I'm currently working on a time series project, and I intend to employ the VMD+CNN technique for forecasting the output.Upon applying VMD to the training data, I obtained a total of 14 Intrinsic Mode Functions (IMFs). Consequently, I constructed my CNN neural network with dimensions (30100, 20, 14, 1), with 20 representing the window size. I achieved an acceptable RMSE of 67 for the training set and 74 for the test set. However, when I assessed the model's performance on new data using a new separate decomposition VMD, I encountered a significantly higher RMSE of 300. This is puzzling because it can't be attributed to overfitting, as the model performed well on the test data previously.
Can you provide any insights into this discrepancy? Shall I devide my data into sub series of 20 steps window then decompose them using VMD ?
