I'm working on a project in which i'm trying to do a pollution forecasting. I googled around and found that LSTM is a good candidate for this task, however, I'm still struggling at how to adapt it to my particular problem.

My problem is as follows :

Features :

  • 109 measurement stations (in the same region) :

    • description (ex : typology, coordinates ...)
  • 5 years of hourly data for each measurement stations :

    • weather data (ex : temperature, pressure, humidity, precipitation ...)
    • preliminary prediction of the target variables

Target :

Hourly measurements of 4 pollutants (O3, NO2, PM10, PM25) for each site

Goal :

Predict the daily value of each pollutant for each measurement site for the next 3 days

My Questions :

Q1. Multi-Site : Is it a good idea to build a different model for each measurement site ? thus, ending up with 109 models and not taking advantage of the fact that close stations tend to have close measurements.

Q2. Multivariate : Should I build a NN for each pollutant ? or just one NN that outputs a vector of 4 ?

Q3. Multi-step : How do you adapt the LSTM to output the prediction the next 3 days ?

Q4. : Do I need to precise a time window ? I've read that LSTMs learn the timewindow by itself

Q5. : Is it better to transform the data by calculating the daily average then make a daily prediction OR make the hourly prediction then calculate the average ?


closed as too broad by Peter Flom Dec 9 '18 at 11:53

Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. Avoid asking multiple distinct questions at once. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.