Abstract: To address strong process continuity, significant cascading effects among processes, complex coupling among process variables, prominent temporal dependencies, and raw-material fluctuations in process manufacturing, this paper proposes a multi-process quality prediction method that integrates hierarchical graph construction with multi-level encoding–decoding. The proposed method establishes a multi-level hierarchical architecture consisting of micro-process graph encoding, macro-process gated fusion, and temporal modeling and decoding, which enables hierarchical representation and step-by-step modeling of the complex characteristics of multi-process production systems. At the micro level, process variables within each individual process are constructed as a graph structure, in which nodes represent different process parameters and edges describe the potential interactions among variables. A micro graph attention encoding layer is developed to adaptively learn the correlation weights among process variables without requiring predefined causal relationships or fixed dependency assumptions. By extracting implicit interaction patterns from historical process data, the proposed layer can effectively characterize nonlinear coupling among multiple process parameters and generate informative process-level feature representations. At the macro level, a chain gated fusion layer is designed according to the sequential structure of the production process. The encoded hidden states of upstream processes are propagated progressively as contextual information between adjacent processes, while gated interactions are then applied to combine this context with the downstream process representations obtained from graph encoding. Through this dynamic fusion mechanism, the proposed model explicitly describes the gradual accumulation and adaptive integration of quality-related information during cross-process transmission. Compared with conventional feature fusion approaches that treat different processes independently or combine process information statically, the gated fusion layer better captures the progressive influence relationships and cumulative effects among multiple production stages. Based on the cross-process and cross-scale representations described above, an improved long short-term memory (LSTM) network combined with a self-attention mechanism is employed for temporal modeling and decoding. The improved LSTM network effectively captures the long-term dynamic evolution of process variables through its gating structure and state update mechanism. Meanwhile, the self-attention module calculates the contribution weights of different time steps with respect to the prediction target, allowing the model to automatically focus on the time periods most relevant to the final quality indicator. Through the integration of graph encoding, gated fusion, and temporal decoding, the proposed framework achieves a unified representation of multi-process and multi-scale information and improves quality-prediction accuracy. Finally, three key processes, namely loose moisture regain, moisturizing and ingredient addition, and drying processes, in a tobacco manufacturing production line are selected for verification. After data preprocessing operations, including data cleaning, mutual information-based feature selection to eliminate redundant variables, and normalization, a multi-process dataset is constructed to predict the key quality indicator of outlet moisture content. Compared with several typical models, the proposed method achieves lower mean absolute error and root mean square error, as well as a higher coefficient of determination, demonstrating its advantages in both prediction accuracy and stability. The results indicate that the proposed model can effectively capture complex parameter coupling within processes, cumulative influence propagation among processes, and dynamic temporal characteristics of process variables, thereby providing a reliable reference for quality control and production optimization in process industries.