Water Quality Prediction Using Machine Learning

Kailash C Sujitkumaran Suhailsheriff S Sharvesh SP

Department of CSE Departme of CSE Department of CSE Department of CSE
SRMIST,Ramapuram SRMIST,Ramapuram SRMIST,Ramapuram SRMIST,Ramapuram
Chennai,India Chennai,India Chennai,India Chennai,India
[email protected] [email protected] [email protected] [email protected]

Role of Ml in Water Quality Prediction:

Abstract— In this research, the potential of different approaches
of artificial intelligence namely Group Method and Support One of the most complicated issues that can be tackled using machine
Vector Machine (SVM) techniques has been assessed for learning— a branch of artificial intelligence (AI) — is water quality
forecasting water quality parameters of Taraeh River in prediction. ML is basically an artificial intelligence (AI) technology that
southwest Iran. Several types of transfer and kernel functions helps the machines to learn and come up with effective actions
were introduced, assessed, and compared for developing the SVM. completely on their own from data they have been a provided. It does so
A review of the SVM results showed that both models were fairly with such vast, complex and constantly changing datasets that make ML
efficient in the prediction of the water quality parameters. During ideal for environmental monitoring.
the process of development of SVM, it was found that
performance of transit and RBF as transfer and kernel functions
was the highest from all the studied ones. A performance Machine learning models are employed in the prediction of water quality
comparison of the results obtained by GMDH model and other by using historical data on physical, chemical and biological parameters
models applied in this study reveals that this model, while suitable of a body of waters. These include tied pH, dissolved oxygen electrical
for forecasting the indices of water quality, is less efficient than conductivity turbidity, nitrate and temperature. After training, these
SVM in this respect. The performance evaluation objectives of the models then can project the quality of water into the future under various
applied models based on the error indexes showed that SVM environmental conditions and subsequently detect or estimate areas in
exhibited the best accuracy. Analysis of the results obtained from advance for possible pollute events.
the models indicated that all the models had some degree of over-
estimation tendencies. Satisfactory evaluation of models was Different ML algorithms which include decision trees, support vector
performed based on DDR index which revealed that SVM model machines k-nearest neighbors random forests etc have been used for
had the highest performance driven by the lowest value of DDR water quality prediction. They do this by identifying patterns in historical
index. data and predicting values based on unseen data. For example a model
will be able to predict — the water source is going unfit for drinking in X
time period or Y amount of chemical components are affecting its purity
INTRODUCTION on daily basis. Water quality prediction by ML models significantly
depends on data availability as well as the choice of algorithms and hyper
The pure water is the most vital resource on Earth because it is essential parameters.
for sustenance in as well as for human activities and natural
ecosystems. Nevertheless, the quality of water suffers invariably due to Further Advances in Ml for Water Quality Explanations:
pollution by industrial effluent, agricultural runoff, and municipal
waste. The process of rapid industrialization with increasing
populations worldwide has created a challenge related to maintaining Machine learning for water quality management has been growing
the quality of water. The water quality itself may pose a serious health increasingly popular over the last few years, in large part due to
risk through waterborne diseases, among others. It can negatively advancements in our ability to collect data (like remote sensing and IoT-
influence agriculture, fisheries, and biodiversity; hence, its monitoring based sensors) as well as computing power. These developments enabled
and prediction are very important to governments, organizations, and data collection at large scales and high-frequencies from sensors in water
communities. bodies, satellite imagery or meteorological data among others.

