Dual-Stage Attention-Based Recurrent Neural Net for Time Series Prediction based off this blog post by Chandler Zuo. I've extended the code to work for multi-variate time-series and added some pre-processing, but given I'm basically copying the code from his post, the copyright probably belongs to him.
The most recent branch, which uses the PyTorch JIT, is called jit.
There is a different implementation by Zhenye-Na, but from what I can tell, it's only single-variate.