This project involved forecasting Sulphur Dioxide concentrations in the atmosphere. The data was collected from certain Air Quality survey sites from the USA. The dataset contained the following details:


Date: The date when that data point was collected.
Source: Source of information- Air Quality Survey.
Site ID: Each survey site has a unique ID.
POC: Stands for “Parameter Occurrence Code”, used to distinguish different instruments that measure the same parameter at the same site (Source).
target: Max SO2 concentration of a given day at a given site.
UNITS: The unit the measurement is taken in - ppb(Parts Per Billion.)
DAILY_AQI_VALUE: Air Quality Index of the given date at a given place.
Site Name: Name of the Survey Site.
DAILY_OBS_COUNT: Total number of observations taken in a day.
PERCENT_COMPLETE: Percentage of the number of hourly observations taken in a day(24 hours).
AQS_PARAMETER_CODE: The unique code for SO2 according to AQS, which is 42401 (Source).
COUNTY: Name of the county where the survey site is located.
SITE_LATITUDE: Latitude of the survey site.
SITE_LONGITUDE: Longitude of the survey site.


Highly correlated values were dropped after observing the correlation matrix (on the right.) After doing a thorough EDA to better understand the data. I ran multiple tests like Rolling statistics to observe Stationarity, then the Dickey-Fuller Test to check the stationarity of the data.

Once it was ascertained that the data was stationary, I performed some pre-processing steps like normalizing the variables, one-hot encoding the values.

Once the data was ready, a simpleRNN model was created to forecast the time series. RMSProp optimizer was used. Since the project didn't require highly accurate prediction, I resorted to this model for faster training. The model trained for 100 epochs and achieved a minimum validation loss of 0.0000944.