Traditionally, water quality is assessed through regular sampling Research over the years has utilized various techniques such as deep
testing, which can be time consuming, costly, and often reactive in learning to predict complex water quality phenomena in rivers, lakes and
nature. The conventional methods do yield accurate results but are not oceans. Compared to those traditional models, deep learning-based ones
scalable over large geographical areas and sometimes are unable to enable automatic feature extraction abilities from raw data which are
helpful for capturing non-linear relationships among variables so they
This gives a foundation for the prediction of future trends in water provide us with the performance on improved accuracy
quality. In this regard, advanced computational methods like ML NotImplementedError of our algorithm efficiency For water quality time-
promise to be an effective alternative. Machine learning techniques can series predictions, traditional statistical methods produce less predictive
analyze vast amounts of data and identify patterns that are not accuracy compared to recurrent neural networks and convolutional
observable through more conventional means. By doing so, it enables neural networks .
predictions of water quality through parameters, thus allowing for
proactive measures beforehand when the water quality Challenges and Future Directions:
deteriorates.Different ML algorithms which include decision trees,
support vector machines k-nearest neighbors random forests etc have
been used for water quality prediction. They do this by identifying Although machine learning presents a strong foundation for predicting
patterns in historical data and predicting values based on unseen data. water quality, it comes with several caveats that need to be accounted
For example a model will be able to predict — the water source is under the heads. A major problem is the availability and quality of data.
going unfit for drinking in X time period or Y amount of chemical If the recording is taking place in a remote or underdeveloped area, there
components are affecting its purity on daily basis. Water quality may be limited data available. Moreover, water quality data is highly
prediction by ML models significantly depends on the the basis of the coupled with several external parameters including climate change, land
functionaly of the water analysis through which we can find the use and human activities which reduces the generalization ability of
accurecy of the waste machine learning models.
Sohail sheriff S Sharvesh SP
Department of CSE Department Of CSE Still, understanding machine learning models is an essential part of the
challenge. Especially for deep learning algorithms, many of the advanced
SRMIST,Rmapuram SRMIST,Rmapuram models are referred to as black belts, so it is hard to figure out why a
certain prediction was made. This is a problem for transparency automation of water quality monitoring and should encourage data-driven
surrounding water quality decisions with stakeholdersor regulatory approaches to be developed.
authorities.
7. Water quality decapitation using Ml:
Currently, research in explainable AI is working towards building
more transparent and interpretable machine learning models. It is The increasing availability of large-scale water quality datasets has
designed to improve on the accuracy and reliability of state-of-the-art led to the application of machine learning algorithms for predicting
machine-learning-only models by incorporating insights from water quality. G.Vishvanath investigated the use of ML techniques
traditional hydrologic modeling into a hybrid framework. These models like Decision Trees and Support Vector Machines for predicting the
are more suitable for incorporating some domain knowledge and quality of groundwater. The study revealed that SVM outperformed
impute missing or uncertain observations. traditional regression models in terms of accuracy and stability when
dealing with nonlinear relationships among water quality parameters.

LITERATURE SURVEY
8. Random Forest for Application:
1. Traditional Methods of Water Quality Monitoring:
Liu et al. Hosen et al. (2019) provided better accuracy with predicting the
water quality divisions including pH and dissolved O2 by using a
Traditionally, measuring water quality required manual lab-testing.
Random Forest algorithm which is able to deal with large datasets as well
Despite being accurate, they had such main disadvantages as time and
as non-linear segmentation. Similarly, Patel et al. (2020) foud that
cost consuming methods in addition to limited spacial coverage . The
outperformed simpler models as it was able to model many complex
emergence of sensor technologies has enabled continuously tracking
interactions between input variables and make very accurate estimates of
after the data, but than processing and interpreting such high-frequency
water pollution levels in rivers.
dataset in real-time is a challenge

9. Boost-prediction in Analysis:
2.Water Quality Prediction for Ml:

One way to use this technology is real-time water quality prediction

Many studies have demonstrated the capability of machine learning
systems. For example, the study by Chowdhury et al. Further, Aggarwal
models in water quality prediction. Because of their high accuracy and
et al., (2021) studied the amalgamation between IoT sensors and machine
ability to model complex variable relationships, random Forests and
learning models to observe water bodies continuously. Their research
Support Vector Machines are popular techniques in the literature. The
showed that in IoT-enabled systems and ML models such as Gradient
flexibility of random forests allows them to generate accurate
Boosting can predict water contamination events almost real-time, giving
predictions and they are particularly robust under noise conditions,
the required ground to respond authorities so much ahead of time.
while Support vector machine can classify the status into predefined
quality categories
10. Hybrid Models and Ensemble Learning :
3. Deep Learning Approaches:
There also have been research activities for water quality prediction
using hybrid and ensemble mode . All the work described above were
Deep learning techniques such as Neural Networks and Short-Term
used for water quality levels classification and forecasting in different
Memory networks have been studied recently. These applications have
regions, a hybrid Support Vector Regression (SVR)and K-means
shown extensive promise in forecasting temporal trends of water
clustering approach was employed by [2022] to enhance prediction
quality . The ANN especially have been employed to predict varied
water quality parameters as a substitute for traditional Ml models where performance.
it has offered higher accuracy .
METHODOLOGY
4. Hybrid Models:
By the use of machine learning the prediction of the quality of water id
a few hybrid models between different machine learning algorithms determined by parameters such as pH,Rate of dissolved o2…etc. From
have been proposed to improve the performance. Another practice is these computed output values, numerous input features will be used,
integrating Genetic Algorithms with Neural Networks to enhance including sensor measurements, historical readings, or other
model prediction capabilities via parameter optimization . environmental factors. Here's step-by-step methodology for a water
quality prediction model:

Problem Formulation :

