The need for sustainability within the ETES (sewage treatment stations), to take advantage of the residue and turn it into raw materials, depends on an important step which is the care with the anaerobic reactor variables. Only with care at the beginning of the sewage treatment process will it be possible to ensure a waste that can be used as a raw material for commercial use or in the station itself. To this end, based on monitoring data obtained in scientific work, convolutional neural networks were trained to model and indicate that the use of artificial intelligence in domestic sewage treatment stations was feasible. Data obtained from two scientific papers were grouped in the same ETE, in the same temporal lapse, the data were refined and normalized. In this article, RNA was applied to predict the methane flow (CH4), the model has good estimates for the data sets. The results obtained show that the developed model provides reliable estimates with response variables and shows that the most effective architecture of forecasts was using convolutional networks with a hidden layer with 8 neurons and sigmoid activation function, this configuration was obtained in the greatest convergence in Prediction of methane flow.
Research Article
Convolutional Neural Network Applied in Anaerobic Reactors of Domestic Sewage
https://doi.org/10.21203/rs.3.rs-4202088/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
Artificial intelligence
Biogas
modelling
Sewage treatment station
UASB.
1) Application of convolutional neural networks to real domestic sewage data.
2) Assistance in decision-making for sanitation companies through artificial intelligence and computational models.
Artificial intelligence (IA), since the decade 1980 has been presenting notorious advances, machine learning, and artificial neural networks, have been applied in many different areas like agriculture, engineering, education, medicine, environment and many other subjects. The number of articles related to IA, increased more than 600% lately (Crew 2020), and the advances are due to the capacity that is related to how to deal with more complex problems, uncertainty, and more dynamics.
Ye, at al, (2019) tell the AI technology in environmental management, presenting many works that prove the viability in prediction and in online learning, in addition to the tendencies in the use of this technology to substitute the math methods due to the trust replies received as a result and in a faster way than the traditional way. (CHOI & PARK, 2001) prove in their studies that the linear model is inadequate to describe the relation in the residual water treatment plants.
Canete, at al. (2016) present a study that ratifies the use of artificial intelligence, RNA, fuzzy systems, and genetic algorithms in Water treatment plants in residual water contributing to better performance in the plant, diminishing the courses of operation and improving the quality in the effluent.
Sanitation is a societal concern and day by day there are more requirements and regulations regarding its efficiency and the treatment of waste, on the other hand, sanitation is considered a complex system to be modeled, especially because it treats itself as a system with a physical complex process, biological and chemistry; moreover, the uncertainty stemming from the entries, and tributaries, is not controlled. (HAMED, KHALAFALLAH e HASSANIEN, 2003).
Most of the modern monitoring systems are offline systems; namely, it is done by the tributary collection and measurement of the other varieties and, finally, the analyses, which make uncertain the decisions taken in real-time and the correction of occasional problems during the treatment. In this way, it is possible to inform that making the right decision at the right moment is fundamental for the reduction of the impact on the environment ((MORENO-ALFONSO and REDONDO, 2001) and biogas production with high levels of methane (CH4). This component provides the necessary calorific for its energy utilization in sewage treatment.
Implementing an online measurement system is an expensive and demanding and requires a qualified team (BROEKE, CARPENTIER, et al., 2014). This is because the environment in a treatment station is extremely hostile to electronic sensors due to the presence of gasses and acids, which are highly corrosive and extremely detrimental to electronic equipment, potentially leading to their oxidation. In this regard, this article aims to present a real-time decision-making tool with a web-based interface based on a computational model that utilizes convolutional neural networks trained with data obtained based on the wastewater treatment plant database containing the necessary parameters for the project's development. Among the options explored, the best choice, due to the quality and quantity of data collection equipment, was the Padilha Sul WWTP (PAULA, 2019) and (HERNANDEZ, 2019). The model has four input layers represented by the physicochemical variables of the influent of the sewage reactor: flow rate (L/s), chemical oxygen demand (COD) (mg/L), total suspended solids (TSS) (mg/L), and volatile suspended solids (VSS) (mg/L). The model allows the operator of the wastewater treatment plant to make decisions before the effluent is discharged into the receiving bodies, and it also allows for adjustments in the process to obtain high-quality byproducts such as biogas, contributing to the sustainability of wastewater treatment plants.
(HUANG and CHEN, 2015), (MJALLI, AL-ASHEH, and E.ALFADALA, 2007), (HOLUBAR, ZANI, et al., 2002), among others, have applied neural networks to treatment plants and achieved success, which motivated this research to apply convolutional neural networks in the modeling of domestic sewage anaerobic reactors. Since the commercialization of byproducts generated in wastewater treatment plants is linked to their quality and their potential for energy generation, making a wastewater treatment plant sustainable. When reusing their byproducts, it is necessary to ensure that this product has good quality and performs as well as its competitors. After all, studies suggest that it is possible to market wastewater treatment plant biogas (Possetti, et al., 2018), (Ministry of Cities, 2015). Biogas, after a purification process, is referred to as biomethane, with a higher calorific value. Depending on the purification process, which involves removing all hydrogen sulfide (H2S), carbon dioxide (CO2), and water vapor, the efficiency can reach up to 99%, achieving calorific value comparable to gasoline, for example (WELLINGER, MURPHY, and BAXTER, 2013).
