The widespread existence of various uncertainties makes the inherently complex refinery production scheduling problem even more challenging. To address this issue, this paper proposes a viable systematic data-driven multistage scheduling optimization framework and develops a corresponding structured modeling methodology. Under this paradigm, unit-level advanced control and plant-level intelligent scheduling are coordinated to jointly deal with endogenous and exogenous uncertainties while ensuring the implementability of the solution. Historical production data are leveraged to learn the operation mode-specific process models and to train the overall stochastic dynamic scheduling model. Operation mode indicative matrices with associated “shift registers” are introduced to describe different operational status of the production units and also the transition processes induced by operational switching. The model is established based on the policy graph representation and its tractability can be achieved by resorting to the stochastic dual dynamic programming algorithm, although it is mathematically a tricky mixed integer model. The scheduling policy is allowed to be trained according to different risk preferences of the decision-makers and can be applied to online interactive sequential scheduling over finite or infinite horizons. Model refining and policy retraining mechanisms based on human feedback equip the scheduling policy with good generalization capability and evolutionary adaptability. An industrial-scale application case study is conducted based on Julia language, and the in-sample and out-of-sample Monte Carlo simulations demonstrate the feasibility and effectiveness of the proposal in this paper.