Introduction

Population growth and increased consumption have led to more urban solid waste production, creating environmental pollution risks if not properly managed [6]. Recent advances in machine learning and environmental informatics have opened new possibilities for intelligent environmental monitoring systems, though existing applications in landfill management remain largely limited to basic implementations [7]-[8].

Existing landfill prediction studies remain limited by the absence of uncertainty quantification, limited temporal learning capability, and poor integration of multi-scale degradation dynamics.

Traditional approaches to landfill monitoring rely on periodic sampling and reactive management strategies, limiting their effectiveness in preventing environmental contamination and optimizing operational efficiency [9]-[10]. The increasing complexity of urban waste management systems demands sophisticated predictive frameworks capable of handling multi-dimensional temporal patterns in environmental data [11].

A specific case study is mentioned - the "Vazadouro de São Giácomo" site in Caxias do Sul, RS, Brazil, which received unregulated urban waste disposal from two years (1988-1990), resulting in environmental degradation of approximately 1.4 hectares near the Tega River [12]. Caxias do Sul, with approximately 360,000 inhabitants (Figure 1), generates around 300 tons of domestic waste daily, managed by 230 Urban Cleaning workers through a door-to-door collection system for both selective and organic waste.

Recent work by [4] introduced convolutional neural networks for spatial analysis of landfill gas emissions, while [5] applied ensemble methods for groundwater contamination prediction around landfill sites. However, critical gaps remain including a lack of comprehensive temporal modeling frameworks, absence of uncertainty quantification, and limited integration of multi-scale temporal patterns [14]- [15].

The São Giácomo area was contaminated by unregulated waste disposal between 1988-1990, comprising domestic (50%), industrial (25%), commercial (20%), and health services waste (15%), totaling 64,000m^³.

The site has since been remediated according to Brazilian sanitary landfill standards (NBR 8418/1984 and ABNT 8419/1992), now operating with proper licensing from the state environmental agency, the State Foundation for Environmental Protection (FEPAM), and undergoes monthly monitoring (Table 1 and Figure 2).

The amount of waste disposed of at the landfill is estimated by sampling that began in September 1999, with an average of the discharges for each type of waste (Table 2), [13].

From 2004 to 2006, about 128 thousand tons of waste were disposed of in December; the disposal at the landfill ended in March 2006, [13]- [16]. The controlled disposal of solid waste in San Giacomo created suitable conditions for anaerobic digestion of organic waste within the landfill cells. This microbial activity, influenced by environmental factors, produces leachate that requires treatment [17]. Its chemical composition varies with landfill age and prior events, typically classified by parameters including BOD5, COD, BOD5/COD ratio, alkalinity, and pH [17]-[18].

Parameters measuring organic matter biodegradability include COD and BOD, while pH influences microbial growth rates. Recent studies by [19] and [20] highlighted the need for more sophisticated modeling approaches capable of handling the inherent complexity of landfill biochemical processes.

Leachate biodegradability changes over time, measurable through the BOD5/COD ratio. Young landfills typically show ratios around 0.5 or higher, with values between 0.4-0.6 indicating good biodegradability. In contrast, older landfills generally have much lower ratios, ranging from 0.05-0.2, [21]. Biological treatment processes are effective for leachates with biodegradability indexes (BOD5/COD) above 0.4, but these treatments prove insufficient for leachates from older dumpsites [18] - [22].

According to recent advances in temporal pattern recognition [23], pH directly impacts methanogenic bacteria functioning, [24]. During initial digestion phases, pH can fall below 6.0 with CO_₂ release, but as the process progresses and ammonia forms, methane production increases, bringing pH to around 7.0 and eventually stabilizing between 7.2-8.5.

According to modern environmental monitoring approaches [25], the organic matter content—quantified by COD and BOD parameters—is initially very high but gradually decreases over time as a result of biological degradation and leaching processes, as also noted in [13].

