Genetic programming (GP) has been widely applied to learn scheduling rules for dynamic flexible job shop scheduling (DFJSS), yet the resulting rules usually generalize only to tasks with similar scales and distributions. When task characteristics change, performance degrades sharply, and fine‑tuning on new tasks often leads to catastrophic forgetting of previously acquired knowledge. To tackle this, we propose a surrogate‑assisted GP framework for lifelong learning that can handle sequential heterogeneous DFJSS tasks with job or batch arrivals while preserving past expertise. The framework comprises three key components: (1) an adaptive two‑stage training strategy that dynamically balances plasticity and stability; (2) a task‑similarity‑based weighting mechanism that regulates the extent of knowledge reuse; (3) an enhanced surrogate model with distance constraints to improve estimation accuracy. Experiments on several benchmark DFJSS scenarios compare the approach with multiple baselines, showing that it effectively mitigates catastrophic forgetting and retains strong adaptability, highlighting the promise of lifelong GP in dynamic scheduling environments.\n\nReview: The study offers a practical solution for continuous learning in GP‑based scheduling, combining training dynamics and surrogate modeling to address real‑world production challenges, and merits further testing on larger and more complex industrial cases.