Scholarlink

<p>This paper presents a new approach to tuning the cost function weights of a nonlinear model predictive controller (NMPC) using offline Bayesian optimization (BO). We propose a recursive weight selection method that integrates BO directly into the NMPC simulation loop and targets an economic cost function. This approach identifies weights that optimize an economic cost function, ensuring that controller performance is aligned with the operational economics of the process. A case study involving an interconnected tank system with nonlinear level and temperature dynamics illustrates the method. Two operating conditions are tested: an undisturbed scenario and a disturbed one, incorporating sensor noise and model–plant mismatch. In both scenarios, the BO-tuned NMPC outperforms Traditional and Satisficing-based weight strategies, yielding smoother control actions. In the disturbed case, cost reductions of up to 4.5% are achieved. The results confirm that offline BO-based tuning offers a viable and robust alternative to manual or online adaptive strategies, especially in systems where online retraining is impractical. Sensitivity tests further highlight the importance of informed search interval selection to ensure convergence and performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Tuning nonlinear model predictive control via Bayesian optimization: a comparative performance analysis

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Maria Alice de F. Marques,
  • Juarez dos Santos Azevedo,
  • Julio Elias Normey-Rico,
  • Marcus V. Americano da Costa

摘要

This paper presents a new approach to tuning the cost function weights of a nonlinear model predictive controller (NMPC) using offline Bayesian optimization (BO). We propose a recursive weight selection method that integrates BO directly into the NMPC simulation loop and targets an economic cost function. This approach identifies weights that optimize an economic cost function, ensuring that controller performance is aligned with the operational economics of the process. A case study involving an interconnected tank system with nonlinear level and temperature dynamics illustrates the method. Two operating conditions are tested: an undisturbed scenario and a disturbed one, incorporating sensor noise and model–plant mismatch. In both scenarios, the BO-tuned NMPC outperforms Traditional and Satisficing-based weight strategies, yielding smoother control actions. In the disturbed case, cost reductions of up to 4.5% are achieved. The results confirm that offline BO-based tuning offers a viable and robust alternative to manual or online adaptive strategies, especially in systems where online retraining is impractical. Sensitivity tests further highlight the importance of informed search interval selection to ensure convergence and performance.