I attended a few meetings in which the city of chicago is attempting to predict the E.coli levels in the water for the next day using data science techniques.

I attended a few meetings in which the city of chicago is attempting to predict the E.coli levels in the water for the next day using data science techniques. It was a frustrating experience since the current day's E.coli level measurement can vary wildly depending on various factors such as time of day, water depth at site of measurement, month in year, weather, etc... or maybe sometimes it's just a bunch of birds that all pooped in the same area on a given day. Plus, each beach has it's own pattern and correlations; some folks started adding rules the USGS has some method of averaging readings, but I was very tempted to try to find the ideal model that could handle the non-iid beach data. someone on the net suggested multilevel hierarchical modelling, which i don'f fully understand. i think part of the issue is that data is already semi-structured, as well as time based. one approach is to flatten the structure (assume iid) and make the data stationary (remove the time component). but i wonder if something like an RNN would handle the multiple time steps better.