For downscaling and forecasting applications, it is often necessary to train/validate separately from applying GPEP for ensemble generation . The trained/validated regression (or ML/DL) model is stored and can then be read back in when needed for out of sample ensemble generation. The current sample station-point cross-validation for the regression does not fully separate training point sites from the final regression model and ensembles, or allow for use in distinct time periods (past or future). This train/predict mode separation is common in many statistical method packages (R, Python).
For downscaling and forecasting applications, it is often necessary to train/validate separately from applying GPEP for ensemble generation . The trained/validated regression (or ML/DL) model is stored and can then be read back in when needed for out of sample ensemble generation. The current sample station-point cross-validation for the regression does not fully separate training point sites from the final regression model and ensembles, or allow for use in distinct time periods (past or future). This train/predict mode separation is common in many statistical method packages (R, Python).