Reliable Backtesting for Continuous Industrial Optimization via Sequential Simulation: A Gas-Fired Ceramic Kiln Case Study

Abstract

Offline evaluation of optimization policies for continuous industrial processes becomes unreliable when each historical timestamp is optimized independently, because a counterfactual control action should modify the plant state observed at later times. This paper studies this issue in a gas-fired mono-layer ceramic kiln, where turbine frequency is adjusted to preserve a target pressure profile while reducing gas consumption. We propose a finite-horizon sequential simulator that propagates a compact process state through auxiliary transition models, reconstructs future optimizer inputs, and enables temporally coherent closed-loop backtesting. The framework combines objective models for gas and pressure compliance with a two-stage constrained optimizer that respects industrial actuation limits. Using the proposed sequential simulation, the optimized policy reached baseline-referenced improvements, computed against a predefined internal improvement threshold, of 20.61% in total gas consumption, 15.17% in pressure-profile compliance, and 38.70% in optimal minutes on the continuity-constrained subset.

Silvia Ventura Cabrejas

Artificial Intelligence Phd Student

PhD student at Universidad Rey Juan Carlos