optimize the process parameters, this study constructed a specialized database for tar-rich coal pyrolysis. A hybrid strategy integrating the Sparrow Search Algorithm and Particle Swarm Optimization algorithm (SSA-PSO) was proposed to optimize the weights and thresholds of a BP neural network, establishing an SSA-PSO-BP prediction model. The resulting SSA-PSO-BP model significantly outperformed standard BP, PSO-BP, and SSA-BP models in predicting three-phase pyrolysis products. On the test set, the model achieved a coefficient of determination (R²) of 0.9147. Its comprehensive evaluation (S_core)decreased by 17.5% and 9.9% compared to PSO-BP and SSA-BP, respectively. SHAP and partial dependence analysis revealed that volatile matter, carbon and hydrogen content, H/C atomic ratio, and pyrolysis temperature are key features influencing product distribution. The model successfully captured nonlinear relationships and interaction effects among variables. Based on the optimized model, the pyrolysis conditions for typical tar-rich coals from Xinjiang, Shaanxi, and Inner Mongolia were optimized. Objective function optimization yielded optimal pyrolysis temperatures of 602.17 ℃, 537.32 ℃, and 630.21 ℃, with corresponding tar yields of 17.43%, 16.77%, and 12.86%, respectively. Pareto front optimization was further conducted using the Non-dominated Sorting Genetic Algorithm II (NSGA-II), which revealed the competitive relationship between tar and gas yields and clarified how different pyrolysis temperatures and coal properties affect this trade-off.This work provides a data-driven method and theoretical basis for accurate prediction and targeted regulation of tar-rich coal pyrolysis.