基于改进PSO-GA-BP神经网络的电弧炉吨钢电耗预测

Prediction of Electric Arc Furnace Power Consumption per Ton of Steel Based on Improved PSO-GA-BP Neural Network

摘要: 鉴于电弧炉冶炼时，难以精确预估吨钢电耗这一状况，特提出运用群体智能优化算法来优化机器学习的电弧炉吨钢电耗预测模型。基于某炼钢厂的生产数据，借助随机森林算法(Random Forest, RF)对输入特征的重要程度进行排序，挑选出贡献度较高的参数，作为BP神经网络的输入特征。同时，运用Pauta法则对原始数据展开预处理，基于GA-BP (Genetic Algorithm, GA)模型构建出精度颇佳的电弧炉吨钢电耗预测模型。此外，采用改良的粒子群算法(Particle Swarm Optimization, PSO)对GA-BP模型实施二次优化。将所构建的改进PSO-GA-BP模型，与传统电弧炉吨钢电耗预测模型、BP神经网络以及未二次优化的GA-BP模型进行比较。结果显示，改进的PSO-GA-BP模型相较于其他模型，具备更高的预测精度以及良好的泛化能力。此外，通过对于改进PSO-GA-BP模型的SHAP (SHapley Additive exPlanations)可解释性分析，打破了传统黑盒模型的壁垒，增强了模型的可解释性，有效缓解了模型的信任危机。

Abstract: In view of the fact that it is difficult to accurately predict the power consumption per ton of steel during the smelting of electric arc furnace, a swarm intelligence optimization algorithm is proposed to optimize the prediction model of electric arc furnace power consumption per ton of steel. Based on the production data of a steelmaking plant, the importance of the input features was sorted with the help of random forest algorithm (Random Forest, RF), and the parameters with high contribution were selected as the input features of BP neural network. At the same time, the Pauta rule was used to preprocess the original data, and a prediction model of electric arc furnace power consumption per ton of steel with good accuracy was constructed based on the GA-BP (Genetic Algorithm, GA) model. In addition, the improved particle swarm optimization (Particle Swarm Optimization, PSO) algorithm was used to implement the secondary optimization of the GA-BP model. The improved PSO-GA-BP model was compared with the traditional electric arc furnace power consumption prediction model, BP neural network and GA-BP model without quadratic optimization. The results show that the improved PSO-GA-BP model has higher prediction accuracy and good generalization ability than other models. In addition, through the SHAP (SHapley Additive exPlanations) interpretability analysis of the improved PSO-GA-BP model, the barrier of traditional black-box models is broken, the interpretability of the model is enhanced, and the trust crisis of the model is effectively alleviated.