Reducing the inputs so drastically and then using a neural network is a weird choice. Time series forecasting usually uses a lot more data before is starts using fancy methods like neural networks. I suspect their granularity is small enough and reactive enough that this forecast isn't doing much work. The grid can probably muddle through sizeable imbalances and other protections kick in if the frequency changes too much.