The analysis conducted for this article was structured with data obtained from two previously published scientific studies conducted at the same plant during the same period. Those showing a mathematical correlation with biogas production were selected among the physicochemical variables studied. Understanding the reactor's behavior and having the ability to be proactive, making decisions as soon as the influent arrives, ensures the quality of biogas, reduces the manager's range of concerns, and enables more effective and timely identification and monitoring of issues during the process. This, in turn, leads to the production of a high-quality, marketable, and competitive product. The capacity for innovation and effective wastewater treatment plant management tends to gain credibility due to (i) sustainability, (ii) financial returns, (iii) ecological responsibility, and (iv) the need to properly manage byproducts generated in domestic sewage treatment.
The European Union has been seeking ways to reduce its energy dependence on Russian imports and supporting the expansion of biogases. There is a forecasted annual growth of 35%, with an estimate that by 2050, the production of over 1 billion cubic meters of low-emission gases, such as hydrogen, biogases, and synthetic methane, will be achieved. This would correspond to nearly one-third of the total demand for gaseous fuels consumed (INTERNATIONAL ENERGY AGENCY, 2022). This continuous effort to explore renewable energy alternatives requires developing new techniques. Computational modeling is already employed in the environmental field; however, the techniques used are distinct, with each implementation encompassing specific characteristics of the authors. Control of physicochemical variables in wastewater treatment plants typically occurs through periodic measurement and laboratory analysis campaigns. The data used in this research, comes from the implementation and monitoring of a Wastewater Treatment Plant (ETE) that participated in a pilot project, the result of a partnership between the Brazilian and German governments aimed at deploying imported electronic sensors and in-depth laboratory analyses. This study has borne fruit in scientific works such as dissertations and theses, where all precautions were taken, such as calibrations and sensor validations, as well as comparisons with other equipment, all validated in scientific research and obtained through equipment designed for measuring wastewater treatment plants, have been calibrated, verified, and validated in scientific works (PAULA, 2019) and (HERNANDEZ, 2019). And the data obtained from these studies served as the basis for the implementation of the artificial intelligence techniques outlined in this article.
Sustainability and the flow of materials in wastewater treatment plants (ETEs) are recurring topics in sector meetings and conferences. The search for a solution and an ideal disposal method for byproducts generated from wastewater treatment is widely discussed in the literature, highlighting the concern with this issue. Therefore, the study conducted in this article contributes to predicting the outcome of the output variable, which is the biogas flow generated in the wastewater treatment plant through domestic sewage treatment. Understanding this correlation, the manager can establish thresholds and monitor the variables to achieve a high-quality product that can be used for energy generation within the plant or commercialized rather than being burned outdoors, as is the case in most ETEs throughout Brazil. This outdoor burning system is inefficient due to atmospheric conditions and consequently has an efficiency rate of between 10–50% in methane destruction (KAMINSKI, POSSETTI, et al., 2021). This leads to limited reductions in greenhouse gas emissions, making the use of this byproduct, even in its early stages, in line with the Sustainable Development Goals (SDGs). Such energy utilization is already taking place in the country (POSSETTI, RIETOW, et al., 2019).
Understanding the operation of the domestic sewage treatment process is essential to make wastewater treatment plants (ETEs) sustainable. There is a need for proactive management, rather than the reactive approach that is common in most cases, where knowledge of the product and the quality of the biogas produced is only gained at the end of the process, through gas chromatography testing or when it passes through some dedicated electronic sensor. At that point, there is nothing more that can be done to improve its quality because it cannot be returned to the reactor. In cases like this, disposal occurs through the flaring process, reducing pollution emissions by 32 times. This process converts methane into carbon dioxide through combustion and is considered environmentally friendly. However, as mentioned earlier, it has low efficiency in reducing greenhouse gas emissions. The use of artificial intelligence techniques in wastewater treatment plants not only brings economic benefits through the reuse of a high-calorific-value product that is currently burned and wasted but also promotes environmental gains by reducing pollutant emissions and aligns with the circular economy and environmental sustainability policy.
This article is organized as follows: it provides a theoretical foundation on domestic sewage treatment and neural networks, as they are the basis of the work. It then details the materials and methods, followed by the results and discussions. Finally, the article concludes with insights on the study.
Basic sanitation, domestic sewage, and public health have a direct connection to the perspective of social and economic development. In 2020, the legal framework for sanitation in Brazil was updated (Brazil, 2020). As a result, basic sanitation in Brazil is now open to private investments and should be accessible to all. Currently, half of the Brazilian population lacks access to the sewage system, and the new legal framework aims to address these shortcomings.
Wastewater treatment plants (ETEs) are designed to treat domestic sewage, and they are intended to remove organic and inorganic compounds. In other words, they separate the solid part from the liquid part and treat both to meet the quality parameters and limits for effluent discharge into water bodies (Jordão & Pessoa, 2011). The treatment process, as shown in Fig. 01, involves three treatment levels: preliminary, primary, and secondary, which consist of a series of steps until the effluent reaches the required level of purity as mandated by legislation.
In Latin American countries, due to climate and economic considerations, there is a preference for anaerobic technology in domestic sewage treatment (CHERNICHARO, RIBEIRO, et al., 2018), (SPERLING, 2016). This preference is because anaerobic reactors require fewer financial resources for implementation and maintenance compared to aerobic reactors. In Brazil, there is a leadership role in the use of anaerobic technology (SPERLING, 2016). All simulations included in this study are based on anaerobic treatment systems.
