Algorithm 3 Dynamic Learning

Input:   $y$: bandwidth slice, S: list of local smoothing algorithms, ${\phi }_i$ : hybrid smoothed LSTM algorithm, k: list of hybrid Smoothed LSTM algorithms (6 in this case),   Output:   $\delta$: statistically significant smoothed LSTM algorithm, E: Forecast Error   Process   01: begin   02: $\delta \leftarrow \varnothing$;   03: for all time steps$t_iϵ y$ do   04: $W_j\leftarrow W_j \cup \left\{t_i\right\}$;   05: if Change is detected = true then//using Anderson–Darling   06: Stop forecasting at $W_j$   07: $\widehat{y}\leftarrow$ smooth $y$ in $W_{j+1}$ using algorithms in S   //algorithm 3   08: for ${\phi }_i$ ink   09: ${\phi }_i\left[ \right]\leftarrow$ Build new hybrid LSTM models ($\widehat{y},\vartheta )$   //algorithm 2   10: E $\leftarrow$ Calculate Forecast error of $t_i in W_{j+2}$ using ${\phi }_i$   11: k$\leftarrow k\cup {\phi }_i sorted with Min\left(E\right)$   12: $\delta \leftarrow$ Find in k the significant ${\phi }_{i }$ with Min(E)   13: If $\delta$ is significantly better than ${\delta }_{i-1}$ //(old-Existed  forecast algorithm) then   14: replace${\delta }_{i-1}$ by $\delta$ else if   15: keep $\delta$   16: endif   17: endif   18: Loop