Predictive Iron Dosing For Phosphorus Removal: A Data-Driven Strategy

Introduction
DC Water owns and operates the Blue Plains Advanced Wastewater Treatment Plant (Blue Plains AWTP) which serves Washington DC and parts of Virginia and Maryland. The Blue Plains AWTP averages 384 mgd treated and has stringent nitrogen and phosphorus limits. The plant performs primary clarification, high-rate carbon removal (secondary treatment) followed by a separate nitrification and denitrification process with methanol addition for nitrogen removal (NiteDenite), effluent filtration and disinfection. Removal of phosphorus is largely accomplished by ferric chloride dosing with three possible dosing locations; upstream of the primary clarifiers, in the secondary treatment process and the NiteDenite process. The effluent phosphorus limits are 0.17 mg/L (average monthly) and 0.34 mg/L (average weekly). With these stringent limits, and the complexity of the process flow, managing iron dosing is crucial for a number of reasons. First, is the occasional overdosing of iron, which may cause P-limitations in the secondary and NiteDenite processes. Second, is the slow response of iron dosing when increasing trends in effluent phosphorus is observed. Finally, due to the experienced increase in ferric costs over the years, Blue Plains AWTP operators and engineers are seeking to optimize the dosing to more efficiencty remove phosphorous and cut down the operational cost. Phosphorus (P) removal using iron salts relies on effective coagulation and flocculation processes. The addition of ferric salts to wastewater results in rapid formation of hydrous ferric oxides (HFO) flocs which then co-precipitate with orthophosphate (OP) while subsequent removal is due to adsorption and complexation. Due to the different mechanisms of P removal and the complex nature of the mechanism, optimization of chemical P removal using iron can be difficult. Furthermore, several factors such as pH, iron dose, mixing conditions, retention time, wastewater composition, and HFO aging can affect P removal. Previous analysis and preliminary modeling by staff at the Blue Plains AWTP suggested auto-correlation of effluent P concentrations with previous week's effluent P concentrations suggesting potential lag in response to ferric dosing. Furthermore, automated control of ferric dosing based on OP analyzers in the influent did not seem to provide any additional benefit based on testing by plant staff. The objective of this study was to use a data-driven approach leveraging advanced analytics, machine learning (ML)-based classification and modeling to provide operational guidance and control strategy for optimizing ferric dosing for chemical P removal to meet effluent discharge limits while stabilizing process performance in secondary and NiteDenite processes while reducing costs.
Methods and Results
Daily monitoring data from 2012-2020 were obtained from the Blue Plains AWTP and uploaded to Seeq®, which is an advanced process analytics platform for time series data. Data cleaning and noise reduction was performed through outlier removal and signal smoothing using the Loess method. Initial analysis showed that the ferric dosing was focused on flow pacing during steady state conditions with manual increases in ferric dosing in response to effluent P spikes (Figure 1). This typically resulted in a delayed response since the effluent P concentrations were lab measured daily data. Influent P spikes did not always result in an increase in effluent P concentration and hence, to avoid overdosing of ferric, manual increases in ferric dosing were employed based on operational experience. Furthermore, there were several influent P spikes (Figure 1) which did not show any corresponding increases in effluent P which makes optimization of ferric dosing a complex exercise. An analysis of P fractions (Particulate P, soluble organic P (SOP) and OP) was undertaken based on measured total phosphorus (TP), total soluble phosphorus (TSP) and OP (Figure 2). This showed that there were several intermittent spikes in effluent P due to spikes in particulate P possibly due to bleeding of solids from the filters, however, this would typically not cause problems with permit violations due to their short-term nature. There were also periods where sustained long-term increases in OP concentrations were observed, which were considered causes of concern since these did not immediately respond to manual increases in ferric dosing at the plant. The overall iron dose (Fe3+:P) ratio can be a key factor in determining efficiency of P removal. A correlational analysis of iron dose (mol:mol) was performed and showed no correlation of iron dose with effluent OP. It also showed that the plant staff have been reducing the iron dose closer to a Fe3+:P ratio of 1 mol/mol in an effort to reduce iron usage and costs (orange points in Figure 3). In order to determine if the dosing location had a significant impact on effluent OP concentrations, the primary Fe: secondary Fe dosing ratio was calculated, and a correlational analysis was performed with effluent OP (Figure 4). The analysis showed no correlation suggesting that primary vs secondary dosing had no change in effluent OP. It also showed that the plant staff had been moving towards a more balanced primary vs secondary dosing approach (Primary Fe:Secondary Fe dose ~ 1) to avoid P-limitations in the B-stage of the process. Preliminary linear modeling was conduced to determine key independent variables controlling effluent P concentrations. Influent TP loads, primary and secondary Fe dose and secondary effluent TSS concentrations were used as predictive variables (Figure 5 and Table 1). Influent TP load (moved forward by 3 days) showed a significant impact (coefficient = 0.014, p < 0.0001) on effluent P concentrations which suggested that effluent P lagged the influent P concentrations by 3 days. Addition of secondary effluent TSS lagged by 7 days showed an increase in predictive capability and had a positive correlation with effluent P (coefficient = 1.18, p < 0.0001). The hypothesis is that secondary effluent TSS is an indicator of colloidal and particulate P entering the NiteDenite process where potential release of P under high SRTs and RAS recycling could result in effluent OP bleed through. The simple linear modeling was able to replicate the baseline effluent P concentrations and some of the peaks, however, it was not able to predict the magnitude and duration of the peaks. Furthermore, preliminary process modeling showed that while it was able overall to predict P removal in the process, the spikes in the OP in the effluent were not simulated. The full presentation will aim to address these gaps using machine learning based modeling and classification approaches to provide predictive capability and operational guidance for ferric dosing. Additional correlations with respect to the change and/or the rate change in effluent OP concentration will be explored to evaluate the impact of several independent variables.