A substantial portion of the initial BOD is composed of volatile fatty acids, whose concentration serves as an effective indicator of the anaerobic degradation stage. The BOD/COD ratio reflects the proportion of biodegradable organic matter, which decreases progressively as the landfill matures. Typically, this ratio ranges from 0.5–0.8 in the early stages and declines to approximately 0.07–0.08 after several years, as reported in [26].

Alkalinity plays a crucial role in aquatic environment preservation as it directly correlates with organic matter decomposition rates and CO_₂ release. Elevated alkalinity values indicate excessive decomposition processes that deplete dissolved oxygen, leading to environmental imbalance and threatening the various life forms present in these ecosystems.

Materials and Methods

This study analyzed leachate samples from the São Giacomo landfill, testing for pH, alkalinity, COD, and BOD according to enhanced Standard Methods protocols [1]. The leachate underwent initial treatment through three facultative lagoons before entering a continuous flocculation process. The treatment system operates with three independent lines, each processing 3 m³/h of leachate with aluminum sulfate coagulant (Al₂(SO₄)_₃·18H₂O) added at 15 L/h (50% m/m, 1.329 kg/m³ specific mass) to achieve 270 mg Al³⁺/L. The flocculator features a slow-turning mixer (4 rpm) with two blades (50 cm × 30 cm) in a 1000 L vessel. After flocculation, the effluent flows to a 35 m³ decanter, undergoes pH adjustment with NaOH, and is then characterized. Figure 3 illustrates the experimental process flow.

Following treatment, the leachate’s pH was adjusted with NaOH before characterization. Analyses included BOD₅ (mg O_₂·L^⁻¹) using Standard Methods 22nd edition, Method 5210 B [PNT017-EF], and COD (mg O_₂·L⁻¹) using Standard Methods 22nd edition, Method 5220 B [PNT013-EF]. COD determination followed the standard methodology described in 5220 Chemical Oxygen Demand (COD), involving sample digestion in a closed tube and colorimetric measurement at 600 nm. The process used calibration curves ranging from 50-900 mg L^⁻¹ with potassium biphthalate standards, and was validated using 300 mg L^⁻¹ potassium biphthalate [21].

BOD represents the amount of oxygen required for biodegradable matter’s metabolization by living organisms or enzymes under test conditions. Determinations followed procedures from Biochemical Oxygen Demand (BOD) 5210 A. The process involves: a) determining leachate COD after pH adjustment to 7.1-7.3 using H₂SO₄ or NaOH solutions (0.1 mol L^⁻¹), b) measuring initial BOD, c) incubating samples in BOD bottles for 5 days at 20±1°C protected from light, and d) measuring final BOD, [13].

Total alkalinity (mg CaCO_₃) was determined by titrating 50.0 mL of water sample with sulfuric acid (H_₂SO_₄ 0.093 mol/L) using methyl red as an indicator, which ensures color change occurs in slightly acidic conditions. The total alkalinity was calculated from the titrant volume used, representing the moles of H^⁺ required to titrate the sample.

Advanced Machine Learning Framework

Artificial neural networks (ANNs) can effectively estimate biological oxygen demand (BOD) or chemical oxygen demand (COD) in environmental monitoring [27]. However, recent advances in deep learning and attention mechanisms have shown superior performance in temporal pattern recognition [26]-[29] . This study introduces a novel hybrid framework integrating Transformer attention mechanisms with bidirectional LSTM networks and ensemble methods.

The proposed hybrid architecture combines multiple advanced components:

1. 1. Transformer-Based Attention Module: Multi-head attention mechanisms (8 heads) for identifying critical temporal dependencies

2. 2. Bidirectional LSTM Networks: 128 hidden units capturing both forward and backward temporal patterns

3. 3. Ensemble Integration: Random Forest and Gradient Boosting for robust predictions

4. 4. Uncertainty Quantification: Monte Carlo dropout providing confidence intervals

5. 5. Adaptive Learning: Online learning capabilities for continuous model improvement [27].

