<p>In real-world high-dimensional systems, both causal dependencies and temporal information of key variables are essential for dissecting the underlying mechanisms governing system dynamics. However, effective approaches to synthesize these two interconnected aspects for deeper insights remain lacking. Here we show a neural network framework, the Causal-oriented Representation Learning Predictor (CReP), which jointly conducts causal analysis and multistep forecasting from a unified perspective. CReP implicitly learns latent causal representations from observed data while simultaneously making multistep predictions, and explicitly interprets the representations to uncover the causes and effects of target variable. The core idea of CReP is to decompose the original space into three orthogonal latent factors, each capturing distinct causal representations: cause-related, effect-related, and non-causal representations of the target variable. The reconstruction-based dynamic causation, generalized through spatiotemporal information (STI) transformation mechanism, provides a theoretical foundation for simultaneously modeling causal interactions via latent representations and predicting future states using the effect representation. Evaluations on three simulation models and two real-world datasets demonstrate CReP’s robust forecasting accuracy and reliable causal insights. As a self-supervised-learning approach, CReP shows significant potential for practical applications and provides a unified framework to reveal intrinsic mechanisms in dynamical systems by integrating causal relationships and temporal information.</p>