The training was run for maximum 60 epochs. Outliers removed from both inputs and targets. Use the configuration file total_early_stopped_target_cleaned_scaled to reproduce the results.
The closer the past day is to the present day, the more attention weight it has. Also the same weekday in the previous week (Time index -7), has similary high weight as the previous day (Time index -1). So yesterday's data and same weekday's data from previous week are most important for model prediction.
Attention weights mostly peak on Wednesday, as seen from the mean values. And it is lowest on Saturday/Sunday. This is because covid cases often peaked at Thursday or Friday. On weekends less cases and deaths are reported, so eventually they have less impact on the model attention.
Train 
Weeekyl patterns are very important as seen from the attentions. That is also evident from the variable importance, as weekly features get most importance. Then past observations of the target variables. Since covid cases/deaths in the past weeks can help learn the trend and predict the future better.
- Static variables
- Encoder variables
- Decoder variables
