Adaptive Soft Sensor Modeling for Coal-Fired Power Plants Based on a Tracking Structured State-Space Model

Accurate and continuous measurement of key process variables in modern coal-fired power plants is critical, yet challenging due to ”concept drift” induced by flexible operating regimes. Frequent start-ups, shutdowns, and rapid load changes introduce complex, variable, and unknown operating conditions, which degrade the performance and adaptability of conventional soft-sensing instruments. To address this measurement problem, this paper proposes a novel soft-sensing model with high accuracy and robust tracking capabilities: the Tracking Structured State Space Sequence Model (TS4). The TS4 architecture is composed of three key innovations for superior measurement performance. First, a Probabilistic Structured State Space Sequence (PS4) model is developed to establish a highly accurate baseline measurement under stable conditions and provide probabilistic uncertainty quantification. Second, a lightweight Probabilistic Kernel Recursive Least Squares (PKRLS) algorithm with a forgetting factor is derived, enabling the soft sensor to rapidly adapt to new operating conditions through online updates. Finally, a dynamic posterior weight updating strategy is designed to intelligently fuse the deep representation of PS4 with the online tracking ability of PKRLS, creating a single, robust measurement model. The proposed soft sensor was validated on operational data from a large-scale coal-fired power plant for the task of power generation measurement. Compared to various baseline models, the TS4-based instrument demonstrates significantly improved measurement accuracy and superior capability in tracking dynamic operating conditions, proving its value for enabling intelligent, safe, and efficient plant operations.