AI on Tap: Revolutionizing Influent Flow Forecasting with Ensemble Learning

[b]Abstract[/b] Accurate influent flow forecasting is critical for resilient operations at water resource recovery facilities (WRRFs), enabling proactive decisions on process control, staffing, and wet-weather readiness. Traditional models often struggle with irregular flow patterns driven by climate variability and land-use changes. This research developed a scalable ensemble machine learning model that produces up to 7-day forecasts using limited historical flow, meteorological, and temporal data. The ensemble model combines four different machine learning (ML) algorithms through a meta-learner with adaptive preprocessing to enhance generalizability across diverse sites. Evaluated across 30 WRRFs (average flows from 0.3 to 146 MGD), the model achieved an average mean absolute percentage error (MAPE) below 11%, including extreme wet-weather events, demonstrating reliability across varied operational scenarios. [b]Introduction[/b] Forecasting influent flows at WRRFs is essential for optimizing operations, ensuring compliance, and improving resilience (Kumar & Goyal, 2025). Accurate forecasts help operators anticipate loading, optimize chemical dosing, allocate staff, and schedule maintenance (Wang et al., 2021; Ye et al., 2025). In combined sewer systems, reliable predictions help mitigate wet-weather impacts, where surges can exceed capacity and compromise effluent quality (Yu et al., 2023; Zhou et al., 2023). Traditional methods such as statistical extrapolation and mechanistic hydrologic models struggle under increasingly irregular conditions driven by climate change and land-use shifts (Zhou et al., 2023). ML offers a data-driven alternative to uncover complex relationships between flow, weather, and temporal patterns (Wei et al., 2023). Most prior ML applications focus on short-term predictions (up to an hour), which are limited in scope and utility for broader planning and operational needs (Wei & Kusiak, 2015). Few ML applications address longer horizons, and those that do are typically data-intensive and site-specific (Ye et al., 2025). To address these gaps, we propose an ensemble ML model for influent flow forecasting, balancing predictive accuracy with scalability across diverse facilities. [b]Research Approach[/b] This research introduces a scalable, data-efficient solution built on an ML model capable of forecasting influent flow up to 7 days ahead. The solution was tested using data from 30 WRRFs, spanned a 3-year period: 2022–2024. The average daily flow (ADF) for the 30 WRRFs ranged from 0.3 to 146 MGD. The maximum daily flow (MDF) to ADF ratio was in the range of 1.3 to 7.5; with the highest peak flow approaching 300 MGD. A summary of the flow and precipitation data is provided in Table 1. The solution described in this research integrates four well established ML algorithms: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) in an ensemble model via a meta-learner. The ensemble model leverages the complementary strengths of each of these algorithms for robust predictions (Figure 1). The tree-based algorithms (RF and XGBoost) excel at capturing the non-linear relationship between the predictors. The neural network algorithms (LSTM and GRU) specialize in describing the temporal dynamics and sequential dependencies in the data. [b]Results and Discussion[/b] The Influent Flow Forecaster Ensemble demonstrated strong generalizability across 30 WRRFs, maintaining an average MAPE below 11% for 7-day prediction horizons and achieving the highest Nash–Sutcliffe Efficiency (NSE; 0.45) among base models (Figure 2). While the average MAPE for the Ensemble and XGBoost are similar, the Ensemble delivers greater reliability, evidenced by a narrower interquartile range, indicating more consistent performance across WRRFs. Figure 3 shows the Ensemble's 7-day ahead predictions for the WRRF with the highest ADF (~663,700 m3/d) and an MDF/ADF peaking factor (PF) of 2.05, where stable flow dynamics enabled strong predictive performance. With only two years of training data, the Ensemble achieved an NSE of 0.73 and MAPE of 5%, demonstrating high accuracy in high-flow environments. To show adaptability across extremes, Figure 3b presents results for one of the smallest WRRFs (ADF ~1,360 m3/d and PF ~5.7), where high variability and weak precipitation-flow correlation complicate forecasting. Despite these challenges, the Ensemble achieved NSE=0.58 and MAPE=4%, underscoring its adaptability and its scalability. A 30-facility evaluation revealed that adaptive preprocessing is a key factor for generalizability of the flow forecasting model. For facilities with distinct flow dynamics, the application of advanced preprocessing techniques led to marked improvements in forecast performance. On average, a 12% increase in NSE was observed across nearly half of the datasets, with some instances achieving gains as high as 43%. Additionally, reductions in MAPE of up to 5% were noted in approximately one-third of the datasets, highlighting the importance of site-specific data conditioning. The Influent Flow Forecaster demonstrates the effectiveness of combining ensemble ML with strategic preprocessing for reliable, scalable influent flow predictions. The approach supports proactive operational planning and resilience under dynamic conditions, outperforming individual learners in both accuracy and consistency.