Accurate prediction of key state variables in complex industrial processes is challenged by complicated inter-variable couplings, operating-condition fluctuations, and pronounced measurement noise. To capture such correlations, traditional deep time-series models often suffer from considerable computational burden; meanwhile, their multi-step forecasting performance commonly degrades as the prediction horizon increases, making it difficult to balance accuracy and efficiency. To address these issues, this paper proposes a factor-aware inverted Transformer for time-series forecasting (FA-iTransformer). First, the proposed model reformulates the embedding scheme for multivariate time series via variable-wise serialization, where the historical trajectory of each variable within a time window is encoded as a basic unit, strengthening the representation of variable dynamics and facilitating the extraction of multi-scale patterns such as trend variations and abrupt disturbances. Second, to reduce the cost of cross-variable interaction modeling, a low-rank attention mechanism based on latent factors is introduced. Through a three-stage “aggregate–interact–distribute” information propagation in the factor space, the model captures shared patterns while preserving variable-specific characteristics, improving computational efficiency and robustness without sacrificing global correlation modeling capability. Finally, a non-autoregressive temporal projection module is employed to enable parallel long-horizon multi-step prediction, mitigating bias propagation in rolling forecasting and reducing inference latency. Experiments on an industrial blast-furnace gas dataset demonstrate that the proposed method consistently outperforms mainstream time-series forecasting models in both prediction accuracy and computational efficiency, and achieves higher efficiency in terms of model size and inference time, providing effective predictive support for online monitoring and scheduling of complex industrial systems.