The proliferation of industrial information technology has made digital transformation and intelligent manufacturing central to modern manufacturing, with digital twin technology enabling real-time optimization and dynamic decision-making. Meanwhile, e-commerce growth has driven fragmented, small-scale, multi-batch uncertainly demand, forcing enterprises to balance delivery time windows, production-line flexibility, and sequence-dependent setups costs, thereby substantially complicating scheduling. To address this, we formulate a mixed-integer programming model for unrelated parallel machine scheduling aimed at maximizing total profit, under constraints including delivery time windows, sequence-dependent setup times, product heterogeneity, and machines capacity differences and so on. On the demand side, deep reinforcement learning is integrated for order forecasting, embedding demand into production planning to balance loads and optimize resources, thereby mitigating stockout and overstock risks and furnishing data-driven planning foundations for digital twin systems. On the solution side, an improved adaptive large-scale neighborhood search (ALNS) algorithm is designed based on the tabu search escape mechanism, with destroy and repair operators tailored to the model's characteristics to enhance efficient real-time solving for complex scheduling problems. Computational experiments validate the effectiveness of the proposed approach, demonstrating improved delivery reliability and capacity utilization, reduced costs and decision-making opacity, and a scientific basis for intelligent scheduling in manufacturing enterprises.