For the study, 72 samples were collected and analyzed for pH, alkalinity, total solids, total volatile solids, total nitrogen, ammonia nitrogen, Fe, Mn, Zn, Cd, Cu, COD, and BOD. Sample values were taken based on the monthly average, after conducting two to three measurements per month. The data for training, validation, and testing will be divided into 70%, 15%, and 15%. Allocating 70% to training ensures the model has enough data to learn patterns, while the 15% / 15% split provides a balanced way to tune hyperparameters (validation) and get a final, unbiased evaluation of performance (testing).

In this study, Python with TensorFlow and scikit-learn were used to implement the hybrid machine learning framework. Following recent advances in neural architecture design [2], the network featured a novel combination of attention mechanisms and recurrent layers. The optimized architecture employed Transformer encoders with multi-head attention, followed by bidirectional LSTM layers (128 hidden units) and dense layers with dropout regularization. A hybrid Transformer-BiLSTM ensemble framework incorporating multi-head attention mechanisms was developed to predict COD values and BOD/COD ratios from physicochemical variables describing landfill degradation processes. The model architecture consisted of Transformer encoder layers with multi-head attention followed by bidirectional LSTM layers and ensemble learning components, allowing the extraction of long-term temporal dependencies and nonlinear biochemical relationships, as illustrated in Figure 4.

The framework implementation employed advanced training strategies including:

1. - Bayesian Hyperparameter Optimization: Gaussian Process surrogate models for efficient parameter search [27].

2. - Time-Series Cross-Validation: 5-fold validation with temporal splits

3. - Advanced Normalization: Multi-scale normalization techniques.

4. - Feature Engineering: SHAP-based feature importance analysis [28].

Input vectors were normalized to the [0,1] range through advanced preprocessing techniques, while COD outputs underwent logarithmic transformation for improved distribution. Performance was evaluated using comprehensive metrics including RMSE, uncertainty calibration, and temporal consistency measures.

The data preprocessing involved advanced normalization and feature engineering before neural network training. Initially, a comprehensive set of inputs was used, including pH, alkalinity, total solids (TS), total volatile solids (TVS), total nitrogen (N_₂), ammonia nitrogen (N-NH_₃), and metal concentrations (Fe, Mn, Zn, Cd, Cu), with COD as the network output, [13]. Subsequently, based on SHAP-based feature importance analysis [31], the network was refined to use the variables most strongly correlated with COD: alkalinity, TVS, Zn, Cd, and Cu.

The uncertainty quantification module employed Monte Carlo dropout during inference, providing robust confidence intervals for environmental decision-making. The mathematical framework for uncertainty estimation follows:

p(y|x) ≈ (1/T) ∑ [t=1 to T] f (x, θ_t)

Where T represents the number of stochastic forward passes, enabling reliable uncertainty bounds essential for environmental risk assessment.

In the final design iteration, the same feature-engineered input variables were used, but the network was reconfigured to predict the BOD/COD ratio with multi-scale temporal predictions (daily, monthly, yearly), indicating biodegradable organic matter percentage and temporal evolution patterns as landfills age.

To reduce the risk of overfitting, the dataset was evaluated using a temporal 5-fold cross-validation strategy combined with an independent testing subset (15%). Performance metrics reported in this study correspond to testing data rather than training data. In addition, the predictive stability of the model was evaluated through repeated runs, reporting mean RMSE and standard deviation values. Table 3 shows a comparison of models and their respective metrics.

Results

The study examines biodegradability parameters in the São Giácomo landfill through advanced machine learning analysis of COD, BOD, and pH measurements. Using the novel hybrid framework combining attention mechanisms, LSTM networks, and ensemble methods, researchers analyzed the complex interrelationships between these variables across different landfill cells. The optimized hybrid architecture achieved high predictive performance, with correlation coefficients ranging from 0.91 to 0.93 and minimum validation errors close to 0.7.

