Scaling/Normalizing time series with extreme values

Question

I am working on a convolutional GANs model for time series simulation. The time series includes the historical time-series demand (sales) data for the retail items. Comprising of different product-store items coming with different sales patterns, the time series data have input values at different scales. There exist some time-series with very large scales when are taken into normalization or scaling, e.g. min-max scaling, they can be seen as extreme values. Indeed, this has negative effects on training of the network. I am wondering what would be the best approach to scale or normalize time series data with potential extreme values. Thanks.