Algorithm 1 Training process for TFFNet
Input: traffic flow matrices:    $\left\{X_t|t=0,\cdots ,n-1\right\}$;
external factors: $\left\{E_t|t=0,\cdots ,n-1\right\}$;
lengths of temporally dependent sequences: $l_c$, $l_p$, and $l_q$;
period parameter: $p$; trend parameter: $q$.
Output: TFFNet model $ℳ$.
1 2 3 4 5 6 7 8 9 10 11 12 13 14	$S\leftarrow \varnothing$  for all available time intervals $t$ ($1\le t\le n-1$) do    $X_c=\left[X_{t-l_c},X_{t-\left(l_c-1\right)},\cdots ,X_{t-1}\right]$    $X_p=\left[X_{t-l_p\cdot p},X_{t-\left(l_p-1\right)\cdot p},\cdots ,X_{t-p}\right]$    $X_q=\left[X_{t-l_q\cdot q},X_{t-\left(l_q-1\right)\cdot q},\cdots ,X_{t-q}\right]$    // $X_t$ is the target value at time interval t    put a training instance $\left(\left\{\left|X_c,X_p,X_q\right|,E_t\right\},X_t\right)$ into $S$  initialize model parameters $\theta$  repeat    randomly select a batch of training instances $S_b$ from $S$    find $\theta$ by minimizing the objective $ℒ\left(\theta \right)$ with $S_b$  until the stopping criteria is satisfied  output the learned TFFNet model $ℳ$