Fig. 2: Architectural overview of Seq-to-Seq model.

In this overview subfigure a depicts the detailed architecture of the encoder and decoder components. The LSTM-based encoder processes historical temporal segments to capture the intricate pattern of battery life cycles. It integrates skip-connection and layer normalization to preserve and stabilize essential key temporal features. The decoder is initialized with the encoder’s final states and applies an attention mechanism to focus on relevant temporal features from the encoder output and enrich the context of its predictions. The attention-enhanced representations are combined with the initial decoder input and subsequently propagated through LSTM layers. A fully connected layer with leaky ReLU activation and a dropout layer—used solely during training and inactive during inference—for regularization follow the LSTM outputs. The model outputs are then fed into three separate fully connected layers for predicting a specific quantile of the future distribution based on the pattern learned during training, thus providing a probabilistic characterization of the forecast. Subfigure b illustrates the integrated Seq-to-Seq model flow, depicting the progression from encoding historical data to multi-output future forecasts. It highlights the sliding-window approach that underpins the model’s capability to handle both the tail-end of historical data and the integration of self-generated forecasts with known future conditions. This process also captures the dynamic training process, which incorporates teacher forcing to enhance the predictive fidelity of the model.