I'm working on a problem trying to predict the peak value in a thunderstorm wind speed record. I have many wind records from different locations across the US (each record should be around 120 points). My question is how to manipulate the LSTM architecture to better handle those peaks.
Example of the used TS record:
(https://i.stack.imgur.com/YY4TD.png)
My basic intuition is to give more attention to the peak values by modifying the loss function or using attention, but I'm not sure how to do that.