Water, sewage, and other substances possess physical, chemical, and biological characteristics. Water for human consumption must meet certain standards as per legislation (Ministry of Health, 2020), and the effluent discharged after treatment into a water body must also meet specified criteria. These characteristics can be altered by natural, microbiological, physical, chemical, and radioactive conditions. Therefore, it is essential to analyze these variables to monitor treatment efficiency and to determine the need for preventive or corrective interventions
Domestic and industrial sewage have distinct compositions due to their differing origins, resulting in heterogeneous compositions. Their chemical compositions are different, which has a direct influence on the constitution of the generated biogas. In this article, only domestic sewage was considered. The variables used for training the neural networks in this study are described in Table 01.
|
Parameters |
Description |
|---|---|
|
Sewage Flow |
Flow rate of sewage over a period. |
|
COD |
Chemical Oxygen Demand, an index of oxygen required to chemically stabilize organic matter. |
|
TSS |
Total Suspended Solids, including both organic and inorganic undissolved solids. |
|
VSS |
Volatile Suspended Solids, the portion of suspended or dissolved solids. |
During sewage treatment, some byproducts are produced, and one of them is biogas generated by the action of fermenting bacteria in the anaerobic reactor that processes organic matter and generates biogas. Biogas is a mixture of gases and can be used as a fuel, but it is still underutilized. In most cases, it is flared outdoors, which is a waste of a byproduct that can have a high calorific value (8,570 kcal/m³) (Constant, et al., 1989), and it could be commercialized. The use of this byproduct not only brings economic benefits but also promotes environmental gains and aligns with the circular economy policy.
For biogas to have a high calorific value, a high concentration of methane gas (CH4) is required, which forms through an ideal chemical process in the anaerobic reactor. Identifying this ideal method is the goal of station managers because harnessing biogas reduces the consumption of natural resources and contributes to ecosystem balance, as waste becomes a raw material for energy production. To achieve economic viability in biogas utilization, it's essential to have a controlled composition with a high methane gas content to ensure a substantial calorific value.
For biogas to have a high calorific value, a high concentration of methane gas (CH4) is required, which forms through an ideal chemical process in the anaerobic reactor. Identifying this ideal method is the goal of station managers because utilizing biogas reduces the consumption of natural resources and contributes to ecosystem balance when waste becomes a raw material for energy production. Economic viability in biogas utilization necessitates controlled composition with a high methane content to ensure a substantial calorific value.
In general, compatible software tools are required to implement artificial intelligence. Among them, the most commonly used languages include C/C++, Java, and Python, among various other programming languages (Hentenryck, 2002). One of the languages that has been growing rapidly is Python, primarily due to its ease of use and open-source software license (NYLEN and WALLISCH, 2017).
Artificial neural networks (ANN) can be considered as an ideal nonlinear heuristic model for making predictions and data classification based on the performance of the network's inputs and outputs (HAYKIN, 2001). This technology takes a leading role (Mjalli, Al-Asheh, & E.Alfadala, 2007), (CHOI & PARK, 2001), (HAMED, KHALAFALLAH, and HASSANIEN, 2003) in the combination of artificial intelligence with treatment plants. Most of the research employs either a single NN model or a hybrid one, including a fuzzy neural network, with both techniques yielding good results.
Artificial neural networks were inspired by the central nervous system of higher organisms, and their operating principle enables them to recognize data patterns through training learning to make generalizations based on acquired knowledge (HAYKIN, 2001). Because of this, artificial neural networks (ANNs) are often used for prediction in the environmental field, and their structure can be divided into three parts: the input layer, the hidden or intermediate layer, and the output layer. The first layer corresponds to the input variables of the system, and the last layer represents the expected output variable. The hidden layers are between (LAUWERS, APPELS, et al., 2013). When the network has one or more hidden layers between the input and output layers, it is considered a Multilayer Perceptron (MLP).
Neural networks are composed of artificial neurons capable of performing mathematical functions, which are often nonlinear. Therefore, due to their complexity and nonlinear behavior, they are suitable for adequately modeling biological processes, such as wastewater treatment. (SILVA, SPATTI, and FLAUZINO, 2010) define an artificial neuron as a simplified model of biological neurons. These artificial neurons are generally nonlinear and have the function of collecting input values and producing a response based on their activation function.
The activation function is necessary to introduce nonlinearity into the network. Without this function, hidden layers do not become as powerful. In the works covered in this study, most authors used the sigmoid activation function, with variations such as the tan-sigmoid. These are generally the preferred activation functions over threshold activation functions, as threshold functions are difficult to train due to the gradient not existing or being zero, making it impossible to use backpropagation techniques or more efficient methods (KUSIAK and WEI, 2012). The sigmoid activation function was used because it describes nonlinear relationships, such as in biogas production (KANAT and SARAL, 2009). It is also present in most references and endorsed by (HAYKIN, 2009) in his book as the most common type of applied activation function.
Another relevant issue is related to the periodic behavior observed in the data. As can be seen in Fig. 2(a), it was identified that the biogas flow has a behavior similar to the COD with an approximately 3-hour delay. This factor was also observed when comparing sewage flow to biogas flow, which exhibits a periodic behavior of 1 hour, as shown in Fig. 2(b). This periodic behavior observed in these relationships motivated the study of convolutional neural networks.
