Electricity generation from renewable source- A leachate based microbial fuel cell machine learning approach

doi:10.21203/rs.3.rs-5119060/v1

Download PDF

Research Article

Electricity generation from renewable source- A leachate based microbial fuel cell machine learning approach

https://doi.org/10.21203/rs.3.rs-5119060/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

This research presents an innovative blend of feature selection and sensitivity analysis techniques, which is an essential yet overlooked aspect in the study of MFCs. The study compared predictive models utilizing various machine learning algorithms to assess the impact of time, dosage, pH and temperature on ammonium nitrogen concentration (NH₄-N) to predict the power density (PD) output of microbial fuel cells using leachate as a substrate for treatment. Evaluation of six machine learning models demonstrates varying levels of predictive accuracy. CatBoost (R2:0.9969, MSE: 48.8430, RMSE:6.9888) emerges as the most accurate model, followed closely by XGBoost (R2:0.9917, MSE:130.1668, RMSE:11.4091) and Random Forest (R2:0.9830, MSE:267.0929, RMSE:16.3430). Time series plots illustrate the performance of different models in predicting PD over a period, indicating good alignment with observed data. Comparison of Mean Squared Error (MSE) highlights significant variations in prediction accuracy, with CatBoost demonstrating the greatest enhancement and precision. The study directly tackles the deficiencies in existing MFC predictive modeling by incorporating the CatBoost algorithm, which provides enhanced accuracy and a deeper understanding of the nonlinear connections between environmental variables and power density.

Leachate landfill

ammonium nitrogen

Ensemble Algorithms

power density

Microbial fuel cells

The exploration of alternate sources of energy has become even more important over the past decade due to the growing concerns regarding the degradation of the environment and the depletion of traditional resources of fossil fuels (Naha et al., 2023). According to the International Energy Agency (IEA), there is an increase of 18% in carbon footprint which is deleterious to the environment (Tian et al., 2023). This is because energy sources are becoming increasingly scarce. Microbiological fuel cells, also known as MFCs, have emerged as a viable contender for the generation of sustainable energy among the wide variety of technologies that are used to generate renewable energy (Naha et al., 2023). Microbial fuel cells (MFCs) represent an environmentally friendly and durable solution harnessing microorganisms to generate electricity. Renowned for their energy efficiency and carbon neutrality, MFCs leverage the abundance of microorganisms that decompose substrates found in various environments like wastewater, sludge, and sea sediments, where bacterial populations thrive. These cells benefit from the collective action of numerous microorganisms in electricity generation (Ramya and Senthil Kumar, 2022). Microbial fuel cells (MFCs) are ecologically benign and commercially feasible choices for the generation of energy and the treatment of wastewater because they utilise microbial populations to catalyse electrochemical processes. This allows them to transform organic materials into energy while simultaneously treating wastewater (Naha et al., 2023). Although there has been significant research in microbial fuel cell (MFC) technology, predictive modeling efforts have largely ignored the incorporation of advanced machine learning techniques capable of managing intricate, nonlinear relationships between environmental variables and power density outcomes. Current modeling approaches often lack accuracy and face challenges with overfitting, particularly when working with the extensive and varied datasets that are usually associated with MFC performance evaluations. Moreover, many of the current models do not provide thorough insights into how crucial environmental factors, such as NH4-N levels, pH, and temperature, interact to affect MFC efficiency. Challenges such as scale-up issues, expensive manufacturing processes, and low electricity generation persist, impeding their broader adoption in real-world scenarios. Opportunities for improvement remain in various aspects of MFC technology to address these shortcomings and unlock their full potential (Ramya and Senthil Kumar, 2022). However, advancements in material engineering, microbial biotechnology, and fabrication techniques hold promise for overcoming these hurdles. By improving conversion efficiency and reducing costs, these advancements could unlock the full potential of MFCs and expand their practical applications (Hoang et al., 2022).

(Ishaq et al., 2024) investigates the effectiveness of a dual compartment MFC (DC-MFC) in treating landfill leachate by removing chemical oxygen demand (COD) while generating electricity. The setup utilizes anaerobic sludge as the anode compartment's inoculum, separated by a Nafon117 membrane, with aerated cathode compartments. The study evaluates performance across varying COD concentrations, emphasizing COD removal, power generation, and Coulombic efficiency. Results show higher COD removal and power density at elevated organic matter concentrations, with the most efficient removal (89%) observed at 3325.0 mg L^− 1 COD concentration. However, efficiency decreased at higher concentrations, correlating with bacterial growth reflected by optical density changes.

