Harnessing Machine Learning for Enhanced A-Stage Process Efficiency: A Pioneering Full-Scale Study

Introduction The shift towards energy- and resource-efficient WWTPs is no longer a privilege. For this reason, the A/B scheme is attracting substantial industrial and scientific interest in which the A-stage role is to maximize carbon diversion to the recovery streams and maintain stable low carbon influent to the B-stage. This is attained by operating the process in a very dynamic manner where high mass and volumetric loading rates are applied. As a result, a key challenge in the A-stage process is sludge settleability and the associated maintenance of effluent quality, particularly effluent suspended solids (ESS) [1]. In an operational-related context, it was reported that return streams can have a noticeable effect on A-stage performance [5]. It was found that bio-augmenting the A-stage process with waste nitrification sludge (WNS) may positively influence process performance, whereas supernatant return flow from solids processing (SPB) may negatively influence process stability [5]. However, this was never quantified or transformed into a control strategy for the return streams This is particularly of operational significance because such variables (WNS and SPB flows) are the key manipulated variables, given that solids wastage in Blue Plains is highly dependent on the capacity of the solids processing facilities rather than process performance. Hence, there is a need to develop a robust control strategy for ESS minimization, based on WNS and SPB. The bottleneck for implementing such a control strategy is to develop a model that can overcom the complex and dynamic nature of A-stage process. To this end, this study presents two data-driven modelling approaches to predict ESS of the full scale HiCS system operating at Blue Plains WWTP. This is the first machine learning (ML) model to be developed for an A-stage process. Methodology Modeling are based on a 2-year dedicated data effort from the WEST train of full-scale HiCS system at Blue Plains advanced WWTP. In addition to primary effluent, HiCS handles return flows such as SPB and WNS. Process parameters that are routinely measured include suspended solids, COD fractions, environmental and operational conditions (i.e., SRT, SOR, SLR, MLSS), advanced settleability measurements (i.e., TOF, and EPS TKN). Such variables represent possible model features (inputs), whereas the output (target) is process ESS, which was consistently monitored. Findings The data considered for this project were collected from Online sensors, DC Water Lab, and Research team. As shown in Figure 1, the obtained features have significant variation in frequency (how frequent these features are measured) and consistency (consistency in measuring features at the suggested frequency). This can be referred to the complexity of the process, which necessitates the consideration of sophisticated parameters like TOF, and LOSS [1]. In addition, sensing associated challenges have a key role to play [3]. To address challenges with the data, two modelling approaches are presented: (1) high frequency features classification (HF-C) model, and (2) all features regression (AF-C) model. In the HF-C model, the problem was approached as a classification problem where ESS was classified into three classes (good: <25 mg/L, action needed: >25 mg/L & 35 mg/L <, and bad: >35 mg/L). In this model, only features with high frequency were considered. Such features are mostly the flow-based features such influent, return streams, and wastage flow rates. AUC-ROC and F1 score along with accuracy were used to assess model performance which considered the effect of the less represented classes. Among different algorithms tested, the performance of four algorithms stands out, as shown in Figure 2a. When further evaluated based on the recall of the minor classes and extent of overfitting, it can be seen from Figure 2b that the support vector machine with radial basis function (SVM-rbf) is the best performing model. For the AF-R, the framework, demonstrated in Figure 3, was developed to circumvent the challenges of predicting dynamic and highly non-linear targets, significant variations in features availability, A-stage complexity, and data imbalance. A three-step data imputation approach where adopted to alleviate the effects of the high missingness in the data. Figure 4 shows the resulting outcome of the target and other features after outlier removal, and imputation. Subsequently, the model was developed in a multi-step fashion where the 7-day and 3-day moving average (MA) were first estimated and used as inputs to predict the Raw ESS. Figure 5 shows the effect of optimizing model architecture. These optimizations included the utilization of a weighted loss function to count for data imbalance and stratified cross validation to equally sample from different periods over the 866 days. These modifications resulted in an R2 of 0.72 and 48% reduction in RMSE (in comparison with average) using RF. This is compared to R2 of 0.2 to 0.3 for the model before optimization. Finally, it was observed that the integration of settling features (TOF, LOSS, EPS COD, and SVI) and operational parameters (SRT, WAS solids, and MLSS), and WNS characteristics (WNS COD fractions) significantly improved model performance. Figure 6 shows an example of the results when such features are introduced to the model where R2 above 0.85 was observed with more than 65% reduction in RMSE.