Predict entire time series

I have been a long time lurker. Posting here for the first time. :slight_smile:

I have sensor data (pressure and temperature measurements recorded as time series) at various locations. I have the latitudes and longitudes of these locations.

Often, pressure isn’t recorded owing to mainly cost related concerns. I have been tasked to predict the pressure given the entire temperature series for a particular location. The data does not seem to have any visible trends.

I have the following data to build a model: latitude, longitude, timestamps, pressure, temperature, sensor type.

What would be a good approach to build a model that predicts the pressure for given times at a particular latitude and longitude?

Here are things I tried with limited success:

RNN/LSTM - I have seen models built for forecasting the same time series. How do I use RNNs to build a spatially dependent model?

WaveNet - I picked up the idea of WaveNet from one of @jeremy’s posts on the forum. I can use WaveNet to generate time series, but, how do I pass a location to WaveNet?

Any suggestions would be sincerely appreciated.


@srib Interesting problem! I am curious, do you have good reason to believe that pressure can be inferred from temperature? What is the dimensionality of your dataset (i.e. number of locations and temperature measurements per location, and what percentage also has pressure)?

You can treat lat, lng and datetime info as categorical variables. Have you tried Entiry Embedding from lesson 5(ColumnarModel)? I found it works well for time series in general, especially when there is a periodic pattern. It may not be obvious, but I believe pressure/ temperature have a periodic pattern to an extent. For instance, knowing if it is night time or daytime, will probably help you guess the temperature better.

A late response, but in case it is useful I would consider adding elevation and humidity additional variables (if you can get them).