Predict Air Quality with Machine Learning

This project follows the Scientific Method. Review the steps before you begin.

Overview

In this project, you will gather air quality data from the EPA website for a location of your choice, preprocess the dataset, and train an LSTM model to predict future AQI values over three different time frames: one week, four weeks, and one year. By comparing the model’s performance across these intervals, you will evaluate whether it is more accurate for short-term or long-term predictions.

1. Gathering Our Data

1. Navigate to the Air Quality Index Daily Values Report resource by the EPA. For the ‘Pollutant’ dropdown, select All AQI Pollutants. 

2. Select a year for the data. Eventually, you will need to download data for five consecutive years (you will return to step 2 later for a different year). 

   1. Tip: Earlier years may have missing data, so it is best to choose more recent years. 

3. Choose a city or county. You can pick your own, but keep in mind that not all locations have complete data.

4. After selecting a location, click Generate Report. The data will be displayed in a table, with missing values represented as dots. If there are numerous missing values (e.g., more than 10 scattered across various columns or entire columns with no data), please select a different location that may have more complete data.

5. Click Download CSV (spreadsheet) to save the data to your computer. Change the year and download again to repeat this for each of the five years (Steps 2-5). 

6. Once you have the data for all five years, you can combine it using either Google Sheets or Microsoft Excel. First, we will add the data into one spreadsheet file. See the video at the beginning of this project for a walk-through of this process 

   1. Google Sheets: 

      1. Navigate to Google Sheets and create a new blank spreadsheet.

      2. For each CSV file, go to File > Import, choose Upload, and select the CSV file. Under Import location, choose Insert new sheet(s). You will need to import each year’s data into a new sheet. 

   2. Microsoft Excel: 

      1. Navigate to the Downloads folder on your computer. 

      2. Right-click on any of the downloaded CSV files and open them in Excel. Then, go to File > Save as and save the file as an .xlsx file. 

7. Next, you will combine all of the spreadsheets into one ‘Master’ sheet. 

   1. Google Sheets: 

      1. Add a new sheet by clicking the ‘+’ at the bottom left. 

      2. Rename it ‘Master’ by right-clicking on the tab and selecting Rename.

      3. Copy the first row (column names) from any of the other sheets and paste it into cell A1 of the ‘Master’ sheet. 

      4. For each of sheets with the year’s data, delete the first row by right-clicking on row 1 and selecting ‘Delete row.’ Then, copy all of the data by holding Ctrl + A then Ctrl + C (Cmd + A and Cmd + C on a MacBook), then paste the data into the the first empty slot on the Master sheet by holding Ctrl + V. (Yes, you will have to do a bit of scrolling). 

      5. If needed, sort the data by right-clicking the first column and selecting Sort Sheet A-Z. 

   2. Microsoft Excel: 

      1. Select Data > Get Data > From File > From Text/CSV. One at a time, choose the other four notebooks. 

      2. Add a new sheet called “Master” by clicking on the ‘+’ icon on the bottom. Rename the sheet by right-clicking on the name and selecting ‘Rename.’ 

      3. For each of sheets with the year’s data, delete the first row by right-clicking on row 1 and selecting ‘Delete row.’ Then, copy all of the data by holding Ctrl + A then Ctrl + C (Cmd + A and Cmd + C on a MacBook), then paste the data into the the first empty slot on the Master sheet by holding Ctrl + V. (Yes, you will have to do a bit of scrolling). 

8. When you are finished creating the ‘Master’ spreadsheet, download the ‘Master’ sheet as a .csv file. 

   1. Google Sheets: 

      1. You can download the data by selecting File > Download > Comma Separated Values (.csv). Rename the file to aqidaily_fiveyears.csv.

   2. Microsoft Excel: 

      1. You can download the data by selecting File > Save as and select the CSV (Comma delimited) (*.csv) option. Rename the file to aqidaily_fiveyears.csv. 

2. Setting Up the Project Environment

1. You will need a Google account. If you do not have one, make one when prompted. 

2. Download the air_quality.ipynb file from Science Buddies. This is the code you will need to process your data. 

3. Within your Google Drive, click on ‘MyDrive,’ then create a new folder and rename it “Air Quality Prediction.” Inside the folder, upload both the air_quality.ipynb file and the aqidaily_fiveyears.csv file you created earlier.

