⚡ electrical engineering

Analysis of Sewage Treatment Plant’s Data to evaluate Quality of Effluent using Experimental Results and Prediction based on Artificial Intelligence Models (Case Study: Tehran Wastewater Treatment Plant)

This study evaluates the performance of Artificial Neural Network (ANN), Fuzzy Logic (FL), and Adaptive Neuro-Fuzzy Inference System (ANFIS) models, enhanced by Principal Component Analysis (PCA), to predict effluent quality at Tehran's wastewater treatment plant, ultimately demonstrating that the ANFIS model offers the highest accuracy for forecasting pollutant removal efficiency.

Original authors: Haniyeh Malek, Majid Ehteshami

Published 2026-09-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haniyeh Malek, Majid Ehteshami

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Water is the lifeblood of modern cities, but as populations swell and clean sources dwindle, the water we flush away is becoming a resource we can no longer afford to ignore. Municipal wastewater treatment plants act as the great filters of our civilization, taking the dirty water from homes and industries and cleaning it until it is safe enough to return to the environment or even reuse for farming. However, these plants are complex machines where chemical and biological processes interact in ways that are often messy and unpredictable. Traditional math struggles to capture this chaos because the relationships between the water's temperature, its acidity, and the amount of waste it carries are rarely straight lines. To manage these plants effectively, engineers need to know exactly how clean the water coming out will be, but waiting days for lab results can be too slow for making quick decisions. This is where the promise of artificial intelligence enters the picture, offering a way to learn from past data to predict the future with speed and precision.

In a recent study focused on a major treatment facility in Tehran, Iran, researchers set out to test whether computer models could accurately forecast the quality of the water leaving the plant. They gathered a year's worth of real-world data, tracking everything from the amount of organic matter and suspended solids entering the system to the temperature and acidity of the water. The team then built three different types of digital "brains" to analyze this information. The first was an artificial neural network, a system designed to mimic the way human neurons connect and learn. The second was a fuzzy logic model, which handles uncertainty by using rules that allow for shades of gray rather than strict yes-or-no answers. The third was a hybrid system called an adaptive neuro-fuzzy inference system, which combines the learning speed of the neural network with the rule-based reasoning of fuzzy logic. To see if simplifying the data helped, they also ran these models twice: once using all the raw measurements and once after using a statistical method to strip away redundant information and keep only the most critical factors.

The researchers discovered that all three computer systems were capable of learning the patterns of the treatment plant, but one stood out as significantly more reliable. The hybrid system, which blends neural networks with fuzzy logic, proved to be the most accurate predictor of the water's final quality. When the team compared the computer's guesses against the actual lab results, this hybrid model made fewer mistakes than the others, reducing the average error by nearly 14 percent compared to the fuzzy logic model and by over 8 percent compared to the neural network alone. Interestingly, the study found that feeding the computer fewer, more carefully selected variables did not always improve the results; in fact, for some models, using the full set of data yielded better predictions. This suggests that while simplifying data can be helpful, the complex relationships in wastewater treatment are rich enough that throwing away information might sometimes hide important clues.

Beyond just predicting numbers, the study confirmed that the treatment plant itself is performing exceptionally well. By calculating how much pollution was removed, the researchers found that the facility eliminates about 96.68 percent of the total suspended solids, which are the tiny solid particles floating in the water. The removal rates for other pollutants, such as the organic matter that consumes oxygen in rivers, were nearly as high. The computer models were able to replicate this high level of efficiency with remarkable closeness, proving that they can serve as a trustworthy tool for operators. Instead of waiting for slow laboratory tests, plant managers could potentially use these models to get an immediate, accurate estimate of how clean the water is, allowing them to adjust the treatment process in real time. The work demonstrates that artificial intelligence is not just a theoretical concept but a practical, efficient, and reliable method for ensuring that the water we reuse is safe for the land and the people who depend on it.

Technical Summary: Analysis of Sewage Treatment Plant's Data to Evaluate Quality of Effluent using Experimental Results and Prediction based on Artificial Intelligence Models

Problem Statement
The increasing global demand for water, driven by population growth and limited resources, necessitates the reliable reuse of treated wastewater for agricultural irrigation. However, modeling wastewater treatment processes is inherently challenging due to the non-linear nature of treatment operations and the variability of influent characteristics (e.g., flow rates, organic/inorganic compound types). Traditional linear mathematical models often rely on hypotheses and probabilities, providing only estimates rather than precise predictions. Consequently, there is a need for advanced, non-linear modeling techniques that can accurately predict effluent quality to facilitate principled planning and operational control without the high costs associated with extensive long-term experimental measurements.

