Algorithm 1 The proposed model training algorithm.

Input: Historical spectrum data sequence Output: Trained long-term joint temporal–spectral network //Construct the dataset 1: construct spectrum matrix ${\mathit{\chi }}_{\mathit{t}-\mathit{n}},{\mathit{\chi }}_{\mathit{t}-\mathit{n}+\mathbf{1}},\dots ,{\mathit{\chi }}_{\mathit{t}-\mathbf{2}},{\mathit{\chi }}_{\mathit{t}-\mathbf{1}},{\mathit{\chi }}_{\mathit{t}}$ from historical spectrum data 2: D←$\varphi$ 3: divide set D into train set $D_{train}$ and test set $D_{test}$ 4: //Train the proposed prediction model 5: initialize all learn-able parameters in long-term joint temporal–spectral network 6: Repeat 7:      randomly select a batch of instances from $D_{train}$ 8:      find $\mathit{w}$, $\mathit{b}$ by minimizing the objective (6) 9: Until the training epochs are met 10: //Test the proposed prediction model 11: for each sample in $D_{test}$ do 12:      fed into the trained proposed model 13:      output the prediction results of that sample 14: end for