Prediction and optimization of the efficiency and energy consumption of an ammonia vacuum thermal stripping process using experiments and machine learning models

In this study, we adjusted the effluent temperature and pH to optimize the ammonia stripping efficiency and energy consumption using quadratic equation (QE) and machine learning models, including multi-layer perceptron (MLP), random forest (RF), and extreme gradient boosting (XGBoost) models. The experimental results for the one-factor-at-a-time method revealed that the stripping efficiency increased with both temperature and pH. However, the energy consumption required to reduce the ammonia concentration by an order of magnitude did not show the same with the experimental result of stripping efficiency; rather, it was lowest when the temperature was 40oC and the pH was 11.5. The response surface for the ammonia stripping efficiency predicted using the QE and machine learning models exhibited a similar trend to the experimental results. Analysis of variance for the QE model revealed that pH, temperature, and reaction time were important factors determining the stripping efficiency. The feature importance analysis revealed that temperature and pH made similar contributions. The RF and XGBoost models also produced relatively reliable results (R2 > 0.98). The validation of RF and XGBoost models using the additional data from optimal conditions (treatment time = 78.456min at pH = 11.079 and temperature = 37.632 °C for RF and treatment time = 62.499min at pH = 11.079 and temperature = 42.895 °C for XGBoost) proved the reliability of both models (observed treatment times were 80.316 and 65.210min, respectively). This study offers implications for designing effective and energy-efficient systems for ammonia removal from anaerobic digestion effluent.