A study of the efficiency of sewage treatment in response to temperate variation by machine learning Open Access

As a vital hub for China's economic development, the Yangtze River Basin (YRB) is densely populated and heavily industrialized. It generates significant quantities of sewage, most of which is collected and piped to sewage treatment plants. The wastewater treatment process (WWTP) degrades sewage contaminants by integrating physical, chemical, and biological techniques, which are influenced by various factors, including temperature, pH, and toxic substances (Wang et al. 2023).

Specifically, the temperature is thought to be one of the pivotal factors impacting the efficiency of the treatment by interfering with the biological degradation of organic materials and the formation of physical and chemical substances in the sewage (Gnida et al. 2016). Biological treatment stages involve the microbial activity of breaking down organic pollutants, including activated sludge processes and bio-filters, which are sensitive to temperature changes (Shatat & Al-Najar 2011; Li et al. 2022). Most microorganisms play the optimal role at 20–45 °C (Kanazawa et al. 2019). Otherwise, deviations from this range may diminish microbial metabolism and decrease the biodegradation rate of organic materials and overall treatment efficiency (Schiraldi & De Rosa 2016). Besides this, temperature has an inverse relationship with the solubility of gases in water, which affects oxygen, a vital factor for aerobic biological treatment. The dissolved oxygen becomes less in water as the temperature rises, which can hamper aerobic microbial activities and affect the efficiency of biological treatment processes; at elevated temperatures, some processes relating to dissolved oxygen, like nitrification (the aerobic conversion of ammonia to nitrate) and phosphorus removal may accelerate (Junhong & Minyu 2019). However, excessively high temperatures will make these processes overly rapid, leading to potential inefficiencies or necessitating adjustments in process control (Hu et al. 2023).

In summary, temperature variations, especially external climatic change, affect the effectiveness of sewage treatment systems by altering the microbiological activity, chemical reaction rates, and physical properties of sewage. The Yangtze River experienced such a climate change, with the annual average Normalized Difference Vegetation Index (NDVI) significantly increasing at the rate of 0.002/a during 1998–2018 (Cui et al. 2024). However, the degree of the efficiency of sewage treatment in response to external climatic variations is still unknown.

There has been extensive usage of the related method to investigate how temperature changes affect the effectiveness of waste treatment, which encompasses a variety of techniques, including computational modelling, empirical investigations, and laboratory-scale trials (Claassen 2019). However, some are insufficient of the traditional techniques for identifying one of the factors contributing to the complex environmental system. In comparison, machine learning has the advantage of evaluating and forecasting complex systems as computational modelling. This allowed for a more thorough understanding of treatment process efficiency in real-world scenarios, which can comprehensively consider operational challenges, environmental factors, and actual influent variability (Zhu et al. 2024). In particular, neural network models are highly valued for their capacity to handle nonlinear interactions and adjust to changing circumstances, which is an effective tool for improving understanding and optimizing sewage treatment procedures (Zhang et al. 2023).

The primary objective of this study is to assess the impact of climate change on the operational efficiency of sewage treatment plants through the application of neural network models. Apart from pinpointing significant sensitivity indicators, the research methodology aims to clarify the relationship between the efficiency of complex sewage treatment processes and climate change.

The study's project site is located in Jiangsu province's Suqian, midway along the Yangtze River; the place experiences a transitional climate between warm temperate and subtropical zones. According to records from the Suqian Meteorological Station, the average annual temperature between 1960 and 2020 was 14.4 °C, and 211 days was the average length of the frost-free season it has. The maximum daily rainfall recorded is 253.9 mm, with an average of 916 mm of precipitation per year. The HEXI wastewater treatment facility (HXWWP), with the geography site (longitude 118.28447342, latitude 33.90750104), is selected as the case of this study.