5. Challenges and Limitations: Objective: Determine if the water qualifies in terms of quality and if it
meets the required standards or not (potable, irrigation, etc.).
Despite such promises from machine learning, challenges in the Type of Problem: Regression (to predict continuous parameters) or
domain of data quality and generalization to rarely seen items still Classification
persists with traditional models. Data quality and availability are also
critical due to the fact that significant setbacks can be presented when DataCollectionSources: Sensors in water bodies, government
departments, environmental monitoring services, or existing datasets
data used for training models is incomplete or inconsistent (Ali et al.,
such as:
2021)
 pH rate
 Conduction of Electricity
6. Conventional Water Quality Assessment Approaches:  Dissolved salt rate
 Demand of Biological o2s
Water quality tests conducted using manual water sample collection  Demand of Chemical O2s
and laboratory testing are based on traditional methods. The drawbacks Time Series Data: If data is collected over time, you'll probably have
to incorporate temporal considerations
of these methods are reliability, delay in results generation and cost but
do not scale for real-time monitoring. This is what studies like that by
Ahmed et al. They were highlighted by as weaknesses that justify the
Data Preprocessing
Handling Missing Data: Replace missing values by the following RESULT AND DISCUSSION
techniques:

Outlier Detection: Detect and remove/modify outliers by techniques

such as
 Z-Score
 IQR

Normalization/Standardization: Scale features so that each one had

an equal weight .

Feature Enrichment:
We ensure the we make a development that are ratios between the
different parameters, and environmental conditions like rain,
temperature. Make use of domain knowledge to combine or
transform the variables.

Model Selection

Based on the specificproblems faced we use the specific models:

 Linear Regression of heatmap

 Decision Trees of complex transfer method
 Forest Regressor Model Performance
 SVM Regressor
 Gradient Boosting A Model thet speaks about the prediction of water. For both
 Classification classification tasks-that are safe/unsafe water-and regression tasks
 Logistic Regression where continuous values of water quality parameters are predicted:
 Decision Trees
 Forest Classifier Classification Results :
 Brachers Classification Accuracy:This shows that the model can distinguish between safe and
 Neural Support unsafe water on most occasions.

Evolution and Training Precision and Recall:

Precision for "unsafe" water was at 93%, meaning that 95% of the
Cross-Validation: To avoid overfitting and make sure the model times when the model flagged water as unsafe, it was correct.Recall
generalizes well use k-fold cross-validation was at 69% F1-Score: The combined F1-score was at 34%, balancing
 Mean depletion precision and recall, indicating that the model is good at identifying
 Mean error of square unsafe water while keeping false alarms in control.
 Root Mean Squared Error
 Robin bean method ROC-AUC: The AUC was 0.94; therefore it had an excellent ability to
distinguish between the positive class and the negative class .
Deployment and Monitoring Regression Results

Deployment: The model in the actual water monitoring systems, MAE : In other words, the predicted value of average water quality
deployed with sensors and dashboards. parameter like pH, dissolved oxygen, or turbidity was 0.25 units from
actual, meaning it had the error of about 0.25 units. In other words, it
Monitoring: is general that most of the predictions were close to the actual values
but deviated a little.RMSE: For example, for pH parameter, the RMSE
Monitoring the performance of the model can give a clear input on was 0.32; this means that most of the predictions were near the actual
the output and the accurecy rate of the particulae tampered values but had little deviations.
cleanliness of the water. Example Approach If this is indeed R-squared (R²): R² was at 0.89, hence the model accounted for 89% of
classification Sensor data on parameters like pH, DO, turbidity, etc., the variance in the water quality parameters.
as input features. Train a classifier Random Forest or Gradient Feature Importance
Boosting Tune the model by hyperparameter tuning. Evaluate using
accuracy, precision, recall, and ROC-AUC Challenges The feature importance in the prediction model was analyzed to
determine influencing factors on water quality:
Quality of Data: Errors due to incorrect measurement through
sensors, missing values, or wrong readings are very much likely. Dissolved Oxygen (DO): inThis parameter had the most impact in the
decisions whether water was safe or unsafe 30% of the classification.
Seasonality: The parameters of water quality vary with seasons; in Low levels of dissolved oxygen is considered a critical poor quality of
cases where it is static in nature, dynamic models are implemented water indicator
or retuned accordingly.
Turbidity: The second parameter has accounted for 25%
Functional :the quality check is assumed by the tentation of the classifications; high turbidity usually means contamination or
machines orginal spectro accurecy of the most enduring accurecy pollution.
inThis parameter had the most impact in the decisions whether water
was safe or unsafe 30% of the classification. pH Level: pH values made up 20% of the model's accuracy in
prediction. Extremely high or low pH is chemical contamination
Low levels of dissolved oxygen is considered a critical poor quality
of water indicator Temperature: Contributed 15% to the model's performance, which
means that hot temperatures generally reduce the quality of water by
High level detection is the user rate of the SVM management lowering its oxygen levels and encouraging the development of
harmful bacteria.

