Temporal Fusion Transformer-Getting wrong seasonality for rolling window

[Manjubn777]

Hi,
I am working on TFT for 15min dataset where my goal is to forecast for 1 year using rolling windows approach. where I forecasts for 6 days, then I pick up the last 4 days of forecasts using mean method((This gives 0.5 quantile values since it is a point forecast)). In every window, I keep on adding 4 days of data to the testing data set, so it forecasts for last 6 days including last 2 days of train data. My issue is when I forecast for 6 days at once(as prediction length is 6 days), I can see the 6 days seasonality however when I use rolling window, I have been getting different seasonality. Do you think it is due to feeding 4 days of data instead of 6 during inference?
Here I have provided inference code for reference. Thanks in advance