The sewage sources of HXWWP contain both residential and industrial sewage. The latter waste comes from a wide range of industries such as building materials, furniture, metal products, thermal power, rubber and plastics, glass products, smelting, textiles and apparel, packaging, printing and paper processing, green food and beverage, electronics and IT, and equipment manufacturing, among others. HXWWP meets China emission standards Class I A discharge regulations by treating 100,000 tons of sewage per day using advanced treatment techniques and the A₂/O process, and the routine monitoring of the discharge water shows the levels of heavy metals meet the requirement of the permissible limits. The data used in this research comes from HXWWP's daily online monitoring records from 2020 to 2022, covering the water quality index both input and discharge water, along with index of sludge volume ratio (%SV), total phosphorus (TP), total nitrogen (TN), ammonia nitrogen (NH₃-N), suspended solids (SS), chemical oxygen demand (COD_Cr), and biochemical oxygen demand (BOD₅).

In this work, a feedforward backpropagation (BP) neural network model is built using MATLAB 2022a's Neural Network Toolbox, which is used for network design, optimization, training, and simulation. There are nine different types of input variables and six output variables in the dataset, which has 15 columns. Ambient temperature and inflow-water indicators, such as COD_Cr, BOD₅, SS, NH₃-N, TP, TN, dissolved oxygen (DO), and sludge volume percentage (%SV), are among the input water quality variables. In the meantime, the output variables are the outflow-water indices, which include TN, SS, NH₃-N, BOD₅, COD_Cr, and TP. In all, 16,440 data matrices were loaded into MATLAB for training.

The dataset has been categorized using a ratio of 70:15:15 for training, validation, and testing. Testing is essential for determining how the model distinguishes data points, validation helps optimize continuous network weight adjustment, and training instructs the network. Following setup, the topology, activation functions for every layer, and selection of training techniques are all part of the model's structure design. Through iterative cycles and weight adjustments, a refined model that skillfully links the input of the water quality with the output counterpart is constructed, enhancing insights on the performance of discriminant analyses across several simulated scenarios. The number of hidden layers and neurons in each hidden layer, the appropriate activation functions for the hidden and output layers, and the optimal training algorithm for optimizing gradient descent are taken when configuring a feedforward neural network.

Too complex networks frequently lead to serious overfitting problems and a decline in learning efficiency (Xipeng 2020). As a result, too many layers will not be considered in this study since the dataset employed for simulation is manageable for its size and complexity and is for a shallower-depth machine learning algorithm. Accordingly, traditional three-layer network architecture with one hidden layer is set, which will be more efficient in adding neurons to the hidden layer rather than creating new hidden layers when improving the model's learning ability becomes required. The importance of choosing the ideal number of neurons for the hidden layer when creating a BP neural network model is emphasized, which is essential for connecting layers within a network, and empirical formulas or the ρ value can be used to determine the ideal number of neurons for the hidden layer (Shi 2013).

Enlightened by this view, this study investigated the possibility of varying the number of neurons in the hidden layer, ranging from 10 to 30, and then determined the ideal arrangement. The output layer of the BP neural network used in this study uses the SoftMax function, the same Sigmoid function the hidden layer takes. The SoftMax function is essential to the regression algorithm and logistic regression, which can expand logistic regression into a multi-class procedure. It makes it very useful for building models with multi-class distinction.

Learning gradients through the BP of errors and training techniques are classified as the Levenberg–Marquardt algorithm, quantized conjugate gradient method, and Bayesian regularization (Bouzid et al. 2024). The training strategy in this study harmonizes the advantages of both the steepest descent and Newton's methods by using the method of the quantized conjugate gradient, an adaptive learning rate methodology, for updating weights and biases. It works only with first-order derivative data, avoiding the steepest descent method's sluggish convergence problem and the need to store and compute the inverse of the Hessian matrix, a known drawback of Newton's approach. As a result, there is little need for as much storage, rapid convergence, improved stability, and no reliance on outside factors.

The outside temperature has been artificially set as 0.2, 0.4, 0.6, 0.8, 1.0, 1.4, 1.8, and 2.2 °C, which reflect the variation of the external temperature change. The well-fitted BP neural network is adopted with the fake parameters to forecast how temperature variation affects treatment effectiveness.

Equation (1) calculates the FI-COD_cr value by deducting the input COD_cr value from the outflow COD_cr value of water. A similar method is used to calculate the other FI indicators.

A higher FI denotes more effective treatment, whereas a lower FI denotes less effective treatment. When the FI value is higher than zero, it suggests that the A₂/O process's treatment efficiency has decreased and that the outflow water quality is lower than the inflow quality.