In MFCs, the levels of ammonium nitrogen (NH₄-N) are an important indication of both the accessibility of substrates and the metabolic processes of micro-bacterial cells. Considering that ammonium nitrogen is a component that is frequently found in organic waste streams, such as leachate from landfills, it is an important parameter to consider when evaluating the performance of MFCs (Bavasso et al., 2016). Nitrogen can alter nutrients for microbial growth and pollutant removal (Wang et al., 2024). The effect that environmental factors like time, dose, pH, and temperature have on the concentration of NH₄-N can offer information that is highly beneficial in understanding how MFCs function and how efficiently they perform. These factors affect the concentration of ammonium nitrogen (NH₄-N), which in turn affects the power density (PD) output of the cells (Naha et al., 2023). The optimal pH range is crucial for the survival and effectiveness of microorganisms, affecting their growth and activity. The pH levels outside this range disrupt microbial function by altering their charge and biochemical reactions. Determining the ideal pH for both electricity-producing and degrading microorganisms is challenging. pH directly influences electron and proton generation, impacting electricity production. Methanogen growth is favoured at higher pH, while neutral to lower pH inhibits methane production. Anode pH significantly influences the performance of microbial fuel cells (MFC) and microbial electrolysis cells (MEC), with varying effects on hydrogen generation and power densities within specific pH ranges. For instance, hydrogen generation is notably higher within a pH range of 5.0 to 6.0, reaching approximately 8 m³ H2/m³/d. Conversely, power densities are considerably greater, up to about 1200 mW/m², when the pH falls between 6.5 and 7.5 (Mullai et al., 2023).

Biochemical reactions are highly sensitive to temperature, with each microbe having a specific temperature range for optimal activity, typically between 35°C and 40°C, promoting electricity generation. According to Arrhenius's law, increased temperature boosts power generation. However, beyond the optimum temperature, microbial enzyme structures weaken, hindering cell activity and reducing performance. Higher temperatures can also stimulate the growth of non-electrogenic microorganisms, competing with electrogenic ones. While mixed inoculum bioelectrochemical systems generally operate well at 40°C, starting at high heat and then lowering the temperature is recommended to mitigate adverse effects. Insulating materials like mineral wool and foam, along with solar energy, are employed to counteract the impact of high temperatures on substrate degradation (Mullai et al., 2023).

To optimize the functionality of Microbial Fuel Cells (MFCs) and enhance energy generation efficiency, it's crucial to thoroughly understand the intricate interplay among these variables. This understanding can be readily attained through modelling techniques (Naha et al., 2023). In addition to experimental approaches, modelling, particularly employing machine learning (ML) algorithms like artificial neural networks (ANN) and support vector machines (SVM), can serve to predict, investigate, and assess the efficiency of Microbial Fuel Cells (MFCs). In today's global landscape, there is a widespread demand for intelligent systems capable of efficiently addressing challenges, especially in industries that produce waste. The conversion of waste into biogas offers significant benefits, both environmentally and economically, as it substantially reduces greenhouse gas emissions while safeguarding air, water, and soil quality. ANN and SVM stand out as widely acknowledged ML techniques utilized across various domains for tasks such as classification, prediction, assessment, and forecasting (Zamrisham et al., 2024).

Alejo et al. (2018) explored two machine learning (ML) approaches, Support Vector Machine (SVM) and Artificial Neural Network (ANN), to predict the effluent composition, specifically Total Ammonia Nitrogen (TAN), in Anaerobic Digestion (AD) processes involving poultry manure. Their findings indicated that SVM achieved higher prediction accuracy (R2 = 0.898) compared to ANN (R2 = 0.875). Similarly, Wang et al. (2021) utilized ANN analysis to forecast alkalinity levels in an Acidogenic Co-Digestion (ACoD) system comprising corn straw, cow manure, and fruit and vegetable waste, achieving an impressive R2 value of 0.995. Moreover, Jaman et al. (2023) investigated biogas production prediction from ACoD of cow manure with molasses residue using a kinetic study and ANN model. Their results revealed superior predictive performance of ANN over the kinetic study, with an R2 value of 0.972. Additionally, Cinar et al. (2022) employed seven ML algorithms, including SVM, to predict temperature fluctuations in the AD process using pellets (animal feed material) substrate. SVM yielded the highest precision result of 0.930, indicating its potential for predicting non-linear and dynamic AD processes. In their study, Lim et al. (2024) utilized Artificial Neural Network (ANN) to forecast biofilm communities within Microbial Fuel Cells (MFCs) and predict power generation from wastewater treatment. The ANN model accurately predicted the total abundances of seven exoelectrogenic bacteria-associated genera in MFCs based on physicochemical properties of the sludge inocula, achieving accuracies ranging from 62–92%. Additionally, another ANN model was developed to integrate biofilm data and forecast power generation from wastewater, demonstrating an 84% accuracy when validated against literature studies. These findings highlight ANN's efficacy in predicting biofilm communities and MFC power generation, eliminating the need for complex biofilm meta-genome analysis and facilitating future parametric investigations and scale-up studies.

Optimizing Microbial Fuel Cells (MFCs) is crucial for maximizing energy production. Oyedeji et al, (2023) developed data-driven models using machine learning techniques such as support vector regression (SVR), artificial neural networks (ANNs), Gaussian process regression (GPR), and ensemble learners to optimize a typical MFC made from polymethylmethacrylate and two graphite plates. Two datasets were used to model power density and output voltage, considering different features like current density, anolyte concentration, and chemical oxygen demand. Hyperparameter optimization techniques including Bayesian optimization, grid search, and random search were employed to refine the models. The resulting models achieved 99% accuracy on testing set evaluations for both power density and output voltage.

