Corrosion problems have threatened long-term safe and stable operation of refining units. At present, refining enterprises mainly use corrosion monitoring and detection to identify equipment corrosion states, which has the shortcomings of narrow identification scope and high cost. The data-driven method avoids these problems, and has the advantage of efficiently predicting corrosion states to support corrosion management decisions. This paper, based on multi-source data, proposes a method focusing on the prediction about key corrosion parameters and establishes prediction models on critical parts of refining unit. Firstly, the application of demand-oriented corrosion prediction method is proposed. Then, according to the process operation parameters and medium analysis data of atmospheric tower overhead circuit, the regression prediction models based on RF are established. Meanwhile, after outlier detection by iForest, the model's parameters are optimized by SOS. In the limited real data, the optimized model achieves the best prediction with RMSE of 0.00611, MAE of 0.00513, and R² of 0.918, and realizes the mining of corrosion parameter sensitivity. Finally, a variety of models are compared. The prediction method proposes in this paper have generalization performance, which can serve as an instruction for equipment safety management and hidden dangers identification.