In the prediction of key state variables for complex industrial processes, multivariable couplings and operating-condition fluctuations lead to complicated correlation structures. Existing time-series forecasting models, when modeling such correlations, not only incur high computational cost but also tend to suffer from the cascade amplification of prediction errors as the forecasting horizon increases, making it difficult to achieve a balance between prediction accuracy and computational efficiency. To address these issues, this paper proposes a factor-aware inverted Transformer time-series forecasting model. Different from attention-based architectures that model multivariable signals jointly, the key innovation of the proposed method lies in reconstructing the orthogonal representation of multivariable data at the foundation data level and realizing cross-variable decoupling in the latent space. Specifically, the model adopts a variable-wise serialization scheme, where the historical sequence of each variable within a time window is encoded as a basic unit, thereby enhancing the dynamic representation of individual variables. To reduce the computational cost of cross-variable interaction modeling, a latent-factor-based low-rank attention mechanism is further introduced, which improves computational efficiency and robustness while preserving the capability of global correlation modeling. Finally, a non-autoregressive temporal projection module is integrated to enable parallel long-horizon multi-step forecasting, reducing bias propagation in multi-step forecasting. Experimental results demonstrate that the proposed method outperforms existing time-series forecasting models in both prediction accuracy and computational efficiency, providing efficient predictive support for the online monitoring and scheduling of complex industrial systems.