Insightful Analysis and Prediction of SCOD Component Variation in Low-Carbon/Nitrogen-Ratio Domestic Wastewater via Machine Learning
1. Introduction
Biological denitrification requires sufficient carbon sources to achieve higher nitrogen removal efficiency [1]. Especially in heterotrophic denitrification, carbon sources provide electron donors for the denitrification process [2]. However, the current domestic sewage in China faces the problem of a low carbon-to-nitrogen ratio (C/N) in influent water. In order to ensure the efficient removal of nitrogen, many studies have proposed that denitrification efficiency can be enhanced by additional dosing of sodium acetate, glucose, methanol, etc. [3,4,5]. However, the addition of external carbon sources requires a profound understanding of the COD components of the influent water in the local wastewater treatment plant (WWTP), as well as the kinetics of degradation and utilization processes [6,7].
Numerous studies have shown that different COD fractions can exert great influence on effluent quality [8,9,10]. For example, one of the most popular software, BIOWIN 6.2, classified COD components into non-biodegradable COD (NBCOD), readily biodegradable COD (RBCOD), slowly biodegradable COD (SBCOD), etc. [11,12,13]. The efficiency of nitrogen removal can only be optimized by adding external RBCOD at the most appropriate time, for example, when NH4+-N is fully oxidized to NO2−-N and NO3−-N. Therefore, the rapid measurement and analysis of the proportion of different COD components in wastewater and its degradation process could provide a certain reference for the precise addition of carbon sources and the enhancement of nitrogen removal efficiency in wastewater plants. Three-dimensional excitation–emission matrix (3D-EEM) spectroscopy has recently been recognized as an effective method to qualitatively and quantitatively characterize dissolved organic matters (DOMs) in wastewater treatment processes [14,15,16]. Guo, L. et al. indicated that the NO3−-N reduction process is closely related to the utilization of SCOD [17], and the soluble tyrosine-like proteins and tryptophan-like proteins were the most easily utilized carbon sources during the denitrification process according to the 3D-EEM analysis [18]. The above studies have identified biodegradable carbon sources through 3D-EEM, which in turn provides us with richer and more comprehensive information than SCOD in wastewater. Tang et al. [19] further assessed the correlation between PARAFAC components and water quality parameters such as COD, DOC, Chlorophyll a, TN, TP, and NH4+-N using Pearson’s correlation coefficient and redundancy analysis (RDA). The results indicated that COD concentration was positively correlated with humic-like substances. Several studies [20,21] have sought to investigate the potential of fluorescence excitation–emission (F-EEM) spectroscopy as an alternative analytical method for assessing the relationship between the presence of crucial drugs of addiction, ammonia, and pH with PARAFAC components. However, the correlation analysis techniques used in the above studies were RDA, canonical correlation, or linear correlation analysis methods.
In actual fact, the intricate mechanisms of fluorescence and the presence of irregular phenomena, which pose challenges for explanation, have raised concerns among researchers regarding the adequacy of linear relationship modeling in characterizing dissolved organic matter (DOM) during the analysis of EEM data [22]. In order to accurately predict the carbon source components and concentration during the different stages of wastewater treatment, it is of vital importance to establish the correlation between SCOD and different EEM components at specific excitation and emission wavelengths. Therefore, in the face of information-rich 3D-EEM data, the application of ML would be promising in establishing a prediction model for components of EEM extracted by PARAFAC and SCOD components, which can effectively reduce the amount of prediction for water quality analysis by extracting the main analytical constituents, and thus provide a technical reference for the operation and control of wastewater plants.
Recently, machine learning (ML), which is regarded as one of the technical approaches of artificial intelligence, has demonstrated unique performance in describing the relationship between various input and output factors [23,24]. The application of ML in fields including water quality prediction [25,26], water source classification [24], source tracing, and the aerobactin assessment of contaminants [27,28] has received significant attention in previous studies. ML involves a variety of algorithmic models, such as artificial neural networks (ANNs) [29], K-nearest neighbors (kNN) [30], decision tree (DT) [31], and gradient boosting decision trees (GBDTs) [32]. These models are implemented to achieve the predictive control of effluent quality such as biological oxygen demand (BOD), chemical oxygen demand (COD), and nutrient concentration. According to previous studies [33,34,35,36,37], a dynamic kernel extreme learning machine was proposed, including 170 samples and eight variables, to predict the COD proportion of industrial wastewater, and achieved a 10-fold cross-validation R2 of 0.708 [38]. Alavi et al. [39] proposed a novel computing algorithm that integrates an intelligent optimization algorithm with a KELM for the prediction of inlet COD concentrations in WWTPs. This study also compared the performance of different algorithms for optimizing real-time COD prediction. Zhao et al. [40] utilized six kinds of machine learning models to estimate the RBCOD and SBCOD in municipal wastewater with an R2 higher than 0.80 by inputting oxidation–reduction potential (ORP) data. Kim et al. [41] established high-performance (>95% of accuracy) ML models to predict the influence of different feeding carbon sources (acetate, glucose, and starch) on the microcosm communities of activated sludge. Therefore, machine learning has demonstrated excellent performance in past research; these algorithms help us to identify the complex relationship between input factors and output factors, and achieve high levels of accuracy and generalization. Nevertheless, they were in a superficial level of ML application, that is, merely focusing on predicting and analyzing simple important features. Moreover, few research studies reported a practical technical reference to assist WWTPs to guide the operation and improve the SCOD removal performance.
To provide a detailed analysis of the degradation, transformation, and removal processes of carbon sources in three WWTPs facing a low C/N problem, we analyzed the degradation process of SCOD in aerobic and anoxic processes by conducting batch experiments in the lab, and identified the key carbon sources that could be easily degraded using PARAFAC technology. In addition, we developed a GBDT-based model related to SCOD and four PARAFAC components through comparing four ML models. We also compared the four models in terms of model interpretation using Shapley additive explanations (SHAP) analysis to avoid the problem of an ML “black box”. Furthermore, to make the developed ML model more practical, we identified PARAFAC components closely related to SCOD removal in wastewater, and attempted to accurately predict the carbon source degradation process of the target WWTPs. The ML model proposed in this study is expected to be a valuable tool for further application in process optimization and accurate carbon source dosing for practical WWTPs.
