自动化学报2025，Vol.51Issue(6)：1329-1346,18.DOI:10.16383/j.aas.c240566
基于域适应物理信息神经网络的时间序列预测方法
Time Series Prediction Method Based on Domain Adaptation Physics-informed Neural Network
摘要
Abstract
Machine learning-based prediction methods usually achieve high fitting accuracy,but often suffer from limited model interpretability and poor generalization performance.In industrial processes,the stability of these methods is affected by the concept drift phenomenon,which makes accurate modeling in complex industrial environ-ments a difficult and challenging task.To this end,we propose a domain adaptation physics-informed neural net-work method based on linear dynamical operator.A linear dynamical operator neural network model is built from the historical data,capturing dynamic properties of multivariate time series data.Then the mechanism model is dis-cretized by the forward Euler method to construct the physical information regularization term,so that the model obeys the constraints of the mechanism.Finally,the distribution of hidden layer state variables under historical and current operating conditions is aligned by the maximum mean discrepancy,and the domain adaptation loss is con-structed to reduce the impact of data distribution changes on the model under variable operating conditions.Exper-iments on several datasets show that the proposed method can effectively improve the model prediction accuracy and generalization performance.关键词
物理信息机器学习/概念漂移/域适应/线性动力算子神经网络Key words
Physics-informed machine learning/concept drift/domain adaptation/linear dynamical operator neur-al network引用本文复制引用
曹力丰,阎高伟,肖舒怡,董珍柱,董平..基于域适应物理信息神经网络的时间序列预测方法[J].自动化学报,2025,51(6):1329-1346,18.基金项目
国家自然科学基金(61973226),山西省科技重大专项(202201090301013),山西省自然科学青年基金(202203021222101),格盟集团科技创新基金(2023-05),山西省研究生科研创新项目(2024KY268)资助Supported by National Natural Science Foundation of China(61973226),Shanxi Province Major Special Program of Science and Technology(202201090301013),Shanxi Province Science Foundation for Youths(202203021222101),Gemeng Group Tech-nology Innovation Fund Project(2023-05),and Shanxi Province Graduate Research Innovation Project(2024KY268) （61973226）
