From historical patterns to hourly dynamics: interpretable deep learning for cycling estimation