A causal representation learning based model for time series prediction under external interference

With the development of artificial neural networks, time-series prediction techniques are becoming more mature. While the prediction of time-series data under external interference remains a challenge. The different distribution of data before and after the external interference and the out-of-distribution of different datasets will lead to poor prediction accuracy, robustness, and generalization ability of the general prediction models. Thus, firstly, we propose a conditional causal representation (CCR) model based on causal representation learning, which solves these problems by extracting interference-related causal representations and learning causal mechanisms subject to external interference. Secondly, we demonstrate that the interferencerelated causal representations should satisfy three properties: independence of non-causal factors and other causal representations, mutually independent dimensions, and causality sufficient for the prediction. Based on these properties, the causal representation abstraction component is designed to extract interference-related causal representations. Thirdly, we demonstrate the conditional structure is equivalent to the causal mechanism if the conditions of the conditional structure are interference-related causal representations, and based on this, the conditional causal network (CCN) component is designed to learn the causal mechanisms subject to external interference. The experimental results show that the CCR has good prediction accuracy, robustness, and generalization ability.