A hybrid risk prediction method for chemical process systems based on enhanced feature engineering and XGBoost

Abstract

Deviations of process parameters from their normal ranges are the primary causes of accidents in chemical process systems. Traditional risk assessment methods largely rely on static probability analysis based on historical data, which struggles to capture the dynamic influence of real-time parameter variations on risk and lacks the capability to predict risk evolution trends. This paper proposes a dynamic risk prediction method for chemical process systems based on enhanced feature engineering and XGBoost. First, key process parameters (KPPs), including temperature, liquid level and flow rate are identified through process analysis. A comprehensive risk indicator is then constructed using Dempster-Shafer (D-S) evidence theory to achieve dynamic quantification of system risk. Second, a dynamic simulation model is established using Aspen Plus, simulating operations under normal, disturbed, and extreme conditions to generate time-series data of KPPs. On this basis, multi-scale sliding window techniques are employed to extract enhanced features, including temporal, statistical, trend, and disturbance features. Finally, an XGBoost-based risk prediction model is developed. A continuous stirred-tank reactor (CSTR) is employed to demonstrate the proposed methodology. The results indicate that the proposed methodology achieves an RMSE of 0.159, MAE of 0.122 and an R² of 0.7237, outperforming traditional methods by significant margins. The results validate the effectiveness of combining enhanced feature engineering with XGBoost for risk prediction in chemical processes.