Results show that the neural network model is well simulated and can provide a reliably predictive analysis.

This decreasing trend implies that as temperatures escalate from 0 to 1.5 °C, the A₂/O process becomes progressively more efficient at degrading COD_cr. However, this efficiency trend slowly plateaus when the temperature increment reaches 2 °C. The trend for BOD₅ is similar to that of COD_cr, and the increases in BOD₅ degradation efficiency become stable, being noticeable around 2 °C , which means the higher temperature can raise the degradation of BOD₅ in sewage. The FI-SS, FI-TN, and FI-NH₃-N indicators all show a linear increase in deterioration efficiency from 0 to 2 °C, and there is a clear linear association between temperature and the ability of the A₂/O process to degrade SS, TN, and NH₃-N. On the other hand, FI-TP has an increasing trend, indicating that the effectiveness of the A₂/O process is decreasing.

The A₂/O process of sewage treatment, an abbreviation for anaerobic–anoxic–oxic, represents a prevalent process in sewage treatment, capitalizing on the biological removal of nitrogen and phosphorus across anaerobic, anoxic, and aerobic phases. The efficacy of sewage treatment is affected by complex factors, including hydraulic retention time, temperature, and pH levels, epitomizing the nuanced interplay within complex environmental-process interaction systems (Viraraghavan & Kikkeri 1990; He-Ping 2015).

Exploring the factors of sewage treatment efficacy is important, as it can help aid the process. In the past, a range of techniques, including basic statistics, correlation analysis, principal component analysis, cluster analysis, analysis of variance (ANOVA), and time series analysis, were employed, and the relationships between indicators like pH, temperature, BOD5, COD_cr, SS, TN, and TP were examined. The results confirmed a significant link between TN and TP and a substantial correlation between BOD5, COD_cr, and SS (Park & Dho 2018), Temperature impacts both phosphorus absorption and nitrogen management, which are related to the efficacy of entire sewage treatment (Yoo & Lee 2015; Han et al. 2024).

As far as the entire procedure is concerned, it is a complex environmental system known as the intricate interaction, which is like a huge ‘black box’. Classical techniques may face great challenges in picking up some index role in the efficacy of such big interactions. Machine learning has shown to be more effective than traditional methods in analyzing complicated datasets, and neural network models, as one of the deep machine learning, provide such a perfect simulation by connecting in/out parameters units and assigning a weight to each counterpart connection to a computer program (Kiiza et al. 2020; Huang et al. 2023).

Among the techniques, BP is the fundamental component of neural network training, which can be achieved by adjusting a neural network's weights depending on the error rate recorded in the preceding epoch (i.e., iteration) (Huang et al. 2023). Additionally, artificial neural networks can be developed to perform human learning, computer speech, and other tasks, as well as create prediction models from huge databases (Chen et al. 2024).

Therefore, the BP neural network is simulated to understand how climate change affects the efficiency of the A₂/O process. HXWWP, a typical A₂/O technique widely applied throughout the Central YRB, is selected as a case study.

First, the model's three-layer neural network architecture comprises nine input layers of water quality index, 10 neurons in the hidden layer, and six nodes in the output layers (Figure 1). After simulation, 20 computation rounds are optimized to produce the predicted results (Figure 2). The error histogram also shows a normal distribution (Figure 3), it has a strong correlation in the regression analysis of the training, validation, and testing phases as well as the entire dataset (Figure 4). The efficacy of the BP neural network is strongly supported by the high accuracy rate of 95.62% of the total samples, demonstrating its ability to forecast changes in sewage treatment indicators in complex systems properly.

Second, to explore the impact of temperature on sewage treatment efficacy, the artificial temperature parameters are set artificially from 0.2 to 2.2 °C. In contrast, the other parameters, including the process type, inflow water quality, and inflow water volume, are kept constant. Then, a predictive result is obtained by combining the artificial neural network with the well-BP neural network simulation model. According to the preliminary research, even a small temperature increase of 0.2 °C causes the COD_cr and SS indicators to rise beyond zero, indicating a potential decrease in the efficiency of their deterioration. The TP and TN levels show little change, indicating that these indicators are not greatly impacted by even a small temperature increase (Figure 5).