Identifying significant parameters influencing power density (PD) output and understanding their relative importance in driving system efficiency is a crucial gap in microbial fuel cell (MFC) efficiency prediction. Feature selection and sensitivity analysis play pivotal roles in addressing this gap, allowing for the simplification of model complexity and the improvement of prediction accuracy. Despite the importance of these methodologies, their application in the context of MFC efficiency prediction remains underexplored. This research aims to bridge this gap by investigating strategies for feature selection and sensitivity analysis to optimize model performance and interpretability, particularly concerning the impact of environmental factors on MFC functionality. Therefore, the aim of this research is to create reliable predictive models that evaluate microbial fuel cell (MFC) performance using leachate as a substrate. Specifically, this research investigates how environmental factors such as pH, temperature, duration, and dosage affect ammonium nitrogen (NH4-N) and power density (PD) output. By concentrating on thorough feature selection and sensitivity analysis, this study enhances existing methodologies by presenting a new framework for optimizing MFC efficiency in sustainable energy scenarios. By employing various machine learning algorithms like CatBoost, XGBoost, and Random Forest, the research aims to evaluate the impact of time, dose, pH, and temperature on MFC functionality. Addressing these gaps is vital for developing reliable prediction models in complex domains such as MFC efficiency.

This study seeks to contribute insights into the development of efficient prediction models for MFC efficiency, addressing the need for reliable technologies in sustainable energy. By elucidating the intricate interactions among critical factors and their influence on MFC performance, the research aims to employ machine learning techniques and conduct detailed feature selection and sensitivity analysis. This approach is crucial for enhancing our understanding of MFC behaviour and for providing valuable recommendations for optimizing MFC operation and design. Ultimately, this research endeavours to facilitate the widespread adoption of MFCs as viable renewable energy sources for wastewater treatment and electricity generation.

In summary, this study aims to fill the gap in current understanding by emphasizing the importance of feature selection and sensitivity analysis in MFC efficiency prediction. Its relevance lies in its potential to advance sustainable energy technologies and provide practical recommendations for optimizing MFC performance. The novelty of this research lies in its focus on utilizing machine learning techniques and conducting comprehensive analysis to enhance predictive modelling in the context of MFCs, thereby contributing to their broader adoption for environmental and energy applications.

2.1. MFC construction, configurations and experimental set up analysis

MFCs (H-type double-chambered sets) with similar features to those described in earlier research (Hassan et al., 2018; Nor et al., 2015) were used with minor adjustments (modifying the reactor's scale). The experiment used laboratory-scale reactors made of Plexiglas MFCs, which were specifically constructed to meet the specifications needed. Each microbial fuel cell (MFC) had a cathode chamber and an anode chamber, with a total capacity of 100 ml (measured 5 cm x 5 cm x 4 cm). Regarding the electrodes, carbon felt with the following dimensions: height (h) = 1.71cm, width (w) = 0.62cm, and length (l) = 1.71cm were used for both the cathode and anode, as seen in Fig. 1. The electrodes were positioned at a distance of roughly 4cm from each other, with a Nafion 117 proton exchange membrane (PEM) between them. The Nafion 117 membrane, manufactured by DuPont in Delaware, USA, underwent pre-treatment by immersing it in a solution of 0.1 M H₂SO₄, followed by 0.1 M H₂O₂, and lastly rinsing it with deionized water. The procedure was conducted for 60 minutes at a temperature of 60°C (Bakhshian et al., 2011). The anode and cathode were completely submerged in acetone for a duration of 24 hours. Subsequently, they were rinsed with deionized water and left to dry in the open air overnight. This process was carried out to enhance porosity and minimise the impact of impurities (Özkaya et al., 2013).

2.2 Power output computations and calculations

The voltage potential was measured at regular 60-second intervals without the presence of an external resistor (known as open circuit voltage) until a steady value was attained. Therefore, 1 kilo-ohm resistors were employed in conjunction with a lab jack U3-V Multimeter, manufactured by Agilent Technologies in California, USA, which possesses a built-in data logging feature. This multimeter was attached to a computer to record voltage, current, and power, as seen in Fig. 1.

The current was determined using the formula I = V/R, where I represents the current, V represents the voltage, and R represents the resistance. The power density was computed using the method P = P/A (Wm^− 2), where P represents power in watts, A represents the surface area of the electrode, and I and V are used to get P (Ghasemi et al., 2013). The current density was determined by applying the formula J = I/A (A/m²), where J represents the current density, I represents the current in amperes, and A represents the surface area of the electrode in square metres. A fixed value of this parameter signifies the attainment of stable conditions. Ultimately, analytical measures, such as the determination of current density and coulombic efficiency, were carried out.

The Coulombic efficiency (CE) quantifies the ratio between the observed number of electrons generated during the oxidation process of organic matter and the theoretically expected quantity (Moharir & Tembhurkar, 2018). The Coulombic efficiency (CE) can instead be expressed as the proportion of organic matter used in energy production (Eq. 1). Certain bacteria that are not capable of producing electricity may employ organic materials for their development and to store carbon. In contrast, bacteria that are capable of producing electricity can use organic matter for their metabolic operations. The low Coulombic efficiency serves as an indicator of the limited existence of electrogenic bacteria.

C_E = \(\:\frac{8{\int\:}_{0}^{\text{t}}\text{I}\text{d}\text{t}}{\text{F}{\Delta\:}\text{C}\text{O}\text{D}\text{V}\text{A}\text{n}}\:\:\times\:\)100% Eq. (1)

Where, ΔCOD (COD_in-COD_out) (mg/L) is achieved through the subtraction of influent COD concentration to effluent COD concentration; F (96,485 C/mol) is the Faraday's constant, where I (mA) is the current across the resistor in the MFC reactor at operation time (t, d); The volume of working liquid at the anode chamber is represented with V_An.

2.3 Modelling approach (Machine learning algorithms)

2.3.1 Ensemble learning algorithms

A unique blend of feature selection and sensitivity analysis is presented to refine the predictive models, boosting both their accuracy and interpretability. This combination fills an important void in predicting microbial fuel cell efficiency, providing a more nuanced approach to understanding how environmental variables like time, pH, temperature, and NH4-N concentration affect power density output. This study Imports necessary libraries for ensemble categories of algorithms models such as scikit-learn, XGBoost, and CatBoost, Random Forest, Gradient Boosting, and AdaBoost to combine multiple base models to enhance predictive accuracy. Using Python, the analysis of these algorithms typically involves preprocessing data, splitting it into training and testing sets, training models with appropriate parameters, evaluating performance using metrics like accuracy or mean squared error, comparing results to select the best-performing algorithm, and optionally tuning hyperparameters for further optimization. This approach facilitates the development of robust predictive models for classification or regression tasks, contributing to improved decision-making in various domains.

2.3.2 Evaluation matrix

This study employed three assessment metrics: mean absolute error (MAE), root mean square error (RMSE), and R squared which are as follows:

\(\:\text{R}\text{M}\text{S}\text{E}=\:\sqrt{[\:\frac{1}{\text{m}}\:}\sum\:_{\text{i}=1}^{\text{m}}\:{(\text{y}}_{\text{i}}-\:\text{ŷ}\text{i}\:\left){\:}^{2}\:\right]\:\) Eq. (2)

\(\:\:\text{M}\text{A}\text{E}=\:\frac{1}{\text{m}}\sum\:_{\text{i}=1}^{\text{m}}\:{|\text{y}}_{\text{i}}-\:{\text{ŷ}}_{\text{i}}|\:\) Eq. (3)

\(\:\text{R}-\text{s}\text{q}\text{u}\text{a}\text{r}\text{e}\text{d}=1-\:\frac{\sum\:_{\text{i}=1}^{\text{m}}\:{(\text{y}}_{\text{i}}-\:{\text{ŷ}}_{\text{i}}){\:}^{2}}{\sum\:_{\text{i}=1}^{\text{m}}\:{|\text{y}}_{\text{i}}-\:{\text{ȳ}}_{\text{i}}|}\) Eq. (4)

where y_i\(\:-\:ŷi\:\) indicate, respectively, the ith actual and estimated rates, and m is the total number of estimates in the test set. \(\:{ȳ}_{i}\) denotes the mean value of the actual value yi.

2.3.3 Model setup and dataset configuration.

The dataset consisted of about 1000 samples, which were divided into three separate sections: training (70%), validation (10%), and testing (20%). By employing this subsection, one may assess the effectiveness of the model's training, fine-tune its hyperparameters using the validation set, and evaluate the accuracy of its predictions using the test set. Table 1 demonstrates that NH4-N concentrations, temperature, pH, and time are the input parameters for each subgroup, whereas power density serves as the output variable.

Table 1

Mean, maximum and minimum values of input and output variables.
Variables		Unit	Average	Minimum	Maximum	SD
Input	Time	S	108030	60	216000	107970
	pH	-	6.88	6.85	7.00	0.0793
	Temperature	^OC	30.21	27.31	33.42	3.0563
	\(\:{\text{N}\text{H}}_{4}-\text{N}\)	mg/L	84.43	0.00159	283.38	145.4949
Output	Power density	mWm^− 2	127.92	0.35	339.41	171.2522

2.3.4 Model approach and data simulation.

The present study applies six machine learning models, namely Random Forest (RF), Gradient Boosting (GB), AdaBoost, Decision Tree (DT), Extreme Gradient Boosting (XGB), and CatBoost, to predict the power density in microbial fuel cells using input parameters. The learning algorithms receive the input variables after passing through many data preparation steps, including preliminary processing, ADF testing, reliability evaluation, standardisation, stationarity, and normalisation. 70% of the dataset was allocated as a training set to generate prediction models for each chosen model. We applied a validation set, constituting 10% of the dataset, to fine-tune hyperparameters, and allocate 20% for testing purposes. Figure 2 illustrates the procedure.

3.1. Optimal Hyper-parameters

Throughout the modelling process, we employed similar ensemble models for each of the training and test sets, ensuring systematic uniformity. The prediction process of the training dataset involves determining the optimal model parameters, which are subsequently utilised to generate predictions on the test dataset. As shown in Table 2, the efficacy of the model is greatly influenced by parameters such as max_depth, min_samples_split, samples_split, splitter, and min_samples_leaf. The max_depth parameter acts as a normalisation component, controlling the depth of the decision tree to prevent overfitting. The choice of max_depth = 3 implies a compromise between complexity and straightforwardness. The parameters min_samples_split and min_samples_leaf influence the quantity of samples utilised, aiding in the control of overfitting. The variability and significance of the model are contingent upon the specific values assigned to the learning rate, n_estimator, and tree-related parameters in ensemble techniques like XGBoost, CatBoost, Adaboost, and Gradient Boosting. Decreasing the learning rate, for example, by setting XGBoost to 0.1, enhances convergence and decreases the probability of overfitting.

Table 2

The parameters of comparison models.
Models	Parameters	Optimal parameters
Random forest	n-estimator	100
	max_depth	3
	Min samples_split	2
	samples _split	2
	Splitter	-
	Min_samples _leaf	1
	n-estimator	200
XGBoost	learning_rate	0.1
	max-depth	3
	n-estimator	200
CatBoost	learning _rate	0.1
	max_depth	6
	n_estimator	200
Decision tree	learning rate	0.1
	max_depth	3
	n_estimator	50
	col sample_bytree	1
	Min_child_samples	20
	Min_child_wight	0.01
	Num_leaves	31
	subsanple_for_bin	2000
	Subsample	1
Adaboost	n_estimator	50
	Min _samples_split	2
	Min_samples _leaf	1
	learning _rate	2
Gradient boosting	n_estimator	100
	Min _samples_split	2
	Min_samples _leaf	1
	learning _rate	0.2
	Alpha	0.9
	max_depth	5
	validation fraction	0.1

The objective is to construct and evaluate predictive models by employing Random Forest, Gradient Descent, Decision Tree, XGBoost, and CatBoost algorithms. These models will be used to analyse the impact of time, ammonium nitrogen concentration (NH₄-N), pH, and temperature on the power density (PD) output of microbial fuel cells in the treatment of leachate. Feature selection and parameter sensitivity analysis are crucial in machine learning to optimise the performance and interpretability of models. Feature selection streamlines the model by eliminating unnecessary or redundant predictors, resulting in a reduction in overfitting and processing requirements, while also improving the model's ability to accurately forecast new data. Conversely, sensitivity analysis assesses the resilience of a model by analysing the impact of changes in input parameters on the output. This analysis determines the key characteristics that have the greatest influence on model predictions. Figure 3 is a heatmap of a sensitivity analysis that highlights the relationship between NH4-N (Ammonium Nitrogen) and temperature (Tempt) with respect to their influence on the power density (PD) output of microbial fuel cells in the treatment of leachate. A positive correlation value of 0.58 between NH₄-N and temperature suggests a moderate positive relationship, indicating that as NH₄-N concentration increases, temperature also tends to increase, or vice versa. This may suggest that temperature variations have an impact on NH₄-N dynamics in the system or that NH₄-N concentration varies with changes in temperature in MFC environments. The absence of strong cross-relations between other parameters (as shown by the blue areas with little or no correlation) suggests that the relationship between NH4-N and temperature has the most noticeable effect on the outcome being analyzed.

Table 3 assesses the performance of six machine learning models in predicting Power Density (PD) based on Ammonium Nitrogen (NH₄-N) concentrations. The evaluation is done using parameters such as R-squared (R²), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The Random Forest (RF) model has a remarkable level of explanatory capability, as seen by an R² value of 0.9830 and reasonably low error rates. Gradient Boosting (GB) is a powerful method, achieving an R² value of 0.9777, albeit it exhibits somewhat greater error rates compared to RF. The performance of AdaBoost decreases, as indicated by an R² value of 0.8801 and increasing error values, which implies less accuracy. The Decision Tree (DT) model performs superiorly than AdaBoost but is surpassed by RF and GB. XGBoost has exceptional performance with an R² value of 0.9917 and exhibits the lowest errors across all models, save for CatBoost. Ultimately, CatBoost surpasses all other models with an R² value of 0.9969, as well as the lowest MSE and RMSE. This indicates that CatBoost provides the most accurate and exact predictions for the given dataset. By employing the CatBoost algorithm, this research addresses significant shortcomings in existing MFC predictive modeling studies. In contrast to traditional algorithms, CatBoost is crafted to manage high-dimensional data with little preprocessing and excels at reducing overfitting through its innovative gradient boosting framework. The capability of this algorithm to capture sophisticated, nonlinear connections between environmental factors and MFC power density leads to a more precise and resilient predictive model, addressing the deficiencies of less advanced methods. Additionally, CatBoost’s effectiveness greatly improves model interpretability, offering deeper understanding of the interactions between variables such as NH4-N concentration, pH, and temperature, which were not fully examined in earlier studies. The table compares several machine learning models for forecasting Power Density based on Ammonium Nitrogen levels. The RMSE metric is used to measure the prediction accuracy of each model. The Random Forest (RF) model has an RMSE (Root Mean Square Error) value of 16.3430, indicating a satisfactory level of precision. On the other hand, Gradient Boosting (GB) demonstrates a somewhat lower level of precision, as seen by its RMSE value of 18.7132. AdaBoost exhibits a significantly greater root mean square error (RMSE) value of 43.3902, suggesting a lower level of consistency in its predictions. The Decision Tree (DT) exhibits superior accuracy compared to AdaBoost, but with a somewhat high Root Mean Square Error (RMSE) of 39.5169. The XGBoost (XGB) model achieves a high level of precision, as seen by its root mean square error (RMSE) value of 11.4091. CatBoost demonstrates superior performance by achieving the lowest Root Mean Square Error (RMSE) of 6.9888. This indicates that CatBoost produces the most precise predictions with the fewest average deviations from the actual values. The R² value of the Gradient Boosting (GB) model differs from that of the Random Forest (RF) model by a slight margin of 0.54%. In contrast, AdaBoost and Decision Tree (DT) exhibit more significant disparities of 11.05% and 8.75% respectively. XGBoost (XGB) has a negligible deviation of only 0.88% from RF, whereas CatBoost marginally outperforms RF with a 1.40% higher R² value, suggesting a little enhancement in explanatory capability. Figure 4 displays the scatter plot of the top three models.