In 1998, a new type of network was developed, which would later be The Convolutional Neural Network (CNN), an architecture using back proto languages forward network with many hidden layers, showing excel proliferating image processing. CNNs are deep neural networks, meaning they have multiple hidden layers with interconnected neurons and are widely used for image recognition and visual tasks. However, very linear so being used for text analysis and various other activities (DATA SCIENCE ACADEMY, 2022). Because they are fully connected networks between layers, they yield satisfactory training re et all criteria to explore combining during training (LECUN, BOTTOU, et al., 1998).
Deep learning has become widely known for popular computer vision experiments, such as autonomous cars, robotics, drones, medical diagnostics, and other modern popular activities.
Time series data are collections of observations occurring over time, implying a temporal relationship between at least one data point in the series and its predecessors, meaning there is a temporal connection, with one element having a relationship with its previous elements. Time-delay neural networks (TDNN) are also known as one-dimensional convolutional neural networks (CNN) (POVEY, CHENG, et al., 2018). The definition of the number of delays, i.e., how many previous data points impact the current data point, is flexible and should be adjusted for each dataset. In these systems, data is no longer a set of independent samples and becomes functions of time.
Once the operation of time series is understood, it is essential to realize that based on this information, it is possible to model various natural phenomena with the aim of predicting future behaviors and understanding the operation of such processes (MORETTIN and TOLOI, 2006).
According to (HAYKIN, 1998), incorporating time into a neural network can be done through an implicit representation. For example, the input signal is expressed in the same way, and the sequence of weights for each neuron connected to the input layer is convoluted with a different input sequence, thereby incorporating the time series of the signal into the network structure.
Another relevant point is the issue of seasonal phenomena, which occur regularly at certain times. In treatment plants, this phenomenon is closely related to population behavior, for example, bathing times, or in other words, the habits and times when people tend to release material into the sewage collection network and the time it takes for this material to reach the treatment plant, as well as the configuration and structure of the pipelines and networks (HERNANDEZ, 2019).
The data in this article was obtained through the digital library of the theses and dissertations of the Federal University of Paraná - UFPR. Both studies referenced (HERNANDEZ, 2019) and (PAULA, 2019) focused on the same subject of study within the same Wastewater Treatment Plant (WWTP) located in the city of Curitiba, State of Paraná, and the research took place during the same period. Based on the available data, the model included the use of CNN-based algorithms with the ability to predict results regarding a biological system, using the following physical-chemical parameters as input variables: sewage flow rate (L/s), COD (Chemical Oxygen Demand, mg/L), TSS (Total Suspended Solids, mg/L), and VSS (Volatile Suspended Solids, mg/L), aiming to predict the biogas flow rate (Nm³/h), as controlling this output variable ensures good production.
(PAULA, 2019) conducted a Spearman and Kendall Tau correlation analysis to examine the relationship between the variables in the liquid phase of the influent and the gaseous phase (biogas production). The aim was to assess the relationship between the biogas produced and the physical-chemical input parameters of the system. It was found that among the solids, VSS showed a strong positive correlation, TSS exhibited a moderate correlation, while SSF had a weak or very weak correlation. It indicates that SSF is unsuitable for predicting biogas production, in addition to the fact that SSF represents the inert fraction of domestic sewage, which is irrelevant as a substrate during anaerobic digestion.
The researched Wastewater Treatment Plant (WWTP) features six modified Upflow Anaerobic Sludge Blanket (UASB) reactors, is considered medium-sized, with a design flow rate of 440 L/s, serving a population of approximately 252,764 inhabitants and its hydraulic retention time is approximately 8 hours (Ross, 2015). In the operational phases, there is a preliminary treatment involving two bar screens, 10 cm and 5 cm in size, a 6 mm mechanical bar screen, a grit chamber, and a Parshall flume.
The database was formatted and normalized to enable the programming language to capture the information so that the data adhered to a range adjusted for orders of magnitude, making all values fall between zero and one. This normalization process allowed for training of the neural network and then validating and testing the results obtained.
The obtained variables refer to the data from the influent and effluent in the liquid phase of the anaerobic treatment process and the gas phase, which corresponds to the final stage of the anaerobic treatment process. The data from the liquid phase of the influent and the gas phase were measured by electronic equipment, already properly calibrated and with their uncertainty analyses studied in the work of (HERNANDEZ, 2019) hourly, obtaining 24 samples per day for 3 consecutive days, for each of the 5 months analyzed. As for the effluent data, laboratory analyses were conducted three times a day: at 8 a.m., 12 p.m., and 4 p.m, for the same three consecutive days in the 5 months of the research.
The quantity of collected data does not represent a limitation for the application of computational models, as can be observed in successful cases (HAMED, KHALAFALLAH, and HASSANIEN, 2003), (CANETE, SAZ-OROZCO, et al., 2016), (CHOI and PARK, 2001), (MJALLI, AL-ASHEH, and E.ALFADALA, 2007), (SAKIEWICZ, PIOTROWSKI, et al., 2020), (ASADI, GUO, and MCPHEDRAN, 2020), (KUSIAK and WEI, 2012), (ALSULAILI and REFAIE, 2021), (NOURANI, ELKIRAN, and ABBA, 2018), and (HEJABI, SAGHEBIAN, et al., 2021). In many cases, the quantity of data is even substantial, but due to measurement problems, the data is incomplete or uncalibrated, leading to their exclusion as they do not represent the actual situation. The benefit of reducing data for modeling is achieving more efficient and enhanced modeling.
