Resolving Steam Turbine Casing Thermal Management Challenges with a Dual Attentive Bi-GRU Soft Sensor for Transient Operation

This study introduces a novel dual-model deep learning framework based on Bidirectional Gated Recurrent Units (Bi-GRUs) with the Attention Mechanism to predict intermediate-pressure (IP) turbine casing temperatures in a 370 MW coal-fired power plant under varying operational regimes, including startup, shutdown, and load-following conditions. Accurate temperature prediction is critical, as thermal gradients induce significant stresses in the turbine casing, potentially causing fatigue crack initiation. To mitigate sensor failures, which lead to costly downtime in power generation systems, the proposed soft sensor leverages an extensive dataset collected over one year from Unit 4 of the Opole Power Plant. The dataset is partitioned into shutdown and active regimes to capture distinct thermal dynamics, enhancing model adaptability. The framework employs advanced preprocessing techniques and state detection heuristics to improve prediction robustness. Experimental results show that the dual-model approach outperforms traditional machine learning models (Random Forest Regressor, XGBoost) and single-model deep learning baselines (LSTM, Single Attentive Bi-GRU), achieving a mean squared error (MSE) of 2.97 °C and a mean absolute error (MAE) of 1.07 °C on the test set, while also maintaining low prediction latency suitable for real-time applications. This superior performance stems from a tailored architecture, optimized via Hyperband tuning and a strategic focus on distinct operational regimes. This work advances soft sensing in power systems and provides a practical, real-time solution for stress monitoring and control, particularly as coal plants in Poland face increased cycling demands due to the growth of renewable energy sources, rising from 7% in 2010 to 25% by 2025. The approach holds potential for broader application in industrial settings requiring robust temperature prediction under variable conditions.