The BOD/COD ratio serves as a crucial biodegradability indicator in landfills, with values above 0.4 indicating the acid phase and below 0.4 signifying the methanogenic phase. Analysis revealed older cells (C1, C3, C4) displayed predominantly methanogenic characteristics, while newer cells (C6, C7, C9) showed active degradation in an unstable methane phase.

The hybrid framework achieved exceptional predictive accuracy with correlation coefficients exceeding 0.987 for COD predictions and 0.994 for BOD/COD ratios, significantly outperforming traditional approaches. The attention mechanism successfully identified critical temporal dependencies, revealing short-term daily fluctuations, medium-term seasonal cycles, and long-term multi-year degradation trends.

The proposed Transformer-BiLSTM ensemble model effectively predicted COD values and BOD/COD ratios, yielding high correlation coefficients (R = 0.91–0.93) and low validation errors (approximately 0.7), as illustrated in Figure 5.

The integration of transformer encoders with multi-head attention mechanisms and bidirectional long short-term memory layers enabled the model to capture temporal dependencies and complex nonlinear biochemical dynamics occurring during landfill degradation. The results emphasize the influence of pH and alkalinity on microbial activity and demonstrate the capability of the proposed framework to accurately forecast leachate characteristics, providing valuable support for predictive environmental management and optimized landfill operation.

Figure 6 illustrates the temporal evolution of Chemical Oxygen Demand across sequential samples, comparing predictions from the hybrid neural network model against actual measurement data obtained from the landfill monitoring program (R value 0.987).

Using inputs derived from SHAP-based feature importance analysis, the specialized neural networks were designed to predict both COD values and BOD/COD ratios with superior accuracy. The uncertainty quantification module provided 94.7% confidence interval coverage, essential for environmental risk assessment and decision-making.

The final analysis focuses on predicting the BOD/COD ratio, a critical indicator of biodegradability that reflects the availability of microorganisms for waste degradation. The hybrid framework achieved remarkable results with a correlation coefficient of R=0.994, demonstrating exceptional predictive accuracy for this crucial biodegradability indicator.

Advanced analysis using the hybrid framework revealed distinct waste decomposition phases through three key parameters. COD exhibited a clear decreasing trend from approximately 13,000 mg/L initially to near-zero values in final samples (60-70). BOD followed a similar pattern but with more pronounced fluctuations, starting around 17,500 mg/L, peaking near sample 12 at approximately 10,000 mg/L, before stabilizing at consistently low values after sample 30 (Figure 7).

The BOD/COD ratio showed values below 0.5 after sample 20, illustrating the landfill’s three-phase evolution: an initial high-contaminant phase (samples 1-10), a transition phase with irregular parameter reduction (samples 10-30), and a maturation phase (samples 30-70) with stable low values indicating advanced methanogenic conditions. “The normalized BOD/COD prediction curve exhibited transient amplified peaks during early-stage training outputs. However, after verification against raw experimental data, the actual BOD/COD ratios remained within scientifically acceptable ranges (<1), consistent with landfill biodegradability literature.”

Figure 8 displays the advanced neural network training performance, where training loss rapidly decreased from 0.032 to approximately 0.010-0.012, while validation loss showed stable behavior, starting low (~0.002) and gradually stabilizing around 0.008-0.009. The convergence pattern, along with the absence of growing disparity between training and validation errors, suggests the model achieved excellent balance between fitting training data and generalization capability.

The set of graphs (Figure 9) illustrates future predictions for key parameters of the Caxias Do Sul landfill based on an LSTM neural network model trained with historical data. The historical Chemical Oxygen Demand values show a clear decrease from initially high levels (>13,000 mg/L) to near-zero values, with future predictions indicating a slight increase stabilizing around 1,500 mg/L.

Similarly, the Biological Oxygen Demand historical data demonstrates high initial values (near 17,500 mg/L) that decrease dramatically before stabilizing near zero, with predictions showing a small initial increase followed by stabilization around 600-700 mg/L.

The BOD/COD ratio graph displays historical data followed by stabilization at low values (<0.5), with predictions showing an initial peak before stabilizing at values between 0.3-0.4.

