Steam injection optimization is critical for enhancing heavy oil recovery, yet conventional methods struggle with dynamic parameter adjustments and data scarcity. This study proposes a novel data-driven framework integrating Dynamic Time Warping (DTW)-based data augmentation and a Patch Time Series Transformer (PatchTST) to address these challenges. Leveraging 3,200 manually labeled cycle records and 1.5 million daily production data points from 276 wells in a Chinese heavy oil field, we developed a DTW-enhanced data generation method. This method aligns daily production sequences (e.g., steam injection, liquid production) with cycle labels through optimal temporal warping, followed by cubic spline interpolation and physics-constrained fragment recombination, expanding the dataset to 130,000 cycles while preserving oil-steam ratio and casing pressure limits. The PatchTST model incorporates two key innovations: 1) A patching mechanism that divides time series into 16-length patches with a stride of 8, reducing sequence length from 128 to 16 tokens while preserving local injection-soak-production patterns, and 2) channel-independent processing that isolates multivariate signals (e.g., pressure, temperature) to minimize cross-channel noise. Comparative experiments against LSTM and CNN demonstrated PatchTST’s superiority. For 5-step predictions, PatchTST maintained superior robustness outperforming LSTM and Transformer, attributed to its dual capabilities in long-term dependency capture and multivariate decoupling. This work establishes a new paradigm for steam injection optimization through synergistic integration of physics-guided data augmentation and advanced temporal modeling, providing a scalable solution for intelligent oilfield management. The proposed framework’s code and implementation details are publicly available, facilitating adoption in similar reservoir engineering scenarios. Future directions include integrating generative adversarial networks for enhanced physical consistency and coupling with reservoir simulators for automated constraint generation.