I have a time series data and I wanted to try wavelet transform to denoise the data and then apply machine learning forecasting algorithm
My forecasting problem: is to predict the next day hourly power consumption .
I am using python, pywavelets library. I have read the library documentation. Unfortunately, I didn't find any source clearly descripting how to properly do the wavelet decomposition for forecasting.
I know it needs to be done very carefully because we need to consider that, in real life, future values do not exist and hence shouldn't be considered in the data to be decomposed .. or it's a case of data leakage.
However, I still not confident in how to apply it well.
I think the procedure should be like this:
- Split data into training and testing
- Reshape data into Xs and Ys
- Training: for each example apply wavelet transform for the Xs and Ys separately considering the data from the beginning.
- reconstruct the components
- Build a model for each component
- Repeat 3 to 4 for the Testing phase
- predict the next value in each component using the trained model
- sum the components to get the prediction
Are the listed steps correct? is it true that for the prediction models we need to train the model on the components, not the coefficients?
Any help is appreciated.
Thanks
question from:
https://stackoverflow.com/questions/65661073/wavelet-for-time-series-forecasting 与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…