The study extends the temperature variation from 0.2 to 2.2 °C to better understand the impact of temperature variations on treatment efficiency. FI values are created in different effluent indicators by the use of artificial neural networks to model changes (Figure 6); it is observed that FI-COD_cr and FI-BOD5 values decrease with increasing temperature, suggesting that the A₂/O process is more efficient at degrading COD_cr and BOD₅, however, this increase in deterioration efficiency peaks at a temperature increase of more than 2 °C. Furthermore, the degradation efficiency for SS, NH₃-N, and TN increases linearly with temperature, while the FI-TP values fluctuate and indicate a decrease in TP degradation efficiency as temperatures rise. These results highlight the complex effects of temperature fluctuations on the efficiency of the A₂/O process across various parameters.

An earlier study found that changes in temperature, as well as other variables, can affect the variety and composition of bacterial populations in activated sludge within sewage treatment systems. Another investigation pointed out that the main environmental factors influencing the microbial community structure are COD_cr, TP, total nitrogen, NH₃-N, and water temperature (Yun et al. 2021).

Consistent with these conclusions, our research also found that FI-COD_cr and FI-BOD₅ values decreased as the temperature rose, suggesting that the A₂/O process is becoming more efficient in degrading CODcr and BOD₅. The phenomenon can be interpreted as the rising ambient temperatures altering the composition of microbial communities and facilitating the degradation of BOD₅ and COD_cr. Moreover, the temperature at which the rise in degradation efficiency peaked is higher than 2 °C, which may be explicated as the activity of the microbial population is inhibited when the temperature is higher than a certain threshold. The degradation efficiency growth is thus weakened. The findings in this study are consistent with previous studies that increase the degrading efficiency of NH₃-N and TN with increasing temperature (Wu et al. 2014; Tingquan et al. 2015).

Another view is that the ideal growing environment has elevated temperatures for developing bacteria, which can break down nitrogen. Additionally, raising the temperature increases the effectiveness of SS removal in sewage (Tingquan et al. 2015), as verified in this research. The efficiency of TP in this investigation becomes weaker with increasing temperature. To find out the fundamental mechanism, the related literature is checked. Some researchers discovered that, under typical temperature settings (<20 °C), the rates of anaerobic phosphorus release and aerobic phosphorus uptake rise with increasing temperature. However, the phosphorus release rate increases while the phosphorus uptake rate of polyphosphate bacteria constantly declines in the 20–30 °C range. An explanation is given that organisms that collect glycogen grow in number while organisms that accumulate polyphosphate decrease as temperatures climb over 20 °C, and then, phosphorus is removed inefficiently (Fei et al. 2022). Accordingly, the decrease in TP breakdown efficiency with temperature elevation is speculated to relate to the effect of temperature on the capabilities of microbial communities for phosphorus uptake and release.

In summary, temperature variations outside the system can complexly impact sewage treatment efficiency in various domains (chemical, physical, and biological processes). More research is needed to understand the complex mechanisms driving such effects fully. In order to develop methods that maximize efficiency across a range of environmental conditions and water quality situations, it is critical to comprehend how climate change affects the complexities of sewage treatment systems. It is suggested that the following research focus on deciphering the mechanisms that underlie these findings, providing a stronger theoretical framework for understanding how temperature variations impact the treatment process's overall effectiveness.

The results of this investigation suggest that the A₂/O process exhibits improved COD and BOD degradation efficiency as temperature rises. However, the rate of increase in degrading efficiency decreases when the temperature rises above 2 °C. Additionally, the temperature increase improves the degradation efficiency for FI-SS, FI-(NH₃-N), and FI-TN. Simultaneously, when temperature rises, the A₂/O process's effectiveness in TP degradation steadily declines. Furthermore, this study's use of neural network models has shown to be highly advantageous for evaluating and enhancing the efficacy of sewage treatment procedures, which will aid in taking preventive measures against the impact of climate change on sewage treatment processes.

The research presented here is supported by China Key Technologies Research and Development Program 2021YFC1809104.

The contributions of the authors are delineated as follows: H.W. conceptualized the study and provided insights for the BP neural network analysis. S.C. and R.H. were responsible for data analysis and revising the manuscript. All authors participated in the review of the manuscript.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.