Algorithm 2 Tr-Predictior

Input:  Suppose ${\displaystyle T=\{t_1,t_2,\dots ,t_n\}}$ is a set containing multiple workload sequences of different lengths and the base learning learner LSTM. Output:  The ensemble transfer model obtains the predicted value through the following formula: $Tem{p}_j=\frac{log\left(\frac{1}{{\beta }_t^j}\right)}{{\displaystyle \sum _{\mathrm{j}=1}^N}log\left(\frac{1}{{\beta }_t^j}\right)LST{M}_j\left(x\right)}$ For t = 1,2,…,S.  1:Call Equations (1) and (3) to obtain the source domain data, recorded as $T_{source}$ (length n);  2:Call Equations (4)–(6) to preprocess data, extract target data as $T_{target}$ (length m), $T_c=T_{source}+T_{target}$.  3:Set the initial weight vector: $W_t=1/\phantom{1m+n}m+n,\left(i=1,2,\dots ,m+n\right)$.  4:Set the total weight of $T_{source}$: $w_i^s=n/(m+n)$. The total weight of $T_{target}$ is $w_i^t=m/(m+n)T_{source}$.  5:Call the AdaBoost-LSTM (Algorithm 1) to train data $T_c$, freeze the weight $w_i^{\mathrm{s}}$ of the top n source data in the process of training and only update the weight of the target $w_i^{\mathrm{t}}$, record the above training model as $mode{l}_t$. (Train hypothesis in $mode{l}_t$ for $LST{M}_t$ is $h_t:x\to \mathbb{R}$.)  6:Use Step 2 in Algorithm 1 to calculate the adjusted error of each instance in the target domain: $e_{\mathrm{i}}^t$, change ${\beta }_t$ in algorithm 1 to ${\beta }_s=1/(1+\sqrt{2log\left(\frac{n}{N}\right)})$, use it to calculate the adjusted error of each instance in the source domain: $e_{\mathrm{i}}^s$.  7:Calculate the adjusted error, ${\epsilon }_t$, of $mode{l}_t$:${\epsilon }_t={\displaystyle \sum _{i=1}^n}{e}_i^t{w}_i^t$; if ${\epsilon }_t\ge 0.5$, stop and set $N=t-1$.  8:Let ${\beta }_t={\epsilon }_t/(1-{\epsilon }_t)$, freeze the weight of the target domain, then just update the weight vector of the new source domain: $w_i^{s+1}=w_i^s{\beta }_t^{e_i^t}/Z_t$.($Z_t$ is a normalizing constant.) End for.  9:return$y^{pre}={\displaystyle \sum _{j=1}^N}Tem{p}_j$.