Project For "Advanced Machine Learning Subject"

Air Quality Prediction with LSTM

Target:

The target of this application is to provide accurate air quality predictions, enabling individuals to minimize exposure to harmful pollutants and make informed decisions for health and environmental sustainability.

Motivation:

Air quality directly impacts human health, environmental sustainability, and overall quality of life. With the increasing levels of urbanization, industrialization, and climate change, poor air quality has become a pressing global challenge. People are more aware than ever of the need to monitor air quality to make informed decisions that can protect their health and the environment.

An air quality prediction application addresses this need by providing accurate and timely forecasts of air pollution levels. It empowers individuals, communities, and policymakers with actionable insights. For example:

• Health Protection: Air pollution is linked to respiratory diseases, cardiovascular problems, and other severe health conditions. A prediction system helps users plan their outdoor activities to minimize exposure during high-risk periods.
• Environmental Awareness: By understanding patterns of air pollution, individuals and communities can take steps to reduce emissions and advocate for greener policies.
• Data-Driven Decision-Making: Industries and governments can leverage predictive insights to implement proactive measures, such as traffic restrictions, emission controls, or public advisories, before pollution levels exceed safe thresholds.
• Climate Change Mitigation: Tracking air quality data over time helps in understanding trends, identifying pollution hotspots, and devising long-term strategies to mitigate environmental impacts.

Application:

This application will allow users to view detailed charts of pollutant levels for several cities across the United States, using data sourced from the Environmental Protection Agency (EPA). These visualizations will help users track air quality trends, compare pollution levels across locations, and gain insights into environmental conditions.

Results

Air Quality Factors Monitored in Major U.S. Cities

Cities:

Denver, Los Angeles, Phoenix, Pittsburgh, Seattle

---

Source of data:
United States Environmental Protection Agency

Key Air Quality Indicators

Factors taken into consideration during prediction process can be seen in the table below:

Factor	Unit	Source	Health Impact	Air Quality Standard (USA)	Environmental Impact
CO (Carbon Monoxide)	Parts per million (ppm)	Fossil fuel combustion (vehicles, furnaces, power plants)	Reduces oxygen delivery in the body, leading to hypoxia	9 ppm (8-hour average); 35 ppm (1-hour average)	Low direct effect, may indirectly contribute to ground-level ozone
NO₂ (Nitrogen Dioxide)	Parts per billion (ppb)	Vehicle emissions, fuel combustion in power plants and industry	Irritates respiratory pathways, increases asthma risk, and lung infections	53 ppb (annual average); 100 ppb (1-hour average)	Contributes to photochemical smog and acid rain formation
O₃ (Ozone)	Parts per million (ppm)	Formed secondary to reactions between NOₓ and VOCs under sunlight	Irritates respiratory pathways, causes coughing, and impairs lung function	0.070 ppm (8-hour average)	Toxic to plants, reduces agricultural yields
PM2.5 (Particulate Matter PM2.5)	Micrograms per cubic meter (µg/m³)	Fossil fuel combustion, industry, vehicle emissions, natural events (e.g., wildfires)	Penetrates deep into lungs and bloodstream, increasing risks of heart, respiratory diseases, and cancers	12 µg/m³ (annual average); 35 µg/m³ (24-hour average)	Reduces visibility, harms aquatic and terrestrial ecosystems
SO₂ (Sulfur Dioxide)	Parts per billion (ppb)	Combustion of sulfur-rich fossil fuels (coal), oil refining, volcanic activity	Irritates respiratory pathways, triggers asthma attacks	75 ppb (1-hour average)	Major contributor to acid rain, damaging plants, soils, and water bodies

USA Regulations

European Air Quality Standards (for reference):

• CO: 10 ppm (8-hour average)
• NO₂: Annual average ≤ 40 µg/m³ (~21 ppb)
• O₃: Maximum 8-hour concentration ≤ 120 µg/m³ (~0.06 ppm)
• PM2.5: Annual average ≤ 25 µg/m³ (WHO recommends 5 µg/m³)
• SO₂: 1-hour concentration ≤ 350 µg/m³ (~132 ppb)

EU Regulations

Regulations for O₃ and SO₂ are much stricter in the USA than in EU. But on the other hand CO, NO₂ and PM2.5 are much stricter in EU than in the USA.

Model structure chosen

LSTM (Long Short-Term Memory) networks are a type of Recurrent Neural Network (RNN) specifically designed to address the challenges associated with learning from time-series data. That is why this type of model is perfect fit (also widely used) in Air Quality analysis.
Air quality data typically consists of time-series sets, where the measurements (CO, NO₂, O₃, PM2.5, SO₂) at each time step depend on previous observations. Unlike traditional RNNs, LSTM networks have specialized memory cells that allow them to remember information over long periods.
Also vanishing gradient is an important issue to address. It's an issue where gradients become too small during backpropagation process, making it difficult to learn long-range dependencies.
Preventing from occurring gradient problems is essential for good air quality prediction, where seasonal patterns, weather conditions, and pollution trends evolve over given periods of time. This is perfectly solved by LSTM architecture, which should improve prediction accuracy.

Parameter	Description	Example
input_size	The number of features in the input data.	1
hidden_size	The number of units in the LSTM hidden layer.	16
num_layers	The number of LSTM layers.	2
output_size	The number of units in the output layer.	1
dropout	The proportion of LSTM units that are randomly dropped during training to prevent overfitting.	0.2
batch_first	Specifies if the input and output tensors should be in the format (batch_size, seq_length, input_size).	True
learning_rate	The learning rate for model training.	0.001
optimizer	The optimization algorithm used for updating weights.	Adam
epochs	The number of training epochs (passes through the entire dataset).	10,15,25,35,50
sequence_length	The number of time steps (data points) the model takes as input.	3

Predictions

All used datasets can be found at the following path:

stronka/Datasets/*.csv

The results comparing the actual state with the predicted state:

stronka/Datasets/graphs/*actual_vs_predicted.png

The training strategy, i.e., the most important indicator that informs how well the model fits the training data, can be found at the following path:

stronka/Datasets/graphs/*training_loss.png

Sample visualisation of results for prediction of Atlanta city:

• CO results

• NO₂ results

• O₃ results

• PM2.5 results

• SO₂ results

Overall, the results look good. The model accurately captures trends in the predicted data. The quality and accuracy of the fitting in characteristic points, such as local extremes on the air quality index chart, could be improved. The model predicted results within the date range:

[2023-09-15 : 2024-01-01]

The prediction was made using purely historical data, but based on the following approach: we have data from several past years, for example, 3 years, i.e., the date range from [X, Y]. We train the data on the period let's say of 2 years: [X, X+2 years]. Then predict the data for the period let's say of 1 year: [Y, Y+1 year]. This approach allows us to assess how the model performs in real-world conditions.
Real-time predictions would require integration with a large amount of quality data from sensors or possibly some commercial API in a paid model.
Additionally, computational resources would need to be increased for such a model to work in real-time. Currently, the model runs on a local CPU:

model.load_state_dict(torch.load('./stronka/PT/O3_PT_lstm_model.pth', map_location=torch.device('cpu')))

However, scaling it to a very high computational performance will not be difficult.

App

The application encapsulates the model's functionality and the ability to predict air quality in the form of a convenient interface and a simple API. The application is built using Flask and Vanilla JS. Sample screenshots from user interface:

:)

Project Executors

• Mateusz Potaśnik
• Szymon Rogowski
• Bartosz Szymański
• Bartosz Pękała