摘要
机器学习方法已经在生物质气化建模中展现出广阔的应用前景。然而,机器学习模型主要依赖于实验数据,并不考虑气化中的反应机理,在数据样本不充分的情况下模型所表现出的实际关联特性与机理规律之间存在严重偏差。为此,提出一种基于物理信息神经网络(PINN)的生物质气化产物分布预测方法,该方法将真实实验数据与先验机理进行无缝衔接,在人工神经网络(ANN)模型中嵌入边界约束和关键参数间的单调性关系,通过自动微分技术进行辅助优化,实现模型的高效训练。结果表明:PINN模型的决定系数大于0.89,均方根误差小于4%,其总体预测精度要优于随机森林(RF)、支持向量机(SVM)和ANN 3种纯拟合机器学习模型;PINN模型能够严格服从边界约束和先验机理单调性关系,表现出更好的可解释性和泛化能力。
Machine learning methods have demonstrated promising applications in biomass gasification modeling.However,machine learning models primarily rely on experimental data and do not consider the reaction mechanisms in gasification.In situations that data samples are insufficient,there can be significant deviations between the actual correlation characteristics exhibited by the model and the mechanistic laws.Thus a method for predicting biomass gasification product distribution based on physics-informed neural networks(PINN)was proposed.This method seamlessly integrated real experimental data with prior mechanistic knowledge,embedding boundary constraints and monotonic relationships among key parameters into the artificial neural network(ANN)model.Automatic differentiation techniques were used to assist optimization,enabling efficient model training.Results show that the PINN model achieves a coefficient of determination greater than 0.89 and a root mean square error less than 4%,the overall prediction accuracy is superior compared to the three models which are purely fitting based on machine learning:random forest(RF),support vector machine(SVM)and ANN.Furthermore,the PINN model strictly adheres to boundary constraints and prior mechanistic monotonic relationship,exhibiting better interpretability and generalization capabilities.
作者
邓志平
任少君
翁琪航
朱保宇
司风琪
DENG Zhiping;REN Shaojun;WENG Qihang;ZHU Baoyu;SI Fengqi(MOE Key Laboratory of Energy Thermal Conversion&Control,Southeast University,Nanjing 211189,China)
出处
《动力工程学报》
CAS
CSCD
北大核心
2024年第5期719-726,共8页
Journal of Chinese Society of Power Engineering
基金
国家自然科学基金资助项目(青年项目)(52306230)。
关键词
生物质气化
机器学习模型
物理信息神经网络
机理约束
biomass gasification
machine learning model
physics-informed neural network
mechanistic constraint