An Engineering Neuro-Fuzzy Decision Support System for Optimizing an Industrial Sulfur Production Process Under Uncertainty

Sulfur recovery units (SRUs) at oil refineries are complex, multi-parameter systems whose Cold Bed Absorption (CBA) reactors operate under incomplete and uncertain process information, which limits conventional control. This study develops an engineering neuro-fuzzy decision support system (DSS) for optimizing sulfur production at the Atyrau oil refinery (Kazakhstan) under uncertainty. The system integrates a mathematical description of the process, whose dynamic part is stated as a general formulation, an expert-derived fuzzy rule base, a first-order Sugeno neuro-fuzzy estimator of sulfur yield and quality, and a multi-criteria fuzzy optimization block based on the Bellman–Zadeh approach, together with an explicit operator-facing decision algorithm and a software interface. The estimators were identified and validated on 200 operating log records of the unit. These records show that the effect of reactor temperature on the sulfur quality index reverses sign with reactor pressure—the correlation is −0.57 below 0.79 kgf/cm2 and +0.75 above it—so that a model additive in the two variables cannot represent the relationship: linear regression on temperature and pressure attains a cross-validated R2 of only 0.06 for the quality index, whereas the neuro-fuzzy estimator attains 0.83 (RMSE 0.72 points, MAPE 0.74%). A regression with an explicit interaction term on the same inputs attains 0.79; in a paired comparison the neuro-fuzzy estimator had the lower error in every repetition, but the difference is not statistically significant (p = 0.07), and the estimator is retained for its rule-level interpretability and its integration with the decision block. For sulfur yield, where the relationship is close to linear once pressure and throughput are accounted for, the corresponding figures are 0.71 and 0.85 (RMSE 0.33 t/day, MAPE 1.35%). The same ordering is reproduced on an independent hold-out test set, under nested and cluster-blocked cross-validation and after removal of the records lying at the unit’s capacity and regulatory bounds, and the expert rule base was revised into an asymmetric form accordingly. A worked application of the optimization block to three operating situations, with the production plan, quality and capacity constraints enforced, raised the model-predicted degree of joint satisfaction of the yield, quality and energy criteria from 0.08–0.25 to 0.49–0.57; in one of the three, the recommended action is to reduce process severity and lower a reheat duty surrogate by about a third rather than to increase output. The resulting DSS, at present a standalone advisory software prototype, supports the operator in analyzing the current process state, evaluating alternative operating modes and forming recommendations, while the final decision remains with the operator.