Alternate: Forecasting Volatile Solids Reduction of Municipal Sludge Using 32 Years of Data

The prediction of performance indicators such as Volatile Solids Reduction (VSR) is key for the accurate sizing of municipal sludge digesters. However, prediction models available from the literature typically lack accuracy, leading to the use of conservative values and inadequate process sizing. The present study aims at developing an empirical model for VSR prediction. This model is based on over 32 years of operational data originating from six industrial plants featuring a diversity of wastewater treatment lines. The model was benchmarked against the Anaerobic Digestion Model No. 1 (ADM1) with or without calibration and displayed the lowest average prediction error (RMSE of 4.2%VSR) and bias (2.3 %VSR). These high performances were obtained when predicting VSR for plants, independent from the empirical model training data (i.e., unseen data), highlighting its applicability for design tasks. During the design phase, the proposed model can promote optimal sizing of biogas networks and ancillary equipment, reducing inefficiencies and operational costs. For operational optimization, the model supports dynamic sludge age control strategies to maximize methane yield, contributing to sustainable wastewater treatment practices. In the future, the proposed model can still be improved upon by integrating new operational data with the already available training data used in this study. INTRODUCTION Anaerobic Digestion (AD) is a crucial technology for sustainable waste management, offering the benefits of biogas production, solids reduction and nutrient recovery. With the depletion of fossil energy resources and a growing global population, AD is increasingly important for achieving energy neutrality in wastewater treatment plants (WWTPs). Volatile Solids Reduction (VSR) is a key performance indicator of AD systems. Accurately predicting the VSR is hence essential for optimizing AD plants sizing and performance. Existing design models often lack accuracy, leading to conservative designs and inefficiencies. This study aims to develop an empirical model for VSR prediction using a total of 32 years of data originating from six industrial plants, providing a more reliable tool for AD system design and operation. MATERIALS AND METHODS Plants Description: The study analyzed data from six WWTPs located in diverse regions, including Oceania, the Middle East, South America, Central America, and Western Europe with data from all plants pooled. These plants employed various wastewater treatment processes, such as Conventional Activated Sludge (CAS) and High Rate Activated Sludge (HRAS), with or without primary settling tanks. The sludge types processed included primary sludge, waste activated sludge, and combinations of both. Data Cleaning and Enrichment: To ensure the accuracy of the model, data anomalies were identified and removed. These included periods of digester shutdowns and startups, as well as anomalous observations based on industry standards. Missing values were imputed using an Expectation-Maximisation algorithm, and outliers were detected and removed using the Mahalanobis distance method. Anomalous periods were further investigated using mass balance checks and consistency verification between process inlets and outlets. Variables Definition & Calculation: Key variables for the model included wastewater temperature, Total Solids (TS) concentration, Volatile Solids (VS) fraction of primary and waste activated sludge, and the ratio of primary sludge to waste activated sludge at the digester inlet. Calculated variables included sludge age, Hydraulic Retention Time (HRT), and VSR. VSR was determined using either mass-balance and Van Kleeck methods. Empirical VSR Modelling: A linear model with feature engineering (Lasso) was developed, incorporating commonly measured process variables known to impact AD performance. These variables are the HRT, sludge age, temperature, primary sludge VS content, and the mass-based fraction of primary sludge in the mixed sludge fed to the AD system. The model was trained using a leave-one-plant-out (LOPO) validation approach, ensuring robustness and generalizability. Mechanistic VSR Modelling -- ADM1: The Anaerobic Digestion Model No. 1 (ADM1) was used for comparison, both with and without calibration. Key parameters, namely overall degradability (which is analogous to Biochemical methane potential), and hydrolysis coefficient, were calibrated using a grid search algorithm. The performance of the ADM1 model was assessed using the same test datasets as the empirical model. RESULTS AND DISCUSSION Empirical Model Performance: The empirical model outperformed the ADM1 model in predicting VSR, both in its calibrated and uncalibrated forms. The empirical model demonstrated lower Root Mean Square Error (RMSE) and bias, indicating higher accuracy and applicability for design tasks. The model's ability to integrate seasonal variability and provide reliable predictions without requiring plant-specific calibration was a significant advantage. Model Assessment: The empirical model's performance was validated across different plants and conditions, showcasing its robustness. The model's integration of seasonal variability, such as wastewater temperature, likely contributed to its superior performance compared to the ADM1 model, which assumes constant feedstock characteristics. The empirical model's ability to avoid overfitting, particularly in cases with limited data, further highlighted its reliability. Implications for AD Systems Design: Accurate VSR prediction is critical for optimal sizing of biogas networks and ancillary equipment. The empirical model reduces the risk of under or oversizing, leading to cost savings and improved operational efficiency. For instance, an 8.7% VSR uncertainty with the unfitted ADM1 model (vs. 2.3% VSR with the empirical model) could lead to a 19% relative error in biogas system sizing, compared to a 5% relative error with the empirical model. This precision is crucial for maintaining stable operation and minimizing costs associated with equipment purchase, maintenance, and efficiency. Implications for AD System Operation: The empirical model supports dynamic sludge age control strategies, enhancing methane yield and contributing to energy neutrality in WWTPs. By optimizing the activated sludge age, the model can significantly impact both water and sludge treatment lines. For example, reducing the activated sludge age from 22 to 14 days at a wastewater temperature of 17°C can increase VSR from 19% to 27%, translating to a 42% increase in biogas production. This optimization can enhance oxygen transfer in the aeration tank as well as improve AD performance, supporting a dynamic control strategy that accounts for seasonal variability. CONCLUSION The empirical model developed in this study represents a significant advancement in predicting VSR for municipal sludge digestion. Built on 32 years of data from six industrial plants, the model offers accurate and reliable predictions, supporting the design and operation of more efficient and sustainable AD systems. It may perform on plants outside its experience envelope, which can be addressed by increased training sets. By reducing the risk of under or oversizing and enhancing operational strategies, the model contributes to the broader goal of energy neutrality in WWTPs. Future improvements can be achieved by incorporating more operational data, expanding the dataset to include a wider variety of industrial plants and operational conditions, thereby improving the model's robustness and reliability.