[Question] Forecasting (Multivariate or Multiple) TimeSeries with multiple covariates

[ZhangAllen98]

In the multi-time-series example, Note that ElectricityDataset contains measurements of electric power comsumption (in kW) for 370 clients with a sampling rate of 15 minutes. For this case, if I want to add each clients with future covariates including temperature, wind speed, and humidity etc, How should I do.

Should I transform the multivariate time series (1 series with 370 components)to multiple time series (370 series with each having 1 component), and for each series to prepare a covariates time series with 3 components. It seems this way is straightforward.

Or just keep the multivariate time series unchanged, and prepare the covariates time series, If so, how should I prepare this kind of covariates time series to let the model know which three covariates belong to which client?

Is there any demo code for this kind of case.
Thanks

[dennisbader]

Hi @ZhangAllen98, yes you should do it as you described here:

Should I transform the multivariate time series (1 series with 370 components)to multiple time series (370 series with each having 1 component), and for each series to prepare a covariates time series with 3 components. It seems this way is straightforward.

For this, simply create a list of single column time series from the electricity dataset:

from darts.datasets import ElectricityDataset
series = ElectricityDataset().load()
series = [series[col] for col in series.columns]

And then create a covariates TimeSeries for each of the series in series and pass them as a list of covariates TimeSeries of the same length (370) to the models.