Table 3

Performance analysis for PD based on \(\:{\text{N}\text{H}}_{4}-\text{N}\) implication
Models	R²	MSE	RMSE
RF	0.9830	267.0929	16.3430
GB	0.9777	350.I832	18.7132
AdaBoost	0.8801	1882.7061	43.3902
DT	0.9006	1561.5875	39.5169
XGB	0.9917	130.1668	11.4091
CatBoost	0.9969	48.8430	6.9888

The time series graphs in Fig. 5 illustrate the effectiveness of several machine learning models in predicting Power Density (PD) within a certain time period. The Observed Power Density (PD), indicated by the black dashed line, serves as the reference point for comparisons against various model predictions. The Random Forest Prediction (green line) closely mirrors the observed data, although there are some inconsistencies at particular points. The XGBoost Prediction (blue line) generally adheres to the observed trend but shows considerable divergence during certain phases, especially towards the end of the series. The CatBoost Prediction (red line) seems to be the model that aligns most closely with the observed data, particularly during the stable trend phase (marked in red). A stable trend section (marked in red) signifies a timeframe where the CatBoost model aligns very well with the observed power density, while a divergence trend section (marked in blue) highlights where the Random Forest model begins to deviate from the observed data. The label "Divergence: RF Deviates" identifies the exact point of deviation between the Random Forest predictions and the observed power density. By visually comparing these models, we can assess how closely each model's predictions align with the actual observed data. CatBoost appears to be the most reliable model throughout the time series, especially during the stable trend phase. Its close correlation with the observed data implies that it may excel at capturing the underlying dynamics of power density. The Random Forest model seems to perform adequately at the beginning but starts to stray from the observed data in the latter part of the time series (around the 30–40 time unit range). XGBoost demonstrates a generally good fit, yet it experiences a more pronounced divergence in the mid-to-late sections. The model's ability to generate accurate predictions seems to improve as it aligns more closely with the peaks and troughs of the observed data. The graphic suggests a possible ranking of the models based on accuracy, with CatBoost leading in precision, followed by XGBoost, and then Random Forest.

However, to validate this ranking, it is necessary to refer to the actual performance measures such as R-squared, MSE, and RMSE that were previously addressed. The evaluation of the Mean Squared Error (MSE) amongst models, with Random Forest (RF) as a reference, demonstrates notable discrepancies in prediction precision. Gradient Boosting (GB) demonstrates a 26.92% rise in Mean Squared Error (MSE), suggesting a moderate increase in prediction inaccuracy compared to Random Forest (RF). On the other hand, AdaBoost's mean squared error (MSE) is significantly greater, surpassing RF's MSE by more than double (150.30%), indicating a substantially bigger average prediction error. Similarly, the Mean Squared Error (MSE) of the Decision Tree (DT) model is 141.58% higher than that of the Random Forest (RF) model, indicating worse accuracy in its predictions. Conversely, XGBoost (XGB) exhibits a 68.94% reduction in Mean Squared Error (MSE) in comparison to RF, indicating a significant enhancement in prediction accuracy. CatBoost exhibits the most notable improvement in prediction accuracy, with a Mean Squared Error (MSE) that is 138.16% lower than that of RF. This signifies a substantial decrease in average prediction errors and establishes CatBoost as the most exact model among the ones analysed (refer to Fig. 6 ). The investigation reveals significant disparities in model performance, with CatBoost and XGBoost demonstrating higher accuracy compared to other models in forecasting Power Density based on Ammonium Nitrogen levels.