Methodology
This study utilized data from the Varamin wastewater treatment plant in Tehran, Iran, which employs an Extended Aeration Activated Sludge (EAAS) process with nitrogen removal. The research focused on predicting the quality of the effluent based on specific input parameters.

Data Collection and Pre-processing:
- Inputs: One year of data was analyzed, including influent parameters (Temperature, pH, BOD, COD, TSS) and aeration basin parameters (MLSS, MLVSS, SVI).
- Outputs: Effluent concentrations of BOD, COD, and TSS.
- Data Processing: Outliers and insignificant data were removed. Principal Component Analysis (PCA) was employed to reduce the dimensionality of the input variables, aiming to identify key components and improve model accuracy while reducing computational complexity.
Model Development:
Three Artificial Intelligence (AI) models were developed and compared:
- Artificial Neural Network (ANN): A Feed-Forward Neural Network (FFNN) using a Multi-Layer Perceptron (MLP) structure with a sigmoid transfer function and the Levenberg-Marquardt training algorithm.
- Fuzzy Logic (FL): A Mamdani-type fuzzy inference system using triangular and trapezoidal membership functions.
- Adaptive Neuro-Fuzzy Inference System (ANFIS): A hybrid system combining neural networks and fuzzy logic (Sugeno type) to adaptively learn from training data.
Experimental Design:
Three distinct scenarios were modeled for each output parameter (BOD, COD, TSS). Each scenario was tested in two configurations:
- Without PCA: Using all 11 original input variables.
- With PCA: Using 4 reduced input variables derived from the PCA method.
  Data was split into training (70%), testing (15%), and validation (15%) sets.
Performance Evaluation:
Model accuracy was assessed using statistical indicators: Correlation Coefficient ( $R$ ), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). Additionally, a "Productivity Index" (Removal Efficiency) was calculated to compare the models' ability to simulate contaminant reduction against actual plant performance.

Key Results

Model Performance: The ANFIS model demonstrated superior performance compared to both ANN and FL models across all scenarios.
- Accuracy: ANFIS achieved the highest correlation coefficients and the lowest error rates. Specifically, the ANFIS model showed a mean error reduction of 13.92% compared to the FL model and 8.22% compared to the ANN model.
- PCA Impact: The application of PCA yielded mixed results. While it improved the accuracy of the ANN model in the first and third scenarios (specifically for BOD and TSS), it negatively affected the performance of other models and parameters, such as increasing the RMSE for COD in the ANN model. The authors note in the conclusions that the use of PCA "negatively affected the prediction performance of some models with an average error of 3.615%," indicating that while predictive capability remained high, the dimensionality reduction did not universally enhance accuracy across all metrics.
Removal Efficiency: The study calculated the removal efficiency for the pollutants. The actual plant achieved a maximum removal efficiency of 96.68% for Total Suspended Solids (TSS). The AI models closely mirrored these results, with ANFIS estimating a TSS removal efficiency of 96.75% (without PCA) and 96.61% (with PCA).
Comparative Analysis: While all three models (ANN, FL, ANFIS) were deemed capable of predicting effluent quality, ANFIS consistently offered the best balance of training speed and prediction accuracy. FL models were noted to be time-consuming in training, and while ANFIS offers speed, it can become complex if the number of inputs is too high due to the exponential increase in fuzzy rules.

Significance and Claims
The paper concludes that Artificial Intelligence Network (AIN) models are a simple, accurate, efficient, and reliable approach for predicting the quality of effluents from municipal wastewater treatment plants.

Operational Utility: The study asserts that these models can serve as effective tools for operators to understand and simulate the non-linear behavior of treatment plants, aiding in decision-making for water reuse and environmental protection.
Methodological Value: The research highlights that while multivariate statistical methods like PCA can reduce input variables, their impact on final prediction accuracy is variable and can sometimes degrade performance (as seen with the average error increase). However, the core finding is that AI models, particularly ANFIS, maintain high predictive validity even with reduced input sets.
Reliability: The close alignment between the models' predicted removal efficiencies and the actual plant performance (e.g., ~96% for TSS) validates the use of these AI tools as substitutes for or complements to traditional monitoring, potentially saving time and resources associated with manual testing (such as the 5-day BOD5 measurement).

The authors emphasize that the selection of a specific model (ANN, FL, or ANFIS) should depend on the specific objectives of the study, but for the purpose of high-accuracy prediction of effluent quality in this context, ANFIS is the most appropriate choice.

More like this