Multi-variate Time-series Prediction models

Question

I have a time series dataset with the following characteristics:

• 5 years duration data
• 50 features in time-series, daily (temperature, humidity, etc)
• Target: predict CO2 level

I'm looking for a time-series multi-variate algorithm to help me predict the CO2 of tomorrow, having temperature, humidity, etc of today. All of these features have their own patters and might/might not be related with CO2.

Nevertheless, what kind of model would you recommend (e.g., RNN) in Python? How would I fit this amount of features to the predictor? How would I define the training window size?

Answer 1

You have 1825 samples here (which is not a lot). Each sample is a single timestep described by a feature vector of dimension 50. The first step is to create training input data in the form of a numpy array with a shape of (1825L,1L,50L) this assuming you plan only to predict based on a single previous day. You may have better luck using the previous 30 days, in which case the data would be (1825L,30L,50L) The data you have (temp,humidity) will be best represented as standardized normalized floats. The training output is a single float so your model will accept (1,50) and output (1) , psuedo-code for a keras implementation:

Make Sequential Model

Add an LSTM layer, 100 input_shape=(1,50)

Maybe a dropout layer, 20%

Output layer Dense (1)