Effluent Excellence: Unleashing Hybrid Model Magic at Fond du Lac

Background and Objective:
Wastewater treatment facilities must meet stringent effluent water quality targets while managing higher flows from extreme weather conditions. They rely on activated sludge models (ASM) for design capacity evaluation (Henze et al., 2000). For smaller utilities, maintaining these models is challenging due to frequent updates, recalibration based on influent flow characteristics, input data quality issues, and manual scenario simulations. Recently, hybrid process models combining predictive machine learning (ML) for influent forecasting, automated data handling for ASM, and auto-calibrated simulations have been proposed to increase modeling tool adoption (Mannina et al., 2019; Zhao et al., 2020). Automatic data storage and processing remain challenging due to variability in lab methods, analyzer drift, non-standard naming conventions, poor data organization, and cybersecurity concerns (Demir & Szczepanek, 2017 and Newhart et al., 2019). This study explores hybrid modeling at the Fond du Lac Wastewater Treatment & Resource Recovery Facility (WTRRF) in Wisconsin. Key facts about the WTRRF (Figure 1):
- Rated for 12 mgd (~44,000 m3/day), serving a population equivalent of 60,000 people.
- Utilizes EBPR processes combined with Chem P removal for phosphorus trimming. Managed extreme flow conditions in 2024, while maintaining low effluent TSS and nutrient limits.
- By 2027, a 6-month mass load allocation equating to an average total phosphorus (Total P) concentration of 0.19 mg/L-P is anticipated. It is aimed to meet this goal while minimizing operational and maintenance costs for chemical phosphorus removal and future capital costs. Fond du Lac WTRRF collaborated with Black & Veatch and Maia Water to design and deploy a hybrid model optimizing operational decision-making around flow and load management.

Approach:
The cloud-based hybrid modeling approach for WTRRF included an ML model for primary effluent flow and load forecasting, followed by an ASM to estimate effluent quality and performance using the application package interface (API) for SUMO22 (Dynamita, France) (Figure 2). Process data was imported, supplemented with live weather data to generate a historical dataset for the ML algorithm. The dataset was used to auto-calibrate a 1-D secondary clarifier SUMO model (Figure 3) to match observed secondary effluent TSS. Next, future influent flows and concentrations were forecasted using XGBoost and input into a larger plant model (Figure 4) with calibrated parameters. The SUMO model simulations provided performance forecasting and soft sensor capabilities. The workflow's modularity enabled new optimization routines, updates, and adoption lab results at varying frequencies. It culminated in a hybrid modeling dashboard as a data-backed advisor to Fond du Lac, offering visualizations, optimization routines, sensitivity analyses, and model training options.

Findings & Significance:
Influent Forecasting: For a 7-day trend of 15-minute interval influent flow, a mean absolute percentage error (MAPE) of 26.9% was observed between forecasted and actual flows (Figure 5). A rain event far exceeded forecasts (Figure 6), impacting flow forecasting accuracy. The MAPE between forecasted annual flow relative to actual was 16.2% (Figure 7). A reasonable match was observed between influent total COD and total P forecasts. The MAPE for COD was 9.4% (Figure 8) and for TP was 7.5% (Figure 9). These forecasts indicate that operations at Fond du Lac WTRRF can be augmented with forecasts for improved decision-making. The current ML forecasting model will be optimized as more data becomes available.

Hybrid model: The forecasted secondary effluent TSS from SUMO matched well with observed data, with an MAPE of 12.58% and absolute differences within ±0.8 mg/L (Figure 10). Sensitivity analysis can demonstrate differences in forecasted effluent TSS quality with clarifiers in/out of service for maintenance planning. Forecasted effluent TSS supports improved decision-making around clarifier operations.

The hybrid model indicated better P removal from EBPR and Chem P versus observations (Figure 11). The difference underscores challenges in predictively modelling EBPR and Chem P at low OP concentrations due to relevant parameters' complexity. While the absolute error between observed and forecasted effluent OP was 0.05 mg/L-P, less than industry standard stop criteria of ±0.5mg/L-P (IWA, 2013), a more accurate match is required. An effluent OP error correction model using XGBoost and several features from influent, operational, temporal, and model-derived data improved effluent OP forecasting. Model-derived features indicate P loading to carbon uptake ratio by polyphosphate accumulating organisms over relevant time scales. The error correction allowed for a simpler regression model factoring system loading history. Figure 12 shows testing data from this model that will be integrated with outlier detection in final deployment.

This approach unlocked soft-sensor capabilities from model results: model estimates for rbCOD in selector effluent were compared to observed settleability revealing potential correlations. Spikes in rbCOD from selector effluent led to a rise in SVI-30 (Figure 13). Further analyses are planned to reveal more correlations using soft-sensor values, opening opportunities to leverage state variables for making informed process decisions.