The deep learning-based method for small target fire detection in complex scenarios mainly faces two challenges. First, collecting video images of small target fires in complex scenarios is costly and difficult, which limits the generalization ability and robustness of the model. Second, small target fire detection in complex scenarios is easily affected by factors such as fire scale, scene type, and lighting conditions, resulting in low detection accuracy. To address the above issues, this paper proposes a small target fire detection model based on S-PGA-YOLOv12 for complex scenarios. Firstly, based on YOLOv12, it integrates the parallel patch awareness attention (PPA) module (for highlighting key information of small targets), the GOLD module (for balancing speed and accuracy), and the small target detection head (Detect-ASFF) module (for adaptively learning the spatial fusion weights of feature maps at different scales). Secondly, to solve the problems of high cost and difficulty in collecting small target fire images in complex scenarios, a dataset construction method based on simulation is proposed. Finally, the small target fire dataset of complex scenarios is constructed through simulation. The effectiveness of the proposed model is verified through ablation experiments, comparative experiments, as well as robustness and generalization analysis. Extensive experiments on the three datasets constructed in this paper demonstrate the effectiveness and superiority of the proposed method.