Finding rare wind patterns from noise

Hey! I’m working on a weather forecast project and a bit stuck getting started with distinguishing meaningful patterns when most of the data is random noise.

A bit of context: I get minute-by-minute data from a weather station with wind speed and wind direction, and I’m hoping to build a system which alerts me whenever a rapid but predictable change is happening. For example, if the wind direction has been steady and speed has decreased for the past 20 minutes, can I expect the wind direction to change in the next 10 minutes? Having this knowledge helps preparing experiments beforehand.

The problem is that most of the sensor data is pretty random. What I’m looking for are rare occurrences of conditions that follow patterns, which would allow predicting the tails of those patterns.

I did the previous course but couldn’t find any example that directly matches my case. Any tips or advice to put me on the right track with how to accomplish this with Fast.AI would be highly appreciated. Thanks a whole bunch! :slight_smile:

This is more a general machine learning question rather than a FastAI question.
Things I would consider when trying to answer this question:

  • Looking at the data, how would you identify where the wind direction changed (label the data as needed)
  • Next, how would you train someone else to answer that (Curate as many training data to train the ML with)

Then I would train a time series model to predict the wind direction change.
or a Linear/CNN model accepting the last X units of time to classify a wind direction changed condition.