Research on fast load ramping methods for thermal power units based on kernel bandwidth adaptive particle filter-MPC

High renewable penetration requires coal-fired units to provide faster and more reliable load-ramping support, while time-varying uncertainties such as coal-quality fluctuation and equipment aging aggravate model-plant mismatch and degrade MPC tracking performance. This paper proposes a kernel-bandwidth adaptive particle filter-model predictive control (KBAPF-MPC) strategy for rapid load-ramping control of thermal power units. The residual prediction error is formulated as an output-equivalent mismatch and estimated online by a particle filter with kernel-bandwidth adaptive diffusion. The diffusion bandwidth is updated from the empirical particle distribution using Silverman’s rule and the effective sample size, providing an adaptive mismatch-estimation mechanism for rapid load-ramping operation. A conditional local practical-stability analysis is provided under bounded mismatch-estimation errors, bounded reference variations, local recursive feasibility, and a practical cost-decrease condition. The analysis characterizes a residual tracking region and establishes a time-average tracking-error bound in terms of the mismatch-estimation-error and reference-variation bounds. The estimated mismatch is embedded into the MPC prediction as a feedforward compensation term while retaining the original optimizer and physical constraints. Simulations on a 350 MW CFB unit with composite parameter drift show that KBAPF-MPC reduces RMSE by 60.7%, 61.9%, and 75.5% compared with conventional MPC, and by 30.7%, 41.5%, and 49.3% compared with UKF-MPC for power output, main steam pressure, and temperature, respectively, while maintaining constraint-compliant actuator responses and real-time feasibility under the considered scenario.