Maybe take a fifteen minutely rolling mean instead.
Maybe only smooth the training target. Don't smooth the PV data when it's used as an input to the model? (Because those rapid ramps are probably caused by small clouds, which we kind of want our model to know about!)
Maybe test all permutations (no smoothing; smooth just the target; smooth all PV data; only smooth PV data when it's used as an input).
5 minutely PV power is very noisy.
Maybe take a fifteen minutely rolling mean instead.
Maybe only smooth the training target. Don't smooth the PV data when it's used as an input to the model? (Because those rapid ramps are probably caused by small clouds, which we kind of want our model to know about!)
Maybe test all permutations (no smoothing; smooth just the target; smooth all PV data; only smooth PV data when it's used as an input).