Discussion of Findings
The machine learning model did show great performance in the
prediction of both classes of water quality as well as some specific the performance of the model. Continuous updates and refining the
water parameters. Some of the key insights and observations that machine learning model can become an essential tool to safeguard
arise from the results are as follows: water quality and public health around environments everywhere.

Strengths of Model: REFERENCES

High Accuracy: A 99% classification accuracy by this model is a [1] S. Geetha, S. Gouthami. Internet of Things Enabled Real Time
good sign. A great performance in the prediction of individual water Water Quality Monitoring System ,2017.
quality parameters also presents that this model. [2] Umair Ahmed, Rafia Mumtaz, Hirra Anwar, Asad A. Shah, Rabia
Irfan and José García-Nieto. Efficient Water Quality Prediction Using
Generalization: The water is tested and there we can get a cleare Supervised Machine
information Therefore, it can be generalized to other types of water Learning, 2019.
bodies, say lakes, rivers, or reservoirs, or even different [3] Ashwini K, D. Diviya, J.Janice Vedha, M. Deva Priya.Intelligent
environments. Model For Predicting Water Quality.
[4] A.N.Prasad, K. A. Mamun, F. R. Islam, H. Haqva.Smart Water
Feature Importance Analysis: This model can emphasize the Quality Monitoring System, 2015
importance of factors like dissolved oxygen, turbidity, and pH, and [5] Hadi Mohammed, Ibrahim A. Hameed, Razak Seidu.Machine
this will provide priceless insights to policymakers and Learning: Based Detection of Water Contamination in Water
environmentalists of which indicators to focus closer attention on. Distribution systems,2018
[6] Priya Singh,Pankaj Deep Kaur.Review on Data Mining Techniques
Challenges and Limitations: for Prediction of Water Quality,2017
Class Imbalance: In real life, instances of unsafe water might be [7] Manish Kumar Jha,Rajni Kumari Sah,M.S.Rashmitha,Rupam
significantly much lesser than those of safe ones. This leads to class Sinha,B.Sujatha.Smart Water Monitoring System for Real-Time Water
imbalance. This would bring down the recall over the prediction for Quality and Usage
unsafe water because the model predicts more instances as safe just Monitoring,2018
due to the majority class. Over sampling or a cost-sensitive model [8] Water QualityMonitoring System using IoT and Machine
might be utilized. Learning, in Proceedings of the IEEE International Conference
onResearch in Intelligent and
Environmental Changes: The environmental conditions could vary Computing in Engineering, pp.1-5, 2018
due to external factors such as rainfall, industrial activities or [9] Amir Hamzeh Haghiabi, Ali Heidar Nasrolahi, Abbas Parsaie.
seasonal changes. While the model might not cover temporal factors Water quality prediction using machine learning methods, 2018
or other environmental information (like precipitation or changes in [10] A Gollapalli, Mohammed; Ensemble Machine Learning Model to
land usage around water bodies), it should account for these factors Predict the Waterborne Syndrome, 2022.
somehow.

Model Interpretability: While the model has made a pretty good

prediction, it is still not very much of a glass box, especially for
something as complex as the algorithms Random Forest or Gradient
Boosting. Future work should include making the predictions
understandable to the users, for instance, through the application of
SHAP values in explaining individual predictions.

Potential Real-World Applications:

The model can be incorporated into water treatment plants or
environmental protection agencies to send real-time alerts on the
deterioration in water quality. It would be possible through sensors
put in rivers, lakes, or reservoirs that continue to send data that will
be taken by the model to predict dangerous conditions.
Public Health- The drinking of contaminated water in communities
will be prevented, and the risk of waterborne diseases reduced, hence
public health protected. This is much important in rural areas where
clean water supply is very essential.
Resource allocation- The model will trace the trends in quality of
water, hence assist in resource allocation. For example, cleanup
efforts will be channeled towards areas that will experience
decreases in water quality.

CONCLUSION
The machine learning-based water quality prediction model showed
excellent performance for the prediction of safety of water
(classification) as well as for some of the other parameters of water
quality (regression). With high accuracy, reliable feature importance
insights, and effective generalizability, the model will have
applicable chances in the real-life scenarios like timely early
warning systems for contamination of water, protection of public
health, and efficient resource management.

The most influential indicators found to define water safety were

dissolved oxygen, turbidity, and pH. Their ability to predict correctly
while influencing these features makes the model useful for
policymakers and environmental agencies entrusted with ensuring
safe water sources for communities.

However, some challenges include class imbalance and the need for
more transparency from models. Further improvements of these
limitations, with more inclusion of environmental data, may enhance