Indoor Air Quality Index Prediction Using Machine Learning Algorithms
Abstract
Indoor air pollution remains under-monitored despite its significant health implications. This project proposes a
machine learning-based Indoor Air Quality Index (IAQI) prediction system using real-time data from low-cost
sensors integrated with IoT technology. The system will stream live environmental data to a cloud platform,
where an adaptive LSTM model will process it to forecast AQI levels. The focus is on enhancing prediction
accuracy through dynamic feature selection and model interpretability. The outcome will be a scalable, user-
friendly IAQI system supporting timely indoor environmental decision-making.
Project Background (Significance/Justification)
Air quality monitoring in indoor environments has not kept pace with technological advances, even though
exposure to indoor pollutants is linked to serious health concerns. Conventional AQI systems are designed for
outdoor monitoring and cannot capture spatial and temporal variations indoors (Shahsavari et al., 2021). The
rise of IoT-based sensor networks has introduced new opportunities, but their application in indoor air quality
prediction is still limited.
Several studies highlight the accuracy of ML algorithms like Random Forest and LSTM for AQI prediction, yet
most are trained on outdoor datasets and are not optimized for indoor environments, which are affected by
irregular activities like cooking, cleaning, or opening windows (Park et al., 2022; Li et al., 2019). Moreover, the
absence of real-time implementation and adaptive models has been a major barrier to practical deployment
(Zhang et al., 2022).
Another challenge is the reliability of low-cost sensors, which suffer from calibration drift and environmental
interference (Castell et al., 2017). While hybrid compensation models have been proposed, they are rarely
integrated with real-time ML pipelines in a scalable, user-centered way.
This project aims to fill these gaps by creating a real-time, adaptive AQI prediction system tailored specifically
for indoor use—an area currently underdeveloped in both research and application.
Project Aim
To develop a real-time, adaptive indoor air quality prediction system that integrates low-cost IoT sensors with
machine learning algorithms to enable accurate, interpretable, and scalable Indoor Air Quality Index (IAQI)
forecasting.
Project Objectives
1. Design and implement a multi-sensor IoT system using devices such as PMS5003 and MQ135 with
ESP32 microcontroller to collect real-time indoor air pollutant and environmental data.
2. Develop and train a machine learning model—specifically an LSTM-based time-series predictor—
tailored to handle indoor air quality dynamics and sensor variability.
3. Integrate the sensor system with a real-time cloud-based data pipeline to enable continuous data
logging, preprocessing, and ML inference for live IAQI prediction.
4. Evaluate and optimize system performance in terms of prediction accuracy, adaptability to
environmental changes, and model interpretability through testing in simulated and real indoor conditions.
Milestones and Tasks
1. Designing the Indoor AQI Monitoring Framework
I. Conduct a review of pollutant-specific indoor monitoring requirements based on WHO and EPA standards.
II. Select suitable low-cost sensors (e.g., PMS5003, MQ135, DHT11) and microcontroller (ESP32) for indoor
deployment.
III. Design the sensor layout and data acquisition flow for real-time monitoring.
2. Implementing Sensor Data Collection and Transmission Pipeline
I. Develop and test Arduino/ESP32-based firmware to collect and transmit sensor data via Wi-Fi.
II. Establish cloud integration (e.g., Firebase, ThingsBoard, or AWS IoT) for real-time data logging.
III. Calibrate sensors in controlled conditions and apply basic environmental compensation techniques.
3. Developing and Training the Machine Learning Model
I. Preprocess collected sensor data (filtering, normalization, and feature engineering).
II. Train LSTM and baseline models (e.g., Random Forest) on labeled AQI data to predict indoor AQI.
III. Evaluate model performance using accuracy metrics (e.g., RMSE, MAE) and real-time responsiveness.
IV. Optimize model for computational efficiency on edge/cloud systems.
4. Integrating and Testing the Real-Time IAQI Prediction System
I. Deploy ML model and integrate it with the live sensor pipeline.
II. Validate predictions against known IAQI levels or simulated pollutant scenarios.
III. Conduct tests in different indoor environments (e.g., kitchen, bedroom, lab).
IV. Assess system interpretability, user feedback potential, and scalability.
V. Prepare final documentation and reflection on system performance.
Project Timeline
Experimental Strategy and Approach
This project aims to design and implement a real-time indoor air quality prediction system by integrating low-
cost IoT sensors with an adaptive machine learning (ML) model. The strategy is broken down into distinct
experimental layers corresponding to the major milestones and objectives outlined earlier.
1. Sensor Network Design and Implementation
To capture relevant indoor air pollutants and environmental variables, the system will use a combination of
PMS5003 (for PM2.5 and PM10), MQ135 (for CO₂, NH₃, benzene), and DHT11 (for humidity and temperature)
sensors. These will be interfaced with an ESP32 microcontroller programmed using the Arduino IDE. The
ESP32 is selected for its low power consumption and built-in Wi-Fi capability, enabling real-time data
transmission.
