If you drill down into the demographic details, Trump has 29% of the black vote (Obama had 93% in PA 2012), and 30% of Hispanic (Obama had 80%).
These are **WTF** type numbers. If sampling was distributed by population OR if these black/Hispanic ratios are in the ballpark of the actual results, Trump wins PA. Combine both of these together, you have a polling error of 10% (like in the Primary PA vote where Trump overperformed polling by 10%).