Predicting Operational Expenditure of Wastewater Treatment Plants Using Machine Learning Techniques

Accurate forecasting of operational expenditure (OPEX) in wastewater treatment plants is critical for financial planning and utility management, yet it remains challenging due to complex, nonlinear interactions between plant capacity, influent characteristics, and fluctuating market prices. This research develops an interpretable predictive framework for OPEX estimation by systematically evaluating multiple advanced machine learning algorithms on a dataset of 2847 operational records (1139 empirical and 1708 synthetic). We compared the predictive performance of decision tree, AdaBoost, random forest, 1D-CNN (serving as an architectural negative control), support vector regression (SVR), multilayer perceptron artificial neural networks (MLP-ANN), and ensemble approaches using a multidimensional dataset comprising influent parameters, operational factors, and economic variables. The predictive models were strictly evaluated using a 90% training and 10% testing data split, prioritizing robust metrics including R², root mean square error (RMSE), and mean absolute error (MAE). Results demonstrated that the MLP-ANN significantly outperformed all other models, achieving an outstanding testing R² of 0.993. In contrast, traditional tree-based ensembles exhibited varying degrees of overfitting, and the 1D-CNN performed poorly on this tabular dataset (R² = 0.568). To resolve the black-box nature of the winning neural network, SHapley Additive exPlanations (SHAP) was applied, revealing that volumetric capacity, energy unit cost, and sludge disposal cost are the dominant drivers of OPEX, reflecting known physical and economic principles. Importantly, this model is evaluated exclusively on historical data and currently lacks validation on external datasets. While the integration of SHAP ensures mathematical interpretability and validates economies of scale, future research must incorporate independent external validation to confirm the framework's broad regional transferability.