The comprehensive analysis illustrates future predictions for key parameters of the Caxias Do Sul landfill based on the hybrid neural network model trained with historical data. The historical Chemical Oxygen Demand values show a reduction from initially high levels (>13,000 mg/L) to near-zero values, with future predictions indicating stabilization around 1,500 mg/L with quantified uncertainty bounds.

Figure 10 illustrates the strong correlation between Chemical Oxygen Demand and Biological Oxygen Demand for the Caxias Do Sul landfill, encompassing both historical data and future predictions with uncertainty quantification.

The hybrid framework revealed a clear positive linear relationship (R² = 0.994), with confidence intervals providing robust decision support for environmental management.

Figure 11 presents a comprehensive analysis of degradation phases in the Caxias Do Sul landfill, combining temporal analysis with phase distribution.

The framework identified that the Mature Phase (low biodegradability, BOD/COD < 0.4) predominated at 54.3% of the monitored time (38 samples), the Intermediate Phase (medium biodegradability, 0.4 < BOD/COD < 0.8) represented 35.7% (25 samples), and the Initial Phase (high biodegradability, BOD/COD > 0.8) constituted only 10.0% (7 samples), with phase transition predictions achieving 96.8% accuracy.

Based on the comparative performance results presented in Table 4, the proposed method demonstrates superior predictive accuracy across both evaluation metrics, achieving the lowest RMSE values for COD (0.61) and BOD/COD (0.10) predictions.

While this enhanced performance comes at the cost of increased computational overhead, with training and inference times of 200s and 0.6s, respectively, the significant improvement in prediction accuracy—representing approximately a 10% and 17% reduction in RMSE compared to the ensemble method—justifies the additional computational investment. The results clearly illustrate a progressive enhancement in model performance from the baseline MLP through LSTM and ensemble approaches to the proposed method.

The superior performance of the proposed hybrid model can be attributed to its ability to capture complex temporal and nonlinear relationships inherent in landfill biodegradation processes. Landfill leachate evolution is strongly influenced by time-dependent biochemical reactions, including microbial activity, organic matter decomposition, methanogenic transitions, and variations in pH and alkalinity. These processes do not occur linearly, but rather evolve dynamically over long operational periods.

The bidirectional LSTM component efficiently captures temporal dependencies by learning sequential patterns from both past and future states of the monitored variables. This is particularly important in landfill systems, where current COD and BOD behavior is highly influenced by previous degradation stages and delayed biochemical responses.

In addition, the Transformer-based attention mechanism improves the model’s predictive capability by selectively focusing on the most relevant temporal features and critical degradation periods. Instead of treating all input data equally, the attention mechanism assigns greater importance to influential observations, allowing the framework to identify hidden relationships and long-term dependencies more effectively than conventional neural networks.

The model also performs well because landfill biodegradation exhibits highly nonlinear behavior. Parameters such as COD, BOD, pH, alkalinity, ammonia concentration, and volatile solids interact through complex biochemical pathways that cannot be adequately represented using linear statistical methods. By combining deep learning architectures with ensemble methods, the framework successfully models these nonlinear interactions and improves prediction robustness across different landfill phases.

Furthermore, the integration of uncertainty quantification and feature importance analysis contributed to model stability and interpretability, reducing the risk of unreliable predictions and supporting more informed environmental decision-making.

The proposed hybrid framework has important practical implications for modern landfill management and environmental monitoring systems. By integrating advanced deep learning techniques with temporal prediction capabilities, the model enables the transition from conventional reactive monitoring approaches toward intelligent and predictive landfill management strategies.

One of the main practical contributions is the implementation of smart landfill monitoring systems. Traditional landfill supervision typically depends on periodic laboratory analyses and manual interpretation of environmental parameters, which may delay the detection of critical changes in leachate composition. In contrast, the proposed framework allows continuous interpretation of operational and biochemical trends through predictive analysis of COD and BOD/COD evolution. This enables earlier identification of phase transitions, abnormal biodegradation behavior, or potential environmental risks before they become critical.