In all the studies mentioned in the previous paragraph, regardless of the quantity of available data, predictions were successfully made using AI tools, and the results were satisfactory. Alternative methods can be used to generate more input data for the models. In cases where various data is missing, appropriate statistical methods, such as data interpolation and extrapolation, can be used to fill in the missing data and then use the synthetic data in modeling. However, in this case, the synthetic data can introduce a higher error level, and using such data may not lead to appropriate results (NOURANI, ELKIRAN, and ABBA, 2018). This is because the period that was not collected during the day regarding the effluent consists of 21 missing samples per day. Data interpolation and extrapolation could solve this problem by generating data for that period. However, in the case of applying these techniques, generating data from only three daily samples would not provide the expected confidence in the model, which led to the exclusion of effluent information from the development of the model.
Among the different types of available networks, the feed-forward-backpropagation (FFBP) network is the favorite among authors because it is considered the most suitable for modeling anaerobic digestion (HOLUBAR, ZANI, et al., 2002). This is due to its ability to correct the weights applied to the layers.
In their research, HAMED, KHALAFALLAH, and HASSANIEN (2003) used BOD and the concentration of suspended solids as the basis for their ANN model. They found that these two parameters were sufficient for modeling, as they are considered good indicators of ETP (Effluent Treatment Plant) performance, according to Droste (2009).
From the data collected hourly, various parameters were obtained, including the collection time, sewage flow rate, COD (Chemical Oxygen Demand), organic load, TSS (Total Suspended Solids), SSF (Fixed Suspended Solids), VSS (Volatile Suspended Solids), sewage flow rate, CH4, CO2, O2, and H2S content in the biogas. Other data that were collected only three times a day were not used in the analysis. To mitigate the sensitivity of the ANN and improve its prediction capacity, the number of parameters applied to the network was reduced to four: sewage flow rate, COD, TSS, and VSS, as shown in Fig. 4. This optimization of the model helps avoid overfitting by eliminating unnecessary data that would introduce noise into the model.
According to (PAULA, 2019) and (HERNANDEZ, 2019), leaks in the biogas lines were identified during a certain period of data collection. Therefore, in this article, on readings after the correction of the identified leak in the WWTP was considered.
To evaluate the model's performance, the Pearson correlation, which is used to calculate the correlation between two data series, and the coefficient of determination (R2), a squared version of Pearson, commonly used to measure model efficiency, were used, as well as the Mean Squared Error (MSE). The R2 indicates the degree of correlation between experimental and predictive values, which can be represented by the equation present in Eq. 1.
Dessa forma obtendo um maior valor de R2 e um valor de MSE tendendo a zero, significa um melhor desempenho durante a predição.
Os resultados apresentados na próxima sessão referem-se a uma média calculada com o valor obtido de cada um dos 10 treinamentos em sequência realizados, para o índice de correlação R2 e valor do MSE.
In this way obtaining a more considerable value to R2 and a value to MSE tending to zero, it means a better performance during the prediction.
The results showed in the next session refer to a calculated media with the value obtained in each one of the ten trainings realized in sequence for the indices of correlation R2 and value of MSE.
Due to the temporal nature of the biological sewage treatment process and the periodic behavior of the data, we opted for the application of a neural network model capable of handling this type of situation, which motivated the use of convolutional neural networks. The hidden layer configuration was selected after a supervised learning process based on the application of various attempts and error analysis, obtaining the best results and following the methodologies from the Xu & Ho (2006) bibliography.
Until now, the use of convolutional neural networks for prediction in WWTPs has not been observed. It is one of the networks that has been growing in its utilization for applying more advanced techniques, especially in treating periodic behaviors and images, which motivated its application and the comparison of results.
A common practice adopted in most of the articles in this review is the trial and error technique to identify the optimal number of hidden neurons (HAMED, KHALAFALLAH, and HASSANIEN, 2003), (CANETE, SAZ-OROZCO, et al., 2016), (ALSULAILI and REFAIE, 2021), (NOURANI, ELKIRAN, and ABBA, 2018). However, an inadequate number of hidden neurons will limit the neural network's ability to model the process. Furthermore, an excessive number of hidden neurons can provide redundant freedom for weight adjustment and result in the propagation of noise present in the database to the model (LINKO, ZHU, and LINKO, 1999). Therefore, it's better to apply various models with different configurations of hidden neurons and then perform a thorough analysis of the model's error to identify which one provides the best prediction of the measured variable, considering the closest approximation and the lowest error rate.
The trained neural network had four neurons in the input layer, corresponding to: sewage flow, COD, TSS, and VSS. The number of neurons and hidden layers was determined through supervised learning techniques. This involved varying the number of neurons in the hidden layer incrementally to find the model that yielded the best prediction results with the smallest errors. The number of output parameters was set to one, and the number of neurons in the hidden layer ranged from 2 to 16.
