kwuking
How to Use TimeMixer and TimeMixer++ for Electricity Forecasting？
Thank you very much for your excellent work on TimeMixer and TimeMixer++. I am currently working on electricity forecasting, and I recently came across your models, which have demonstrated impressive performance. We have conducted some experiments using your models and found them very promising.
However, I have a few questions and would truly appreciate your guidance:
Normalization and Prediction on Original Scale
We noticed that both TimeMixer and TimeMixer++ perform normalization during the prediction process. For electricity forecasting, the data often exhibits significant fluctuations. I don't quite understand why the output needs to be normalized. I can't use this result in my scenario. Could you advise on how to make predictions directly on the original data scale without normalization?
Handling High Variability in Electricity Data
Electricity data is highly volatile, and I need to forecast power consumption data for the next three months. Are there any specific techniques or adjustments you would recommend for your models to better handle such volatility and long-term forecasting?
Thanks for your support of our work. I have just taken over this project, and my understanding of time series might not be very deep yet. As I understand it, normalizing data has been a common practice in the time series community since the introduction of Informer. However, this might differ significantly from real-world data, where evaluations are often conducted on the original scale. You can set this in the data loader to inverse the data back to the original scale. Regarding the volatility in electricity forecasting, there might currently be a lack of effective solutions. You can try using our model first, and if the results on your data are not as expected, you can refer to more models in TSLib for testing. Thank you again for your support.
Hi, @kevinliu2000, @JasonAlexTan Thanks for your effort I think TimeMixer++ is the best option for AI forecasting. So I am going to use it in the crypto trading field. I want to predict 3 ~ 6 hours token price But I can't set parameters correctly. And I tested with some parameters, but didn't get a good result. The most problems are input_size, output_size and dataset size. There are 1 min, 5 min, 15 min, 30 min, 1 hour candle datas. I used 10MA to reduce noise. Please help me