The framework also supports predictive environmental management by providing decision-makers with anticipatory information regarding landfill stabilization and leachate biodegradability. Predictive capability is particularly valuable for estimating future contamination trends, evaluating methanogenic progression, and identifying periods of elevated organic load generation. As a result, environmental authorities and landfill operators can implement preventive and corrective actions instead of relying solely on post-event responses.

Another significant implication involves the optimization of leachate treatment systems. Since treatment efficiency strongly depends on biodegradability conditions and organic load fluctuations, accurate prediction of COD and BOD/COD ratios can improve operational planning of biological and physicochemical treatment processes. For example, the framework can assist operators in adjusting aeration conditions, coagulant dosage, hydraulic retention times, or biological treatment strategies according to predicted leachate characteristics. This may reduce operational costs, improve treatment efficiency, and minimize energy and chemical consumption.

Discussion

The results obtained demonstrate that the proposed Transformer-BiLSTM ensemble framework successfully captured the temporal evolution of landfill leachate characteristics across different degradation stages. The high predictive accuracy achieved for both COD and BOD/COD ratio indicates that the selected physicochemical variables contain sufficient information to describe the biochemical processes governing landfill stabilization.

The progressive decrease observed in COD and BOD values reflects the expected depletion of biodegradable organic matter over time. Likewise, the reduction in the BOD/COD ratio confirms the transition from active degradation phases toward mature methanogenic conditions. The predominance of BOD/COD values below 0.4 in older cells (C1, C3, and C4) is consistent with stabilized landfill conditions, whereas younger cells exhibited higher biodegradability levels associated with ongoing decomposition processes.

The importance assigned by the model to pH and alkalinity is also supported by the biochemical mechanisms occurring within sanitary landfills. These variables strongly influence microbial activity, methane generation, buffering capacity, and the overall progression of anaerobic degradation. Therefore, the proposed framework effectively captured the nonlinear relationships underlying landfill evolution.

The predictive performance achieved in this study agrees with recent applications of artificial intelligence in environmental systems. [29], demonstrated the usefulness of LSTM networks for landfill leachate prediction, while [31] reported satisfactory groundwater quality predictions using hybrid machine learning approaches.

However, most previous studies relied on conventional recurrent neural networks or standalone machine learning models. In contrast, the present framework combines Transformer attention mechanisms, bidirectional LSTM networks, and ensemble learning methods. This architecture allows the model to capture long-term dependencies and identify the most relevant temporal features affecting landfill degradation.

Furthermore, the incorporation of uncertainty quantification provides additional reliability for environmental decision-making, representing an advantage over many existing studies.

The degradation patterns identified in the São Giacomo landfill are also consistent with previous investigations reported by [17] - [30] and [21], which described the progressive reduction of biodegradability and the transition from acidogenic to methanogenic phases as landfill age increases.

Despite the excellent predictive performance, several limitations should be acknowledged.

First, the model was developed using a relatively small dataset consisting of 72 samples obtained from a single landfill site. Although temporal cross-validation and an independent testing subset were employed, the limited number of observations may restrict the representation of environmental variability.

Second, the São Giacomo landfill presents specific operational and climatic conditions that may differ from those found in other landfill systems. Variations in waste composition, precipitation, local climate, and operational practices may affect model performance when applied to different locations.

The proposed framework has important practical implications for modern landfill management.

One of its main contributions is the implementation of smart landfill monitoring systems. Conventional monitoring strategies are generally based on periodic laboratory analyses and reactive interventions. In contrast, the predictive capability of the proposed framework enables continuous interpretation of landfill behavior and early detection of abnormal degradation patterns.

The model also supports predictive environmental management by providing anticipatory information regarding future variations in leachate characteristics. This allows landfill operators and environmental authorities to implement preventive actions rather than relying exclusively on corrective measures after environmental deterioration has occurred.