In the training phase, the number of neurons in the hidden layer was adjusted, starting with 2 and exponentially increasing to 16 neurons. Few neurons in the hidden layer can lead to underfitting, meaning a fit that is below expectations, while an excessive number of neurons can cause overfitting, which results in excessive fitting and imprecise predictions. As shown in Table 1, the ideal adjustment was achieved with 8 neurons. The average value obtained after 10 simulations was 0.93, with a maximum value of 0.96 and a minimum of 0.90.
| RNA C (2) | RNA C (4) | RNA C (8) | RNA C (16) | |
|---|---|---|---|---|
| MSE | 8.21E− 04 | 1.54E− 03 | 2.08E− 04 | 8.06E− 06 |
| Average R2 | 0.77 | 0.82 | 0.93 | 0.32 |
A scatter plot was constructed to analyze the coefficient of determination R2, which was calculated to assess the variations between the inputs (red line) and outputs (blue points) in order to identify the best combinations. The R2 value was 0.93, as shown in Fig. 5, indicating that the presented model can explain approximately 93% of the total variance in the data.
In order to verify if 8 neurons were indeed the best option, simulations were conducted with 7 and 9 neurons. The results obtained were, for 7 neurons, an average R2 of 0.54 and an MSE of 7.32E-04, and for 9 neurons, an average R2 of 0.73 and an MSE of 1.49E-03. This demonstrates that the simulation with 8 neurons yielded the best average results.
In Fig. 6, you can observe the proximity of the lines in the graph between the predicted data and the real data. The values are plotted in a non-normalized way to facilitate the identification of real values with greater disparity, as in the normalized plot, the values are adjusted to scale. Identifying the points that deviated, an attempt was made to find any abnormal behavior in the data to justify the lower accuracy of the prediction. It was identified that peaks that deviated from the data pattern were cases of less accurate prediction.
Figure 7 shows the graph that indicates the loss in training data and the loss in the validation set during the trained epochs of the model, acceptable indices that support the other results presented.
(HAMED, KHALAFALLAH, and HASSANIEN, 2003) in their study using neural networks to predict the performance of effluents in wastewater treatment plants in Cairo, using the backpropagation and feed-forward algorithms, obtained R2 values ranging from 0.45 to 0.81. Meanwhile, (CANETE, SAZ-OROZCO, et al., 2016) achieved R2 values between 0.88 and 0.92 with their multilayer perceptron neural network for effluent prediction. (HOLUBAR, ZANI, et al., 2002) applied various neural networks to predict methane production using feedforward-back networks in a treatment plant. They obtained the lowest R2 value when simulating pH with a value of 0.82 and the best result when predicting biogas production with a value of 0.90. Other authors such as (KANAT and SARAL, 2009), (AKBAS, BILGEN, and TURHAN, 2015) also achieved excellent results in their predictions, although inferior to those obtained with convolutional networks.
Parameterizing the proposed model in a language with logical inputs similar to human expert thinking tends to be a tool with potential for use. Therefore, the expected result is the widespread use of this tool in wastewater treatment plants. To make the model more accessible and user-friendly, a web-based user interface was developed. Users can input data and instantly obtain the predicted biogas flow, even if they lack programming language or technical skills.
The tool is accessible on multiple platforms as a web page. Users can input data, and in real-time, they receive the predicted biogas flow result based on the pre-configured model, as seen in Fig. 8.
The application runs on a local web server, using the Uvicorn platform that runs on Python. The development of the pages and the form was done using HTML, with a web application using a Python web development framework called FastAPI, all of it running on the selected and trained model.
Developing applications like this, integrated with the technological transition of automated data collection in treatment plants, helps avoid the need for post-analysis by a specialist and provides real-time information on production after the treatment process. Automation through electronic sensors for input variables and computational analysis by a properly trained computational model signifies a significant technological transformation in managing treatment plants and monitoring quality indicators. Efficiently managing this transition is challenging for managers and government agencies seeking sustainability and environmentally responsible management. This transition represents a technological and mindset barrier as it demands more automation in the process, infrastructure investments, and trained personnel familiar with technology beyond traditional sanitation methods, which can still be unfamiliar to many technicians working at these plants. This shift takes one out of their comfort zone, bringing new challenges and more complex technical issues.
The primary barrier to adopting this technology is the influence of biological behavior, which varies in Latin America due to climate differences. Therefore, the model would need to be calibrated and adapted to the climate of each region, which would require data collected from the treatment plants to feed the model's training process.
Treatment plant managers, who are responsible for decision-making, should consider this study as a valuable resource to gain an overview of technology adoption in the treatment process. It aims to facilitate proper environmental management, promote the circular economy, and contribute to achieving the Sustainable Development Goals (SDGs) by reducing the environmental impact of climate change by 2030.
In this article, Convolutional Neural Networks (CNN) were applied to predict methane (CH4) flow rates using data collected from a plant in Curitiba, Paraná, Brazil, in July to November 2018. However, only data from months without leaks were used. The model provides good estimates for the datasets. The predictions made had acceptable errors, but it is important to note the geographical aspect for data validation when considering the tool's potential use in other regions.
Predicting flow rates in treatment stations is important, especially with the new legal framework for sanitation approved, which may lead to an increased load on these stations. After all the analyses performed in the development of this work, preliminary tests were conducted to assess the feasibility of applying a computational model to the obtained data to support the credibility of the potential for renewable energy generation in treatment Plants.
The Convolutional Neural Networks model with one hidden layer, eight neurons, and a sigmoid activation function showed the highest convergence in predicting methane flow rates in domestic wastewater anaerobic reactors. Finding error rates trending towards zero and convergence rates close to 100% indicates consistent results from the modeling, even in cases where some deviations occurred. This provides credibility to the results of biogas production.
