Use of last time step state variables for training

[dsgoll123]

For practical reasons, we use the land surface model state variables from the restart file of ORCHIDEE which records the values for Dec 31st. Such type or restart files are common for land surface models. It would require code modifications to the LSM code to be able to record all state variables in time (e.g. record 365 daily values over a year ). This can easily become a bottleneck for the application of the tool to the various LSM and ORCHIDEE versions.

A shortcoming of this approach is, that the value of Dec 31st might lack information (e.g. state variables with strong seasonality). We believe, the issue might be minor as it is more important for us to predict slowly changing state variables which take long time to spin-up. The other state variables spin-up quickly in the 're-run' land surface model simulation

[ma595]

@dsgoll123 we need to discuss this in the next developer meeting.