Abstract:

To improve the level of process safety management in refining and chemical enterprises and enhance dynamic monitoring and trend warning capabilities for key safety indicators, a multi-model fusion method for process alarm data prediction was proposed. This approach integrated three time series forecasting models: Autoregressive Integrated Moving Average (ARIMA), double exponential smoothing, and particle swarm optimization(PSO)-based support vector regression (SVR). The method effectively modeled and predicted diverse safety indicators by addressing their trend, autocorrelation, and nonlinear characteristics. Initially, outliers in the raw indicator data were processed. Three types of forecasting models were then constructed and their prediction results were computed. The optimal model for trend prediction was automatically selected based on error comparison. Finally, an empirical analysis was conducted using the time-averaged alarm count indicators from a refinery enterprise over one year. The results show that the proposed method dynamically adapts to varying data characteristics, with selected model prediction errors consistently remaining below 0.1, significantly outperforming the existing requirements for alarm magnitude accuracy. This method effectively enhances the accuracy and flexibility of safety indicator prediction in refining and chemical enterprises and enables the timely identification of potential risk indicators.

Key words: refining and chemical enterprises, process safety, safety management, indicator data, time series, support vector regression (SVR), exponential smoothing method