4.1. Conclusions

This research illustrates the efficiency of employing machine learning algorithms for predictive modelling to evaluate the influence of crucial factors on the performance of microbial fuel cells (MFCs). The study examines the impact of time, dose, pH, and temperature on the concentration of ammonium nitrogen (NH₄-N) and the resulting power density (PD) output. It offers helpful information for enhancing the efficiency of microbial fuel cells (MFCs) in sustainable energy generation and wastewater treatment. By using feature selection and sensitivity analysis approaches, the clarity and accuracy of predictive models are improved. This helps in identifying the key aspects that drive MFC performance.

The assessment of six machine learning models reveals different degrees of predicted accuracy, with CatBoost being identified as the most precise model, closely trailed by XGBoost and Random Forest. The time series plots demonstrate the models' ability to accurately predict PD over time, highlighting their capacity to capture the temporal dynamics and trends in MFC functioning. The evaluation of mean squared error (MSE) highlights the substantial improvements in predictive accuracy attained by CatBoost, showcasing its superiority in modelling MFC efficiency. This research makes important strides by presenting a unique blend of feature selection and sensitivity analysis techniques that enhance the performance and clarity of predictive models for MFC efficiency. Importantly, this method tackles a less explored domain in microbial fuel cell research, providing a more accurate means of assessing the impact of crucial environmental variables. Additionally, the study reveals that the CatBoost model yields the most precise predictions (R²: 0.9969, RMSE: 6.9888), outpacing conventional models like XGBoost and Random Forest, thereby enhancing predictive abilities for intricate systems. In conclusion, this investigation not only furthers the advancement of sustainable energy technologies but also directly fills significant voids in microbial fuel cell (MFC) predictive modeling. Through the use of CatBoost, the research presents a more precise and interpretable model that adeptly captures the intricate relationships between environmental elements and power density output. This strategy addresses the shortcomings of earlier studies, which faced issues with overfitting and lacked accuracy, thus establishing a more dependable groundwork for future MFC research and applications. The knowledge acquired from this study may be used to guide the development and improvement of MFC systems, hence promoting their extensive use as viable sources of renewable energy for wastewater treatment and power production. Additional investigation in this field might examine supplementary factors and modelling methods to improve the forecasting powers and practicality of MFCs in real-life scenarios.

4.2. Recommendations

Based on the results and knowledge gained from the present research, a number of suggestions can be made to further enhance the predictive modeling of microbial fuel cell (MFC) efficiency and advance sustainable energy technologies.

The CatBoost algorithm should be further investigated for its applicability in additional studies and real-world scenarios connected to MFC, due to its exceptional prediction accuracy demonstrated in this work. This involves evaluating its effectiveness and durability by analysing its performance on various substrates, under different working circumstances, and in different MFC configurations.
Although this study primarily examined time, dose, pH, and temperature, subsequent studies might investigate the inclusion of supplementary variables such as substrate composition, microbial community dynamics, and environmental conditions. Incorporating these characteristics into predictive models has the potential to offer a more thorough comprehension of MFC performance and improve the accuracy of predictions.
Validation and verification experiments should be conducted using other datasets to evaluate the applicability and dependability of the prediction models created in this work. This will aid in confirming the accuracy of the models' predictions and verifying their suitability for use in various MFC systems and environmental circumstances.
Continuous monitoring and optimisation of MFC performance over an extended period is crucial for maximising energy production efficiency and ensuring sustainable operation. Researchers should use monitoring methods and optimisation strategies based on findings from predictive modelling to consistently enhance the performance and reliability of MFCs.

· Funding: There was no financing for the study.

· Competing interests: The authors state that they have no other financial or personal affiliations or interests that would have affected the research presented in this paper.

