| In refinery operations, compressors ensure process continuity and a reliable hydrogen supply. Therefore, early fault detection is essential for operational safety and economic performance. This study presents a Digital Twin (DT) framework at the equipment level for short-term forecasting and residual-based anomaly detection. This framework is applied to a reciprocating net gas compressor. We employ recurrent neural network (RNN) architectures based on long short-term memory (LSTM) and gated recurrent unit (GRU) layers to model the compressor's normal behavior using multivariate process time-series data collected at a 15-minute sampling interval. Model perfor-mance is evaluated using mean absolute error (MAE) and root mean square error (RMSE). The GRU-128 architecture is identified as a suitable configu-ration. Anomaly detection is performed through residual analysis, where de-viations between actual measurements and model predictions are monitored using exponentially weighted moving average (EWMA) control chart by considering control limits. Beyond offline validation, the proposed approach is deployed in a real-time operational environment using the Dataiku plat-form. Residual-based indicators are continuously visualized on an interactive dashboard, enabling automatic alert generation when control limits are ex-ceeded. Detected anomalies are validated simultaneously against operational records, demonstrating temporal consistency between model alerts and actual process deviations. These results confirm that DT framework ensures coher-ence between offline analytical studies and real-time deployment. This framework provides reliable early-warning capability and practical decision support for predictive maintenance in industrial compressor systems. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.