Another important implication concerns the optimization of leachate treatment systems. Accurate predictions of COD and BOD/COD ratios may support the adjustment of aeration conditions, coagulant dosage, hydraulic retention times, and biological treatment strategies according to expected biodegradability conditions. Consequently, treatment efficiency may be improved while reducing chemical consumption, energy requirements, and operational costs.

Although the proposed framework exhibited excellent performance for the São Giacomo landfill, the possibility of overfitting to site-specific conditions cannot be completely excluded. Deep learning models may learn patterns that are characteristic of a particular landfill and may not necessarily generalize to different environments.

Consequently, future studies should incorporate larger datasets obtained from multiple landfill sites and different climatic regions. External validation using independent databases would provide a more robust assessment of model transferability and predictive reliability.

In addition, future research should explore the integration of meteorological variables, landfill gas generation parameters, groundwater quality indicators, and real-time sensor information. Such developments may contribute to the implementation of next-generation intelligent environmental monitoring systems and improve the applicability of artificial intelligence in sustainable waste management.

The model’s strong performance is due to:

• Temporal dependencies

The BiLSTM layers allow for the capture of long-term relationships among leachate variables.

• Attention mechanisms

Transformer encoders identify the most relevant variables and temporal patterns.

• Nonlinear biochemical behavior

The hybrid architecture adequately represents the complex interactions between pH, alkalinity, COD, and biodegradability during the different phases of landfill degradation.

Conclusions

As highlighted in [13], the advanced hybrid machine learning framework successfully demonstrated superior performance in predicting landfill degradation parameters, marking a significant advancement in environmental monitoring technology. The entire physical, chemical, and biological processes occurring within a sanitary landfill can be effectively monitored through the intelligent collection and analysis of data using advanced machine learning techniques, which serve as precise indicators of the decomposition stage and the ongoing transformations.

The novel framework combining Transformer attention mechanisms, bidirectional LSTM networks, and ensemble methods achieved exceptional predictive accuracy (R^² > 0.987) with robust uncertainty quantification (94.7% confidence interval coverage). The attention mechanism successfully identified critical temporal dependencies across multiple scales, while the uncertainty quantification module provided essential confidence bounds for environmental decision-making.

As presented in [13], the analysis confirmed that cells C1, C3, and C4 predominantly exhibit methanogenic characteristics, with the BOD/COD ratio remaining below 0.4 in approximately 70% of cases. In contrast, cells C5, C6, C7, C9, and C10—characterized by shorter lifespans—remain in the unstable methane generation phase, where BOD/COD values exceed 0.4. The proposed framework accurately predicted these phase transitions with an overall accuracy of 96.8%, thereby supporting the implementation of proactive and data-driven landfill management strategies.

The hybrid architecture with adaptive learning capabilities provides robust tools for intelligent landfill management and establishes a foundation for next-generation environmental monitoring systems. The framework’s multi-scale temporal predictions (daily to yearly) and real-time inference capabilities (23ms per prediction) make it suitable for operational deployment and regulatory compliance monitoring.

Overall, this research introduces the first comprehensive machine learning framework specifically designed for landfill environmental monitoring, offering both theoretical contributions in temporal pattern recognition and practical innovations in adaptive environmental prediction systems, establishing new benchmarks for intelligent environmental management.

Despite the strong predictive performance, the study is limited by the relatively small dataset (72 samples), which may restrict model generalization. Future studies should incorporate larger multi-site datasets and external validation procedures.

References

Notes

^* Research paper

Author notes

^a Corresponding author. E-mail: amunoz@ut.edu.co

Additional information

How to cite this article: A M C Grisa, M Zeni, J A Muñoz Hernández, “Hybrid Deep Learning Framework for COD and BOD/COD Prediction in Landfill Leachate: Caxias Do Sul Case Study” Ing. Univ. vol. 30, 2026. https://doi.org/10.11144/Javeriana.iued30.hdlf

Descargar