Is this approach right to solve a forecasting problem using LSTMs?

I'm trying to forecast some values with information about many variables from the past, there's a description of the data:

  1. It contains information over the time from 5 sensors for different variables ie:

|sensor | var1 | var2 | var3 | target | date |
|sensor1| ... | ... | ... | ... | ... |
|sensor2| .. and so on

here's what I've done:

  1. Split the data into train, validation and test set chronologically, the first 80% records belong to training and 10% for test and 10% validation, each sensor has the same quantity of records of the others on each set on the same date range.

  2. then I created the windows first grouping by sensor, then selecting all the variables (except the target) from the 5 next records as X and from the 5th record I selected the target value as Y (I do this because target value is not only dependent from previous records but current records), after that I move to the next record and do the same till I complete the records for the sensor and till I complete all the records, It's sliding window method.

  3. Finally I fed it into a LSTM but the MAE is about 40% for training and validation sets, the architecture I'm trying is: LSTM(128),LSTM(64),LSTM(32),Dense(1)

Am I doing something wrong? should I add to the Xs the value of target variable from previous periods and don't do sliding window but something like selecting the 5 next values and the target and move not 1 but 5 records away?

Read more here:

Content Attribution

This content was originally published by Salacho65 at Recent Questions - Stack Overflow, and is syndicated here via their RSS feed. You can read the original post over there.

%d bloggers like this: