Hi Paweł Machał

You can estimate the statistically significant spectral peaks in the spectrum or spectrogram and then use the estimated coefficients corresponding to the peaks for forecasting purposes. You may want to look at the following documents for more information:

Article Wavelet Low- and High-Frequency Components as Features for P...

Article A Survey on Change Detection and Time Series Analysis with A...

Hope these references are useful

Best regards,

Ebrahim

Dear Paweł Machał ,

Look the link, maybe useful.

Article Electricity Peak Load Demand using De-noising Wavelet Transf...

Regards,

Shafagat

DWT is a method for signal analysis/processing/synthesis. If your intention is to forecast from the DWT coefficients, you will need some method for this. They can be statistical methods such as ARMA/ARIMA or deep learning methods such as LTSM, among others.

Good afternoon, thank you very much for your answers and answers regarding the work undertaken. my further steps will be according to the diagram below.

the diagram shows:

1) demand

2) approximation at level 13 obtained in the wavelet transformation using the db45 wavelet

3) Seasonality was obtained by fitting the description of A13 with a sinusoidal function.

then the residual signal was obtained by difference. the next step will be to decompose the obtained residue by means of a wavelet that will extract from the signal, for example, the weekly seasonality.

What do you think about such work algorithms?

I understand that this algorithm is adequate. You need to be concerned about the chosen level of decomposition. Because it will interfere with the seasonality you want to infer. I have a video about it but it's in Portuguese: https://www.youtube.com/watch?v=LNm_d6Te76U&t=2015s

Dear Paweł Machał

it attached a good reference

Article Wavelet Transform Application for/in Non-Stationary Time-Ser...