The potential adoption of this more accurate model by sanitation companies could lead to a transition from occasional campaigns to the use of a computational model based on neural networks. This could be a cost-effective way for governments and state-owned sanitation companies to make treatment plant more sustainable.
After the analysis, this study's contribution is seen as an incentive for improving the processes in these reactors, allowing for greater efficiency in treatment plant by allowing operators to make informed decisions. The goal is to use models in the station and potentially develop methods for adjusting variables whenever a deficit is detected that impacts biogas production. Automation of decision-making processes through electronic or mechanical means may be possible to maintain economically viable biogas production with the highest possible methane content and better calorific value, around 6 kWh/m3.
Machine learning is an emerging technology and is becoming increasingly popular. However, each model and architecture has its advantages and disadvantages in the modeling process. It remains an idea for future work to integrate models, mitigate the weaknesses of each individual model, and arrive at the best-performing model.
In both scenarios and the networks developed, the principle of parsimony to optimize the capacity of the ANNs was maintained. This principle states that simpler models with fewer parameters are a more appropriate choice than more complex models with more parameters (Lawrence, Giles, & Tsoi, 1996).
The application of artificial intelligence techniques, including neural networks, is a valuable tool for the operational management of treatment plants. They are cost-effective and allow efficient prediction of the physical-chemical variables involved in the process. This contributes to the strengthening of sustainable treatment plants and supports the transformation of post-treatment by-products into marketable products or efficient self-consumption, leading to significant environmental gains.
Funding
The authors declare that they have received no funding.
Availability of data and materials
The datasets used and/or analyzed during the current study are available from
the corresponding author upon reasonable request.
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable
Competing interests
The authors declare that they have no competing interests.
- AKBAS, H.; BILGEN, B.; TURHAN, A. M. An integrated prediction and optimization model of biogas production system at a wastewater treatment facility. Bioresource Technology, 2015.
- ALSULAILI, A.; REFAIE, A. Artificial neural network modeling approach for the prediction of five-day biological oxygen demand and wastewater treatment plant performance. Water Supply, 24 Agosto 2021. 1861–1877.
- ASADI, M.; GUO, H.; MCPHEDRAN, K. Biogas production estimation using data-driven approaches for cold region municipal wastewater anaerobic digestion. Journal of Environmental Management, 2020.
- BRASIL. LEI Nº 14.026. Diário Oficial da União, p. 135, 2020.
- BROEKE, J. V. D. et al. Compêndio de sensores e monitores e seu uso na indústria global de água. Relatório Final. [S.l.]. 2014.
- CANETE, J. F. D. et al. Soft-sensing estimation of plant effluent concentrations in a biological wastewater treatment plant using an optimal neural network. Expert Systems With applications, 2016. 8–19.
- CHERNICHARO, C. A. D. L. et al. Contribuição para o aprimoramento de projeto, construção e operação de reatores UASB aplicados ao tratamento de esgoto sanitário – Parte 1: Tópicos de Interesse. DAE, São Paulo, v. 66, n. 214, p. 5–16, November 2018. ISSN 0101–6040.
- CHOI, D.-J.; PARK, H. A hybrid artificial neural network as a wastewater treatment process. Water Research, November 2001. 3959–3967.
- CONSTANT, M. et al. Biogas end-use in the European community. [S.l.]. 1989.
- CREW, B. Artificial-intelligence research escalates amid calls for caution. Nature, v. 588, n. S101, 2020.
- DATA SCIENCE ACADEMY. Deep Learning Book. [S.l.]: ebook, v. 1, 2022.
- DROSTE, R. L. Theory and Practice of Water and Wastewater Treatment. India: Wiley, 2009.
- HAMED, M. M.; KHALAFALLAH, M. G.; HASSANIEN, E. A. Prediction of wastewater treatment plant performance using artificial neural networks. Environmental Modelling & Software, 15 October 2003. 919–928.
- HASSOUN, M. H. Fundamentals of Artificial Neural Networks. Cambridge: Bradford Books, 1995.
- HAYKIN, S. O. Neural Networks: A Comprehensive Foundation. United States Edition: Pearson, 1998.
- HAYKIN, S. S. Redes neurais: principios e prática. Porto Alegre: Bookman, 2001.
- HAYKIN, S. S. Neural Networks and Learning Machines. 3º. ed. USA: Prentice Hall, 2009. ISBN 978-0131471399.
- HEJABI, N. et al. Evaluation of the effluent quality parameters of wastewater treatment plant based on uncertainty analysis and post-processing approaches (case study). Water Science & Technology, 15 February 2021. 1633–1648.
- HERNANDEZ, O. A. D. Avaliação de sistemas de medição para controle de processo em tempo real numa Estação de Tratamento de Esgoto sanitário que utiliza reatores UASB. UFPR. Curitiba. 2019.
- HOLUBAR, P. et al. Advanced controlling of anaerobic digestion by means of hierarchical neural networks. Water Research, 2002.
- HUANG, Y. W.; CHEN, M. Q. Artificial neural network modeling of thin layer drying behavior of municipal sewage sludge. Measurement, 26 June 2015. 640–648.
- INTERNATIONAL ENERGY AGENCY. World Energy Outlook. INTERNATIONAL ENERGY AGENCY. 31 member countries, p. 524. 2022.
