| Abstract: |
Predicting process variables in industrial plants is often approached through either physics-based or data-driven models, yet their relative effectiveness strongly depends on the information content of the available measurements. This paper presents a persistence-aware benchmark of physics-based, data-driven, and hybrid predictors for washing-water pH estimation in a full-scale geothermal steam-washing plant using a 19-month industrial dataset comprising 21 asynchronous sensor signals. The benchmark compares a grey-box model based on acid-base recombination with online parameter identification, long short-term memory (LSTM) networks, gradient-boosted regression trees, and the persistence baseline. Results show that, at prediction horizons of 50 and 120 minutes, none of the considered predictors outperforms persistence. Although the grey-box model closely tracks the measured pH and the LSTM accurately reproduces the training dynamics, both fail to provide additional predictive information on real plant data. A combination of simulation-based verification, lagged-correlation analysis, multivariate predictability tests, and parameter identifiability analysis demonstrates that this limitation originates from the dataset rather than from the prediction models. The study further shows that the physics-based model becomes advantageous at multi-hour prediction horizons, where its error remains nearly constant while persistence progressively degrades, yielding up to a 43% reduction in prediction error after 24 hours. |