Calibration will be performed under controlled indoor conditions using standardized pollutant concentrations
and cross-verification against known AQI data from reliable indoor monitors. The sensors will be tested for
noise, stability, and responsiveness using ambient stimuli such as incense, cleaning sprays, or stove emissions to
simulate real conditions.
2. Real-Time Data Logging and Cloud Integration
Sensor readings will be streamed via MQTT or HTTP protocol to a cloud-based platform (e.g., Firebase, AWS
IoT, or ThingsBoard). Each reading will be time-stamped, tagged by location, and stored in JSON format for
consistency. A [Link] or Python-based backend will handle basic preprocessing, including outlier removal,
missing data imputation, and normalization.
This infrastructure will simulate a realistic real-time deployment environment and will also serve as the live data
pipeline for feeding the ML model during inference.
3. Machine Learning Model Development and Training
The core predictive engine will be an LSTM (Long Short-Term Memory) network, chosen for its effectiveness
in modeling time-series data with temporal dependencies. The model will be trained on a dataset constructed
from historical indoor readings collected over multiple days, annotated with AQI classifications based on WHO
or EPA standards.
Data preprocessing will involve:
Rolling window averaging to smooth sensor readings.
Feature selection based on pollutant correlations with AQI.
Conversion of sensor values into derived AQI subindices using the EPA breakpoint formula.
Training will be done in Python using TensorFlow or PyTorch. A baseline Random Forest model will also be
implemented for comparison. Metrics such as RMSE, MAE, and classification accuracy will be used to evaluate
model performance. Hyperparameter tuning will be performed via grid search or Bayesian optimization.
4. Deployment, Inference, and Evaluation
The trained LSTM model will be deployed either on the cloud or at the edge (on a Raspberry Pi or ESP32-
compatible module with TensorFlow Lite support). A lightweight dashboard will be created to display live AQI
levels and model confidence.
To test real-world adaptability, the system will be deployed in at least two different indoor environments (e.g.,
kitchen vs. study room) with contrasting pollution profiles. Model adaptability will be evaluated based on
prediction drift, response time to sudden environmental changes, and system stability.
A user-facing component (e.g., mobile alert system or color-coded dashboard) will be included to assess the
human-centric aspect of the system.
Expected Outcomes
1. Development of a real-time indoor AQI prediction prototype that integrates low-cost sensors with an
LSTM-based machine learning model, optimized for indoor environmental dynamics (Park et al., 2022;
Zhang et al., 2022).
2. Enhanced model accuracy and interpretability through dynamic feature selection and real-time
adaptation to sensor drift and environmental variability, contributing to more trustworthy and actionable
outputs (Molnar, 2020; Xie et al., 2019).
3. Deployment of a cloud-connected monitoring system with live AQI visualization and alert capabilities,
suitable for home, office, or classroom applications, supporting data-driven health decisions (Kumar et
al., 2019).
4. Validation of sensor performance and system robustness in real-world indoor settings with varying
pollutant levels, demonstrating the feasibility of cost-effective, scalable IAQ solutions.
5. Contribution to research on adaptive and human-centric AQI systems, providing insights into
challenges of real-time deployment, data noise mitigation, and end-user usability in ML-powered
environmental monitoring tools.
References
Tan, H., Othman, M.H.D., Kek, H.Y., Chong, W.T., Nyakuma, B.B., Abdul Wahab, R., Hoh Teck, G.L. & Wong,
K.Y., 2024. Revolutionizing indoor air quality monitoring through IoT innovations: A comprehensive systematic
review and bibliometric analysis. Environmental Science and Pollution Research, 31, pp.44463–44488.
Yu, W., Nakisa, B., Loke, S.W., Stevanovic, S., Guo, Y. & Rastgoo, M.N., 2024. Indoor PM2.5 forecasting and
the association with outdoor air pollution: A modelling study based on sensor data in Australia. arXiv preprint
arXiv:2405.07404
Yee, J.J., Yousaf, J. & Harseno, R.W., 2024. Next-generation indoor air quality management: An integrated IoT-
and deep learning-based approach for real-time monitoring and prediction. SSRN Electronic Journal.
Berkani, M.R.A., Chouchane, A., Himeur, Y., Ouamane, A. & Amira, A., 2023. An intelligent edge-deployable
indoor air quality monitoring and activity recognition approach. arXiv preprint arXiv:2311.03920.
Mohammadshirazi, A., Nadafian, A., Monsefi, A.K., Rafiei, M.H. & Ramnath, R., 2023. Novel physics-based
machine-learning models for indoor air quality approximations. arXiv preprint arXiv:2308.01438.