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 60-120 points). My problem now is how can i train the same model on these different records taking into account that every sample should be independent. Also, any insights about the best architecture to follow to better predict the peak value? An example of the record is given below.
I'm trying to predict the peak values ( 5 points in green ) given the previous value ( 10 points in orange).
I have found some suggestions to concatenate the records into one big record but there would be a jump in the concatenation point and would share the cell state between the independent records. Based on some research, I found that the LSTM reset states every batch so I broke each record ( 60 points ) into sequences (lag = 10 points and prediction = 5 points ) which resulted in 46 sequences. Then I selected the batch size to be 46 but I'm not sure if this is correct or not and if there are any better criteria