Closed-loop Identification of a MSW Grate Incinerator using Bayesian Optimization for Selecting Model Inputs and Structure

TL;DR

This work addresses the challenge of obtaining low-order closed-loop models for MSW grate incinerators amid strong disturbances from fluctuating fuel composition. It introduces SysID-HP, a methodology that combines cross-validation, L2 regularization, and Bayesian optimization to sequentially select model inputs and structure, yielding robust, interpretable transfer-function models validated on industrial plant data. Two models are developed: a Basic Model focusing on steam-load dynamics and a Comprehensive Model that adds subprocess-level insights through algebraic relations and an auxiliary input for steam generation. The results show high predictive accuracy (R^2 ≈ 0.92) with meaningful confidence intervals, and the study demonstrates the potential of these low-order closed-loop models for model-based superordinate control and broader application to disturbed, over-actuated systems.

Abstract

The creation of low-order dynamic models for complex industrial systems is complicated by disturbances and limited sensor accuracy. This work presents a system identification procedure that uses machine learning methods and process knowledge to robustly identify a low-order closed-loop model of a municipal solid waste (MSW) grate incineration plant. These types of plants are known for their strong disturbances coming from fuel composition fluctuations. Using Bayesian optimization, the algorithm ranks and selects inputs from the available sensor data and chooses the model structure. This results in accurate models with low complexity while avoiding overfitting. The method is applied and validated using data of an industrial MSW incineration plant. The obtained models give excellent predictions and confidence intervals for the steam capacity and intermediate quantities such as supply air flow and flue gas temperature. The identified continuous-time models are fully given, and their step-response dynamics are discussed. The models can be used to develop model-based unit control schemes for grate incineration plants. The presented method shows great potential for the identification of over-actuated systems or disturbed systems with many sensors.

Figures (7)

• Figure 1: Grate incineration setup and control, main components and sensor measurements.
• Figure 2: Overview of the system identification procedure with cross-validation and Bayesian optimization for hyperparameter tuning.
• Figure 3: Validation experiment $1/7$. Top: dimensionless steam load setpoint. Bottom: measured steam load and Basic Model output with 95% confidence interval.
• Figure 4: Schematic structure of subprocesses and measurement signals in a waste incineration plant.
• Figure 5: Validation results comparing the dimensionless Comprehensive Subprocess Model outputs and confidence intervals with corresponding measurements (validation experiment $1/7$).