A data-driven method for dynamic resilience prediction of chemical process systems integrating dynamic simulation and attention Bi-LSTM

Abstract

Chemical process systems are susceptible to disruptions that can affect production efficiency and even turn into accident with casualties. To reduce the impact of disruptions on chemical process systems, this study presents a data-driven methodology for evaluating and forecasting system resilience. Critical process parameters (CPPs) governing system performance are first identified, considering both safety constraints and production efficiency. Subsequently, potential disruptive events are systematically identified and generated. Then, dynamic simulations are performed to generate transient response data. Finally, a resilience assessment method considering the CPPs is proposed, and an attention Bi-LSTM model is developed to learn the temporal dependencies between disruptions characteristics and system resilience, thereby allowing the prediction of the system resilience. The results indicate that the proposed methodology can accurately predict the variation over time of system performance and resilience under disruptive scenarios, with a root mean square error (RMSE) of 0.0035 and a mean absolute error (MAE) of 0.0032. The error between the predicted value and the actual value of resilience does not exceed 0.005. The proposed methodology demonstrates robust capability in predicting resilience of chemical process systems under diverse disturbance scenarios. It provides a systematic approach for proactive resilience management in chemical process systems, with implications for enhancing operational safety and informing maintenance strategies.