4. Double-click on the air_quality.ipynb file. This should automatically open in Google Colab. 

   1. Read the Troubleshooting Tips and How to Use This Notebook sections. Follow the instructions you find in that section.

   2. Run the blocks under Importing Libraries to ensure you have access to all the functions we will use for this project and the aqidaily_fiveyears.csv data you uploaded to your Drive. We are importing quite a few libraries for this project, so do not worry if this takes a bit to run. 

3. Loading the Data into a Pandas DataFrame

1. (Code Block 3A) Run this code block to create a DataFrame, which is like a table that will be used to load and manipulate the data in the notebook. You will see the data from your .csv file populate in a table below the code block.

4. Preprocessing

1. (Code Block 4A) Run this block to convert the ‘Date’ column from text to a datetime format. This ensures the machine learning model recognizes the data as dates, allowing it to correctly process time-based patterns instead of treating the values as plain text. 

2. (Code Block 4B) Run this block to display the Air Quality Index (AQI) data for the five years you selected. This will give you a complete view of the dataset you have downloaded. 

3. (Code Block 4C) Next, we will drop features that we think will be uninformative for modeling. In this case, we will be dropping ‘Site Name (of Overall AQI)’, ‘Site ID (of Overall AQI)’, ‘Source (of Overall AQI)’, and ‘Main Pollutant.’ We only want to use the pollutants to predict the AQI, so information related to the location is not important. 

4. (Code Block 4D) In this code block, we are filling in the NULL, or blank, values (if any) with forward fill. Forward fill is commonly used in time series data (like predicting future AQI) by filling in NULL values with the last known observation. This technique makes sense in many time series contexts because often, the most recent observation is the best estimate of the next value until a new observation is recorded. Run this code block. 

5. (Code Block 4E) In LSTM, the algorithm relies on scaling data with different values so that it is easier for the LSTM to work with. Click on this link to learn more about why we normalize our data. 

   1. For instance, if you are trying to compare the ages of people and their salaries, the LSTM might focus too much on the salary (because the numbers are bigger) and ignore the age of the people (because the numbers are smaller compared to the salary numbers). 

   2. Normalization helps to put all the variables on the same level so the LSTM can learn from the data more easily. 

   3. As with the previous step, we have provided the code to normalize certain columns from our Pandas DataFrame. Add in the names of the columns that we will be normalizing under the #TODO comment. We will be normalizing the variables ‘CO,’ ‘Ozone,’ ‘PM10,’ ‘PM25,’ and ‘NO2.’

6. (Code Block 4F) This code block defines a helper function that formats the data into sequences that the LSTM can process. Since LSTMs excel at learning sequential patterns, breaking the data into windows of time helps the model learn how each time step in a sequence influences the next one. This technique makes the LSTM capable of predicting future values based on past information. Run this code block.

7. (Code Block 4G) In this code block, we define a variable called WINDOW_SIZE that determines how many consecutive time steps are grouped together into a “window” or sequence. This will be the number of time steps, in other words, how many days, the model will use to predict the next value (e.g., WINDOW_SIZE = 5 means the model will predict the 6th day AQI based on the previous 5 days). Here is where you will be changing the WINDOW_SIZE variable:

   1. Increasing the WINDOW_SIZE means you are using more past time steps in each sequence to predict the next value. This can give the LSTM more context to learn from (capturing longer-term dependencies), but it also increases the complexity of the model because it has more input to process (it will also be slower to train). 

   2. Decreasing the WINDOW_SIZE means each sequence uses fewer time steps to make predictions. This simplifies the input for the model but might not capture enough of the pattern, especially if the data has longer-term dependencies.

5. Split to Train, Validation, and Test

We will split our dataset into train, validation, and test to avoid overfitting and ensure the model is tested on data it has never seen before. Click on this link to learn more about splitting out data into train, validation, and test. 

1. (Code Block 5A) We have provided the code to split the dataset into training, validation, and testing parts. This block will also normalize our data. In LSTM, the algorithm relies on scaling data with different values so that it is easier for the LSTM to work with. Click on this link to learn more about why we normalize our data. Run this code block.

