Polymer Optimization Using Machine Learning

BACKGROUND Jacobs operates two wastewater treatment plants for the City of Vancouver Washington, across the Columbia River from Portland Oregon. The West Side WWTP is a 28 MGD capacity plant that uses a biological nutrient removal process, followed by centrifuge dewatering and incineration. The project team is led by Travis Capson, with Sam Mielcarek as operations manager and Paul Muirhead as maintenance manager. Optimizing polymer dosage for solids thickening or dewatering is an especially challenging problem in our industry, with at least a dozen potential variables to consider when trying to maximize dewatered solids at the lowest overall cost: feed solids concentration, flow rate, polymer dosage and concentration, temperature, feed solids characteristics and blends, machine settings, etc. The site operations staff already fully understood the complexity of the challenge but was forced to spend extra effort tracking and making frequent changes to various settings in the system every day, all while balancing the competing demands of polymer usage cost vs. the cake solids dryness and fuel cost at the incinerator. DISCUSSION A machine learning model that matched the biggest drivers of polymer dosage against actual dewatered cake solids content values multiple times per day, over several years, was created. This involved leveraging recent upgrades to the SCADA control systems designed by Jacobs at the plant to support the capture of real time data for accurate dosage predictions. Starting in May 2023, push notifications for dosage recommendations went live to field staff. Gradually at first, but with increasing confidence, the staff have been following the predictive recommendations and seeing reduced polymer usage and utilization throughout FY 24 is shown in Figure 1. The predictive data science approach to polymer dosage control is savings around 10%-20% - an improvement on a budget of over $700,000 per year in polymer. The intent behind this optimization strategy is to reduce polymer consumption while maintaining appropriate cake solids concentrations for downstream processes and off-hauling. The plant has indicated a target concentration of 26-28% as optimal based on their experiences and the figure below provides a summary of where the system operates below range and when it has operated within range historically. Although It is understood that polymer may not be the only factor for contributing to differences in cake TS concentrations, for simplicity, prior to deployment of the optimization of a preliminary opportunity assessment was done to check these ranges as a function of polymer dose and simplify this to 'overdosing' and 'underdosing'. Based on the historical data going back to January 2021, 54% of the time the system operates within the expected range, 30% of the time above the expected range (overdosing 30% of the time), and 15% of the time below the expected range (underdosing 15% of the time). and thus the opportunity stemmed from being able to cut the 30% overdosing value to being within the desired Cake TS range. A multi-headed convolutional neural network (MH-CNN) approach was used with a temporal component with lookback features (as predictors, response, control variable -- polymer feed) and lookforward features (response and control variable) technique. For each prediction, the model looks back at a period of data and based on what it observes predicts what will happen over the lookforward period. This allows for the MH-CNN to do its own lag engineering by identifying the phemonena within the lookback window are most predictive for lookforward dynamics an schematic illustrating this architecture is shown in Figure 3. In addition to the data provided from the facility, an acoustic sensor was installed on the external wall of pipe carrying post-centrifuged cake. The theory behind this was to use the vibrations and noise measured at different frequencies when concentration of cake varies in order to predict cake TS. This was a non-intrusive install and sensor was glued to the pipe at the discharge point of schwing pump. The location of this was set such that it was installed close to a wall to minimize readings of ambient and non-relevant noise signals. Figure 4 shows the location of this installation. The site has robust sampling plan of measuring total solids at various locations in the process and this data is used to validate and continuously improve the results of the models. Figure 5 shows the results of the model's ability to predict cake TS concentrations. The test set shows a Pearson's R correlation value of 0.59 with a p value less than 0.05 indicating a statistically significant correlation. The variable importance figure shows the importance of the relative features used to train the model where WAS feed rates and the installed acoustic sensor showed significant importance on the model's ability to predict cake TS concentrations. This case-study demonstrates significant progress in the successful implementation of AI at a treatment facility. The strong engagement with operators has advanced their skillset such that they are submitting abstracts to their local operator's conferences at their own volition.