Leveraging machine learning to enable real-time phosphorus quantification and process control in wastewater treatment facilities using semi-specific multi-sensor arrays

Abstract

Phosphorus (P) is an essential element for all life forms, meaning that P is both an important ingredient in fertilizer, which is therefore continuously mined for this purpose, and can be a driver of nutrient pollutant (eutrophication) in surface waters. As such P removal is a key goal of municipal wastewater treatment to ensure protection of receiving waters. Enhanced biological phosphorus removal (EBPR) is an activated sludge process that uses alternating anaerobic and aerobic conditions to stimulate polyphosphate accumulating organisms (PAOs) to take up polyphosphate from the bulk liquid, therein treating wastewaters. EBPR can replace costly and unsustainble chemical-dosing approaches to treatment, which both use metal supplies and generate hazardous solid waste, but also raises potential for resource recovery and reuse (i.e., P can then be recovered from PAO cells in a separate recovery process). However, oxygenation (which can account for 50% of treatment plant costs) is required to promote PAO uptake of P, and therefore energy- (and therefore cost-) efficient operations of these processes requires an ability to track P removal in real time - i.e., to prevent the costs associated with over-treatment. A major challenge in achieving this is the lack of commercial options for online real-time measurement of P. This work proposes an alternative (soft sensor'') approach wherein the data from an array of commercially available sensors for other chemical parameters that provide proxy (either direct or indirect) measurements are fused using a tuned algorithm to estimate P behavior in EBPR processes. Critically, this innovation takes advantage of a known aspect of the PAO metabolism in which P uptake is actually related to poly-P synthesis inside the cell, requiring simultaneous uptake of Mg2+, K+, and PO4{3-}. The existence of commercial sensors for these co-transported ions, and for other bulk chemical properties, suggests a feasible pathway forward. Therefore, this work leverages data from comprehensive lab experiments and custom-designed EBPR-informed machine learning (ML) approaches to achieve two key goals: (1) design of a soft sensor for real-time detection of the timepoint at which the P removal process is complete in an operating EBPR process and (2) design of a soft sensor to provide real-time quantification of phosphorous concentration in an EBPR process. Data collected from a highly-instrumented lab scale reactor (key sensors: conductivity and ion selective electrodes for K+, Ca2+, hardness, Na+, Cl-, NH4+) operated undernormal'' and ``exceptional'' (altered influent chemistry representing real-world variability that would challenge sensor accuracy) conditions are used to evaluate which and how many sensors are required to achieve each of these goals and to assess which machine learning algorithms are most appropriate for the purpose. Classification approaches (support vector machines, nonlinear logistic regression, random forest, and Bayesian classification) were evaluated for process detection, and regression approaches (multiple linear regression, partial least square, support vector regression, random forest, and artificial neural network) were evaluated for real-time quantification of phosphorous. For process detection two key results emerge: (1) the optimal soft sensor design, which achieves 100% accuracy, contains only a single sensors (K+ ISE) as long as the monitored parameter is sensor signal slope (rather than absolute output), regardless of the algorithm used for data fusion and (2) this design can be resilient even in the face of variable inflow conditions, which are common in municipal wastewaters. This highlights the importance of predictor data format and selection over than the ML approach used. For the more challenging quantification problem, the soft sensor with 7-sensor suite fused using artificial neural networks had best performance (RMSE of ~0.14 log[M] for test data, RMSE of ~0.5 log[M] for robustness analysis), while a 5-sensor version (NH4+, K+, Ca2+, Cl-, and hardness ISEs) fused using random forest performed almost as well (RMSE ~0.25 log[M] for test and 0.46 log[M] for robustness data). The developed soft sensor for real-time online estimation of P concentration has a detection limit of approximately 0.2 mg/L (6.3× 10-6 mol/L) which provides a measure sufficient to serve the needs of operating wastewater treatment facilities at approximately 10x lower cost than current technologies.--Author's abstract