   1. For instance, if you are trying to compare the ages of people and their salaries, the LSTM might focus too much on the salary (because the numbers are bigger) and ignore the age of the people (because the numbers are smaller compared to the salary numbers).
   2. Normalization helps to put all the variables on the same level so the LSTM can learn from the data more easily. We will normalize the variables ‘CO,’ ‘Ozone,’ ‘PM10,’ ‘PM2.5,’ and ‘NO2.’

2. (Code Block 5B) This code block prints out the sizes of X_train, y_train, X_val, y_val, X_test, and y_test. If you see the X_train size as (216, 5) that means there are 216 samples (days) in the training data, each with 5 features (e.g. CO, Ozone, PM10, etc.)

6. Training the Model

1. (Code Block 6A) We have provided the code to create an LSTM model. Run this code block. 

2. (Code Block 6B) We have provided the code to set up a model for training and saving. It uses a ModelCheckpoint to save the best version of the model during training to a file called ‘model.keras.’ Run this code block.

3. (Code Block 6C) This code block trains the LSTM model using the training data we split earlier. Run this code (this is like pressing play to let the computer do its job and learn from the examples we have given it). Don’t worry if this takes a while to run! We gave it a lot of data. 

7. Evaluating the Model

1. (Code Block 7A) This code block will reload the saved model. Run this code block.

2. (Code Block 7B) This code block prints a table with the actual values of the predicted days on the first column and the values that the model predicted on the right column. At a glance, how close are the model’s predictions?

3. (Code Block 7C) This code block displays a graph of the predicted values next to the actual values for AQI. Are the lines close to each other?

4. (Code Block 7D) This code block calculates the Mean Absolute Error (MAE) to assess the accuracy of the model’s predictions.

   1. The MAE shows how close predictions are to actual values by averaging the size of errors. A lower MAE means better prediction accuracy.

   2. For example, an MAE of 13 indicates that, on average, the model’s predictions differ from the actual values by 13 units. That means that if the model predicts the AQI to be 100, the actual AQI may be anywhere between 87 and 113. 

5. (Code Block 7E) This code block calculates the Mean Squared Error (MSE) to assess the accuracy of the model’s predictions. 

   1. The MSE measures the average squared difference between predicted and actual values. It provides insight into the accuracy of the model by penalizing larger errors more heavily, as the differences are squared. Lower MSE values indicate better model performance. 

   2. For example, an MSE of 325 means that, on average, the squared difference between the predicted and actual values is 325. This metric is particularly useful when you want to minimize large errors, as the squaring emphasizes their impact, making it ideal for regression tasks where larger deviations are more significant.

6. (Code Block 7F) This code block calculates the R² (R-squared) value, which is a statistical measure that assesses how well the regression predicts the actual data points.

   1. R² = 1: Perfect prediction. The model accurately predicts future AQI levels with no errors. 

   2. R² = 0: Poor prediction. The model’s predictions are no better than simply guessing.

   3. 0 < R² < 1: This range indicates how much of the variation in future AQI levels the model can explain. For example, an R² value of 0.75 means the model can predict 75% of the variation in AQI beyond what would be predicted by just using the average value. 

   4. How accurate were your model’s predictions? How close is your R² value to 1? 

8. Record the Metrics and Retest for Different Timesteps in the Future

1. Currently, the model is set to predict the Air Quality Index (AQI) seven days into the future. Begin by recording the current performance metrics–Mean Absolute Error (MAE), Mean Squared Error (MSE), and R²–in the table below. 

2. Next, update the model to predict AQI for a longer period by modifying the WINDOW_SIZE parameter in Code Block 4F to 28, which will forecast four weeks ahead. Once updated, run all the code blocks again by selecting Runtime > Run all. 

3. After the code execution completes, locate the updated metrics (MAE, MSE, and R²) in Code Blocks 7D, 7E, and 7F. Record these new values in the table. 

4. Repeat steps 2-3 but change the WINDOW_SIZE parameter to 365 to predict AQI one year into the future.

5. Finally, reflect on the results: Did the model maintain accuracy when predicting different time frames, or does it perform better with shorter-term predictions compared to longer-term ones?

Table 1. Comparison of MAE, MSE, and R² metrics for LSTM models predicting AQI at 1 week, 4 weeks, and 1 year timeframes.