Very interesting article, thx for sharing this!

Have you tested it on more than 1 minute interval? I mean in real world you need to predict within factory lets say in range of 10-30 minutes, if supporting remote sites even hours. Did you try to predict lets say 1 hour into the future? What could be bottleneck and possible tuning such approach. On the other hand in my case I have dataset for past 4 years with very detailed markings on different actions within devices (maintainances, runtime, warning, error, etc.)