Alejo, L., Atkinson, J., Guzmán-Fierro, V., & Roeckel, M. (2018). Effluent composition prediction of a two-stage anaerobic digestion process: machine learning and stoichiometry techniques. Environmental Science and Pollution Research, 25, 21149-21163.
Bakhshian, A., Hamidi, F., & Ezati, M. (2011). Relationship between organizational intelligence and entrepreneurship among university educational managers. The Journal of Mathematics and Computer Science, 3(4), 413-421.
Bavasso, I., Di Palma, L., & Petrucci, E. (2016). Treatment of wastewater in H-type MFC with protonic exchange membrane: Experimental study of organic carbon and ammonium reduction with electrochemical characterization. Chemical Engineering Transactions, 47, 223–228. https://doi.org/10.3303/CET1647038
Cinar, E., Kalay, S., & Saricicek, I. (2022). A Predictive Maintenance System Design and Implementation for Intelligent Manufacturing. Machines 2022, 10, 1006.
Ghasemi, M., Ismail, M., Kamarudin, S. K., Saeedfar, K., Daud, W. R. W., Hassan, S. H., ... & Oh, S. E. (2013). Carbon nanotube as an alternative cathode support and catalyst for microbial fuel cells. Applied Energy, 102, 1050-1056.
Hassan, H., Jin, B., Donner, E., Vasileiadis, S., Saint, C., & Dai, S. (2018). Microbial community and bioelectrochemical activities in MFC for degrading phenol and producing electricity: Microbial consortia could make differences. Chemical Engineering Journal, 332, 647–657. https://doi.org/10.1016/j.cej.2017.09.114
Hoang, A. T., Nižetić, S., Ng, K. H., Papadopoulos, A. M., Le, A. T., Kumar, S., & Hadiyanto, H. (2022). Microbial fuel cells for bioelectricity production from waste as sustainable prospect of future energy sector. Chemosphere, 287, 132285.
Ishaq, A., Said, M. I. M., Azman, S. B., Houmsi, M. R., Isah, A. S., Jagun, Z. T., ... & Abubakar, U. A. (2024). The influence of various chemical oxygen demands on microbial fuel cells performance using leachate as a substrate. Environmental Science and Pollution Research, 1-16.
Jaman, K., Idrus, S., Wahab, A. M. A., Harun, R., Daud, N. N. N., Ahsan, A., ... & Uddin, M. A. (2023). Influence of molasses residue on treatment of cow manure in an anaerobic filter with perforated weed membrane and a conventional reactor: Variations of organic loading and a machine learning application. Membranes, 13(2), 159.
Lim, B., Park, P. M., Eun, D. M., Kim, D. W., Kang, C., Jeon, K. J., ... & Youn, J. S. (2024). Development of an Atopic Dermatitis Incidence Rate Prediction Model for South Korea Using Air Pollutants Big Data: Comparisons Between Regression and Artificial Neural Network. Korean Journal of Chemical Engineering, 1-11.
Moharir, P. V., & Tembhurkar, A. R. (2018). Effect of recirculation on bioelectricity generation using microbial fuel cell with food waste leachate as substrate. International Journal of Hydrogen Energy, 43(21), 10061-10069.
Mullai, P., Vishali, S., Sambavi, S. M., Dharmalingam, K., Yogeswari, M. K., Raja, V. V., ... & Rene, E. R. (2023). Energy generation from bioelectrochemical techniques: Concepts, reactor configurations and modeling approaches. Chemosphere, 342, 139950.
Naha, A., Debroy, R., Sharma, D., Shah, M. P., & Nath, S. (2023). Microbial fuel cell: A state-of-the-art and revolutionizing technology for efficient energy recovery. Cleaner and Circular Bioeconomy, 5, 100050. https://doi.org/10.1016/j.clcb.2023.100050
Nor, M. H. M., Mubarak, M. F. M., Elmi, H. S. A., Ibrahim, N., Wahab, M. F. A., & Ibrahim, Z. (2015). Bioelectricity generation in microbial fuel cell using natural microflora and isolated pure culture bacteria from anaerobic palm oil mill effluent sludge. Bioresource Technology, 190, 458–465.
Oyedeji, M. O., Alharbi, A., Aldhaifallah, M., & Rezk, H. (2023). Optimal Data-Driven Modelling of a Microbial Fuel Cell. Energies, 16(12), 4740.
Özkaya, B., Cetinkaya, A. Y., Cakmakci, M., Karadağ, D., & Sahinkaya, E. (2013). Electricity generation from young landfill leachate in a microbial fuel cell with a new electrode material. Bioprocess and biosystems engineering, 36, 399-405.
Ramya, M., Vardhan, K. H., & Kumar, P. S. (2022). Metal mixed biochar electrodes for the generation of electricity with high power density in microbial fuel cell. Sustainable Energy Technologies and Assessments, 53, 102549.
Tariq, S., & Dan, S. (2023). Sensitivity Analysis and Feature Selection for Drilling-Oriented Models. Journal of Energy Resources Technology, 145(12). https://doi.org/10.1115/1.4062382
Tian, X., An, C., & Chen, Z. (2023). The role of clean energy in achieving decarbonization of electricity generation, transportation, and heating sectors by 2050: A meta-analysis review. Renewable and Sustainable Energy Reviews, 182, 113404.
Wang, Y., Huntington, T., & Scown, C. D. (2021). Tree-based automated machine learning to predict biogas production for anaerobic co-digestion of organic waste. ACS Sustainable Chemistry & Engineering, 9(38), 12990-13000.
Wang, X., Li, S., Mi, R., Dong, Y., Jiang, J., Guan, X., ... & Zhou, Z. (2024). Performance enhancement, bacterial communities optimization and emerging pollutants elimination by microalgal-bacterial consortium for treating aquaculture pond sediments. Journal of Environmental Management, 359, 121013.
Zamrisham, N. A. F., Idrus, S., Harun, M. R., Ab Razak, M. S., & Jaman, K. (2024). Biogas production by integrating lava rock, red clay & ceramic bio ring as support carrier in treatment of landfill leachate with liquidised food waste. Biochemical Engineering Journal, 204, 109221.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Electricity generation from renewable source- A leachate based microbial fuel cell machine learning approach

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

2.1. MFC construction, configurations and experimental set up analysis

2.2 Power output computations and calculations

2.3 Modelling approach (Machine learning algorithms)

2.3.1 Ensemble learning algorithms

2.3.2 Evaluation matrix

2.3.3 Model setup and dataset configuration.

2.3.4 Model approach and data simulation.

3. Results and Discussion

3.1. Optimal Hyper-parameters

4. Conclusions and Recommendations

4.1. Conclusions

4.2. Recommendations

Declarations

References

Additional Declarations

Status:

Version 1