DCAST: A Spatiotemporal Model with DenseNet and GRU Based on Attention Mechanism

Algorithm 1

DCAST algorithm training process.

	Input: historical observations: ; length of short-term dependence, period rule, and long-term dependence: , , ; span of period rule and long-term dependence: , ; external features:
Output: DCAST, model
//Generate samples from historical crowd flow observations
(1)
(2)	For all available time interval do
(3)
(4)
(5)
(6)	Put a training instance into
(7)	End
(8)	Divide into and
//Train the model
(9)	Initialize all the parameters in DCAST
(10)	Repeat
(11)	Randomly choose a batch of samples from
(12)	Find by minimizing the objective (10) with
(13)	Until stopping criteria is met
(14)	Output the learned DCAST model