- JORDÃO, E. P.; PESSOA, C. A. Tratamento de Esgotos Domésticos. Rio de Janeiro: ABES, 2011.
- KAMINSKI, G. et al. Combustão direta de biogás em queimadores. Cadernos Técnicos Eng Sanit Ambient |, Rio de Janeiro, v. 1, n. 1, p. 94, 2021.
- KANAT, G.; SARAL, A. Estimation of Biogas Production Rate in a Thermophilic UASB Reactor Using Artificial Neural Networks. Environ Model Assess, 19 June 2009. 607–614.
- KUSIAK, A.; WEI, X. A data-driven model for maximization of methane production in a wastewater treatment plant. Water Science & Technology, 2012. 1116–1122.
- LAUWERS, J. et al. Mathematical modelling of anaerobic digestion of biomass and waste: Power and limitations. Progress in Energy and Combustion Science, v. 39, n. 4, p. 383–402, 2013. ISSN 0360–1285.
- LAWRENCE, S.; GILES, C. L.; TSOI, A. C. What Size Neural Network Gives Optimal Generalization? Convergence Properties of Backpropagation. Washington. 1996.
- LECUN, Y. et al. Gradient-based learning applied to document recognition. Proceedings of the IEEE, v. 86, p. 2278–2324, November 1998. ISSN DOI: 10.1109/5.726791.
- LINKO, S.; ZHU, Y.-H.; LINKO, P. Applying neural networks as software sensors for enzyme engineering. Trends in Biotechnology, 1 April 1999. 155–162.
- MINISTÉRIO DA SAÚDE. Portaria nº 2.914 de 11 de December de 2011. Ministério da Saúde, 13 August 2020. avaliable em: <http://bvsms.saude.gov.br/bvs/saudelegis/gm/2011/prt2914_12_12_2011.html>.
- MINISTÉRIO DAS CIDADES. Guia técnico de aproveitamento energético de biogás em estações de tratamento de esgoto / Probiogás. Brasília. 2015.
- MJALLI, F. S.; AL-ASHEH, S.; E. ALFADALA, H. Use of artificial neural network black-box modeling for the prediction of wastewater treatment plants performance. Journal of Environmental Management, May 2007. 329–338.
- MORENO-ALFONSO, N.; REDONDO, C. F.-A. Intelligent waste-water treatment with neural-networks. Water Policy, 28 Junho 2001. 276 – 271.
- MORETTIN, P. A.; TOLOI, C. Análise de séries temporais. São Paulo: Blucher, 2006.
- NOURANI, V.; ELKIRAN, G.; ABBA, S. I. Wastewater treatment plant performance analysis using artificial intelligence – an ensemble approach. Water Science & Technology, 10, 2018.
- NYLEN, E. L.; WALLISCH, P. Neural Data Science. [S.l.]: Academic Press, v. ISBN 9780128040430, 2017.
- ORTH, F. A. Slide Share. CORSAN, 2007. Avaliable em: <https://pt.slideshare.net/Joao256/tratamento-de-esgoto-doado-pela-corsan-ao-municpio-de-montenegro-em-2007>. Access em: 20 Abril 2018.
- PAULA, A. C. D. Avaliação integrada do desempenho de reatores anaeróbios do tipo UASB tratando esgoto doméstcco em escala real. UFPR. Curitiba. 2019.
- POSSETTI, G. R. C. et al. Contribution for improving the design, construction and operation of UASB reactors treating sewage – Part 5: Biogas and fugitive methane emissions. DAE, v. 66, n. 214, p. 73–89, 2018.
- POSSETTI, G. R. C. et al. Chapter 8: Energy recovery from biogas in UASB reactors treating sewage. In: CHERNICHARO, C. A. D. L.; BRESSANI-RIBEIRO, T. Anaerobic Reactors for Sewage Treatment: Design, Construction and Operation. [S.l.]: IWA Publishing, 2019. Cap. 8.
- POVEY, D. et al. Semi-Orthogonal Low-Rank Matrix Factorization for Deep Neural Networks. Interspeech. [S.l.]: [s.n.]. 2018.
- ROSS, B. Z. L. Escuma de reatores anaeróbios tratando esgotos domésticos em escala real: produção, caracterização e proposição de parâmetros para seu gerenciamento. Curitiba. 2015.
- SAKIEWICZ, P. et al. Innovative artificial neural network approach for integrated biogas wastewater treatment system modelling: Effect of plant operating parameters on process intensification. Renewable and Sustainable Energy Reviews, 20 Fevereiro 2020.
- SILVA, I. N. D.; SPATTI, D. H.; FLAUZINO, R. A. Redes neurais artificiais para engenharia e ciências aplicadas - Curso pratico. São Paulo: Artliber, 2010.
- SPERLING, M. V. Urban Wastewater Treatment in Brazil. Inter-American Development Bank, 2016.
- WELLINGER, A.; MURPHY, J.; BAXTER, D. The Biogas Handbook: Science, Production and Applications. 52. ed. Philadelphia: Elsevier Science, 2013.
- XU, J.; HO, D. W. C. A Node Pruning Algorithm Based on Optimal Brain Surgeon for Feedforward Neural Networks. Advances in Neural Networks, 2006. 524–529.
- YE, Z. et al. Tackling environmental challenges in pollution controls using artificial intelligence: A review. Science of the Total Environment, 5 September 2019. 28.
No competing interests reported.