Unified PF-LSTM Improves Data-Driven Process Simulation Accuracy.

Data-driven process simulation aims to create realistic case trajectories from historical event logs without needing an explicit model of the underlying dynamics. While deep sequence models can capture temporal dependencies, event logs often provide only a partial view of the process state, such as activity completions without start times. This means a single observed history could align with multiple latent process conditions, a challenge standard recurrent models struggle with. This research introduces the Unified Particle Filter LSTM (Unified PF-LSTM) to address this limitation. The PF-LSTM maintains and sequentially updates a weighted set of recurrent-state hypotheses, effectively capturing the multiple plausible latent conditions. It summarizes this "particle belief" using a weighted mean and learned features, which are then used to predict the next activity and the current activity's sojourn time. Trained end-to-end on real-world emergency department datasets, the Unified PF-LSTM consistently outperforms existing data-driven baselines. It shows particularly strong gains in accurately reproducing process routing, activity durations, and overall system-level behavior, especially in scenarios where event logs offer only incomplete process state information.