Real-time Monitoring of Petroleum Hydrocarbons in Groundwater using Hybrid Machine Learning Architectures

Abstract. Monitoring petroleum hydrocarbon (PHC) plumes in groundwater is essential for managing oil contamination but is often hindered by high costs. We evaluated machine learning (ML) frameworks that estimate concentrations of benzene, ethylbenzene, and xylenes (BEX), using affordable, in situ water quality parameters (iWQPs) as inputs: pH, dissolved oxygen, electrical conductivity, and oxidation-reduction potential. Due to a scarcity of field data, we trained and tested models on high-resolution virtual data generated by a reactive transport model. We compared a long short-term memory (LSTM) network against classical algorithms (multiple linear regression, random forest, support vector regression, XGBoost) and an LSTM-XGBoost hybrid. Model performance depended on the underlying geochemical relationship between iWQPs and BEX. Accurate predictions (R² ≥ 0.80, MAPE < 2.3 %) were achieved when iWQPs were strongly correlated with BEX degradation (e.g., as a primary electron donor); the LSTM model yielded predictions within a 5 % error margin for 70 % of the test cases. Performance declined sharply (R² < 0) during periods where iWQPs were correlated with non-volatile dissolved organic carbon, another component of dissolved PHC. Incorporating hydraulic head data improved accuracy by informing the model of groundwater flow dynamics. While the LSTM model struggled to extrapolate beyond its training data (e.g., during extreme flow events), it reliably detected the direction of concentration trends, providing a valuable trigger for adaptive monitoring. We also demonstrated how a hybrid Kalman filter could successfully capture concentration trends after source removal through recursive updating. Our proposed ML framework provides BEX level estimation for improved groundwater monitoring.