نوع مقاله : مقاله پژوهشی
عنوان مقاله English
نویسندگان English
Introduction: The increasing contamination of water resources has intensified the need for efficient approaches to evaluate and optimize pollutant adsorption in wastewater treatment. Biochar is a promising adsorbent, but experimental determination of equilibrium adsorption capacity (Q_e) is time-consuming and costly. This study therefore develops a data-driven artificial intelligence model, framed as a laboratory-data-based soft-sensor prototype, for predicting Q_e in biochar-based adsorption systems.
Materials and Methods: Eight predictive models—five machine learning algorithms (Random Forest, Extra Trees, XGBoost, Gradient Boosting, and Multi-Layer Perceptron) and three deep learning architectures (1D-CNN, EnhancedResNet, and FT-Transformer)—were trained and evaluated using the qe_biochar_ec dataset containing 3,757 experimental records. Inputs include operating conditions and physicochemical descriptors such as solution pH, temperature, contact time, initial pollutant concentration, adsorbent dosage, specific surface area, pore characteristics, elemental composition, feedstock type, and pollutant type. Performance was evaluated using R^2, RMSE, and MAE, and SHAP was used for model interpretation. A FastAPI interface and Streamlit dashboard were also developed as a prototype deployment layer; direct connection to online plant sensors was not evaluated in this study.
Results: The results demonstrated that the Extra Trees algorithm achieved the best predictive performance, with an R^2 value of 0.8899, an RMSE of 20.49 mg g^(-1), and an MAE of 7.61 mg g^(-1). Furthermore, FT-Transformer achieved the highest performance among the deep learning models, with an R^2 of 0.8587. SHAP analysis enabled the identification of the most influential factors affecting adsorption capacity and provided greater insight into the model's predictions.
Conclusion: The results show that AI models can predict equilibrium adsorption capacity with useful accuracy and may reduce reliance on repeated laboratory equilibrium tests. The present system should be interpreted as a laboratory-data-based predictive soft-sensor prototype rather than a validated online industrial sensor. Its use for real-time monitoring, process control, or digital-twin applications requires prospective validation with time-resolved pilot- or full-scale plant data and integration with online measurements and plant information systems.
کلیدواژهها English