Multi-step prediction of main furnace temperature in Alumina roasting based on time-frequency dual-domain fusion large foundation model

In the gas suspension roasting process of alumina, the main furnace temperature is the core variable reflecting the operating conditions. Accurate multi-step ahead prediction of this variable can provide a prerequisite for the implementation of fault warning and model predictive control, thereby stabilizing the phase transformation rate of products and reducing production risks. However, the inherent characteristics of the roasting process, including strong multivariable coupling and nonlinear dynamic evolution, increase the difficulty of prediction. Existing data-driven models are prone to autoregressive error accumulation and often lose high-frequency dynamic details, resulting in overly smooth prediction curves that fail to provide effective guidance for actual production. To address the above issues, this paper proposes a data-driven multi-step prediction method based on the Time-Frequency dual-domain Large Foundation Model (TF-LFM). With feature decoupling and time-frequency dual-domain optimization as the core, this method uses the Seasonal-Trend Decomposition Procedure Based on Loess (STL) to decouple complex process time-series signals into trend components and periodic components. Subsequently, the model leverages a large language model to extract long-range trends for temporal reasoning, and introduces a large speech model to reconstruct the frequency-domain features of high-frequency periodic fluctuations, so as to compensate for the defect of dynamic detail loss in traditional models. Under the constraint of the time-frequency joint loss function, the dynamic synergy and restoration of the dual-domain prediction results are completed. The verification results based on actual production data from a large aluminum plant show that the proposed method overcomes the problems of error accumulation and dynamic detail loss. Compared with the 3 time-series models for comparison, the prediction error of the proposed method in 32-step long horizon prediction is reduced by about 60.5% compared with the optimal comparison model, which provides reliable model support for fault warning and refined control of the roasting process.