ANNs for Air Quality Assessment and Pollution Forecasting — Paper Overview
Using Artificial Neural Networks for Air Quality Assessment and Pollution Forecasting
Introduction
In recent years, air pollution has become a major concern for highly populated areas around the world. According to The World Health Organization’s (WHO)[1] 2016 estimates, air pollution in cities and rural areas causes 4.2 million premature deaths each year. In 2019, 99% of the world’s population lived in places where air quality levels exceeded the WHO [1] guideline limits. As this situation worsens, researchers are finding ways to mitigate the negative impacts caused by certain pollutants.
This article is inspired by this paper: [2] Air quality assessment and pollution forecasting using artificial neural networks in Metropolitan Lima-Peru. I will be drawing on concepts from this paper, focusing on how ANNs can be useful for air quality assessment/pollution forecasting and how researchers use statistical analysis to supplement their results.
Concepts from the Study
Focus: The study focused on predicting PM10 levels in Lima, Peru, a city with high levels of air pollution in South America. [4] PM10 is a pollutant that has significant negative impacts on public health. With factors such as rapid population growth, high levels of industrial activity, and a lack of previous research in Lima, there is a need for advanced methods to address this issue. To address this, the researchers used multilayer perceptron (MLP) and long short-term memory (LSTM) models to forecast PM10 values. (Models covered in the next few sections)
Data Collection: Hourly data was collected from 5 meteorological and air quality monitoring stations in Lima, covering 2 years.
Goal: Predict one hour ahead of PM10 concentrations using MLP and LSTM neural network models.
What are ANNs?
Artificial Neural Networks (ANN) are a subset of machine learning and their structure is inspired by the study of brain architecture. ANN is a model that is made up of layers of interconnected “neurons”. Each of the neurons processes and transmits information, allowing the network to learn how to make intelligent decisions.
[3] Each of the nodes has an associated weight and if an output of a node exceeds a certain threshold value, the node is activated and sends data to the next node; otherwise, no data would be passed to the next layer. These neural networks rely on training data to improve accuracy over time and are then used for various AI/ML tasks.
There are many different types of ANNs, including feedforward networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short-term neural networks (LSTMs). Each type is usually suited for different types of tasks. ANNs are well-suited for most prediction and forecasting tasks because of their ability to handle large amounts of data and non-linear relationships within data.
Multilayer Perceptron (MLP) Model
A multilayer perceptron model is a type of neural network that falls under “feedforward” networks. A feedforward network implies that information is passed through the network in one direction from the input to the output layer. It consists of multiple layers of neurons, or “perceptrons”. This is similar to the diagram pictured earlier in the article.
So, how can MLP differ from any other neural network? One way they are in their architecture and the type of learning they perform. MLPs are supervised learning models and are “feedforward” networks, as I mentioned earlier. This differs from recurrent neural networks (RNNs) in which information is processed in directed cycles instead of in one direction. MLPs may also differ from other neural network models in the type of activation function they use. The activation function (sigmoid, tanh, leaky reLU) determines the output of a neuron given its input.
Long Short-Term Memory (LSTM) Model
LSTM is a type of RNN that is well-suited to learning long sequences of data. As mentioned in the previous paragraph, RNNs process data by using looping connections between units (hence “recurrent”) instead of in one direction like MLP information processing. Because of their looping connections, RNNs (including LSTMs) are considered to be feedback networks. LSTMs are better at capturing long-term dependencies in the data than traditional RNNs.
Highlights from the Study
To understand the complex factors at play in the study, the researchers conducted statistical analyses of PM10 levels, meteorological variables, and air pollution forecasting models. They gathered results through the following methods:
Descriptive Analysis of PM10
The researchers graphed histograms that show the behavior of PM10 in every season. They found that in the probability distribution, the pollutant concentrations are skewed to the right, indicating critical contamination levels. All stations surpass the PM10 limits set by the WHO.
Analysis of Correlations of Meteorological Variables:
A moderate positive correlation (0.39) was found between temperature and PM10 and a moderate negative correlation (-0.38) between relative humidity and PM10.
It was also found that the stations located in the highest area of the city have the highest concentration of PM10, and vice versa. This doesn’t imply a significant correlation, but mostly meteorological influence.
Characteristics of the Most Contaminated Station:
The station with the highest levels of PM10 contamination is known as “HCH.” This area is characterized by high levels of vehicular traffic and traffic congestion, poorly maintained roads, high levels of industrial activity, and a shortage of green spaces. Vehicle traffic is found to contribute 21% to PM10 concentrations, while the dust from poorly maintained roads contributes 54% to air pollution. Industrial activity is also directly linked to increased PM10 generation, with HCH brick industries lacking the necessary technology to mitigate air pollution.
Results from the Air Pollution Forecasting Models (MLP and LSTM):
In the study, two methods were used to evaluate the performance of machine learning models for predicting PM10 trends: hold-out (HO) and blocked-nested cross-validation (BNCV). The HO method involved dividing the data into training, validation, and testing subsets, while the BNCV method involved training and testing the model using a fixed window that moved forward in time, with the model being iteratively retrained on all the data up until the current point in time. The researchers found that the models evaluated using the BNCV method showed slightly better performance than those evaluated with the HO method. Both the multi-layer perceptron (MLP) and long short-term memory (LSTM) models were able to accurately learn and predict PM10 trends and predict the regular component of PM10 contamination levels, but their performance was impacted when stations were subject to unpredictable external sources of pollution.
Evaluation:
4 performance metrics were used in evaluating the forecasting ability of the models: mean absolute error, root mean squared error, symmetric mean absolute percentage error, and spearman’s rank correlation coefficient.
Takeaways
- Artificial neural networks (ANNs) are a type of machine learning model that can be used for air quality assessment and pollution forecasting.
- There are several types of ANNs, including feedforward networks, convolutional neural networks, recurrent neural networks, and long short-term memory networks.
- MLP and LSTM models were used to predict PM10 concentrations one hour in advance.
- Deep learning neural networks like the LSTM have the potential in aiding public policies that prioritize improving air quality and building more sustainable cities.
- While both the MLP and LSTM models predicted accurately, the study found that the LSTM model with the BNCV method would be better adapted to data from monitoring stations in the cases of extreme values.
Sources
[1] Calidad del aire ambiente (exterior) y salud. (2021). Retrieved 6 January 2023, from https://www.who.int/es/news-room/fact-sheets/detail/ambient-(outdoor)-air-quality-and-health
[2] Cordova, C., Portocarrero, M., Salas, R., Torres, R., Rodrigues, P., & López-Gonzales, J. (2021). Air quality assessment and pollution forecasting using artificial neural networks in Metropolitan Lima-Peru. Scientific Reports, 11(1). doi: 10.1038/s41598–021–03650–9
[3] What are Neural Networks? | IBM. (2023). Retrieved 6 January 2023, from https://www.ibm.com/topics/neural-networks
[4] Integrated Science Assessment (ISA) for Particulate Matter (Final Report, Dec 2009) | Risk Assessment Portal | US EPA. (2023). Retrieved 6 January 2023, from https://cfpub.epa.gov/ncea/risk/recordisplay.cfm?deid=216546
