Question about F-Block

[iganggang]

Dear author, Could you please tell me which model F-Block uses?

Thank you very much.

[YHYHYHYHYHY]

The F-Block utilizes a customized Transformer for the frequency domain. This customization is achieved through the CompEncoderBlock, which is defined in ./layers/ComplexLayers.py.

Should you have any additional questions or need further clarification, please feel free to ask.

[iganggang]

I recently encountered a question about the hybrid model. I use VMD to decompose the raw data, get some modes, and then input the patchtst individually, but the result MSE is from 0.4 to 0.8. The predicted values ​​of peaks and troughs are worse than patchtst.

Could you please tell me if the idea is wrong? I use some signal processing methods to decompose raw data. then I get some modes in time and frequency domain. then I do not know how to use these components to predict.

[YHYHYHYHYHY]

Sorry, I don't quite understand your problem. Do you mean that through decomposition, you obtain modes in the time domain (trend perhaps?) and in the frequency domain (seasonal maybe?), and you want to build separate models to predict these two parts? For example, using an F - Block for seasonal prediction and a T - Block for trend prediction?

Since ATFNet doesn't employ any decomposition methods, I'm uncertain if this would enhance the forecasting performance.

[iganggang]

Thank you for your insights. Yes, I’m exploring the idea of decomposing the time series signal into its trend seasonal and residual components. Then, all components use T-Block for prediction. i wonder if the input is simple, the output will be better. But when I try this, the results are the opposite. here is the comparison with pred and true values. as is the same as to last question. the peak is low.

[YHYHYHYHYHY]

It seems that the peaks you mentioned in the previous two responses are different, with the former one in the time domain and the latest one in the frequency domain. As I'm uncertain whether incorporating decomposition for targeted forecasting is effective in general scenarios, I suppose the situation you described might be due to the following possible reasons: 1) The unique characteristics of the training data. 2) The decomposition methods might not be accurate enough. 3) Applying decomposition may not be suitable for every time series.

Thank you for presenting your insights, and we encourage you to explore further in this regard and share more interesting findings in the future.