Fire detection systems deployed on edge devices face critical challenges in balancing detection accuracy with energy consumption. While recent deep learning approaches have improved accuracy, most studies overlook the strict power constraints essential for real-world deployment in resource-limited environments such as IoT networks and solar-powered systems. This study proposes a multi-dataset training strategy combining the FCMI-Fire (indoor), D-Fire (outdoor), and Roboflow (mixed) datasets to enhance cross-domain generalization. We selected the lightweight YOLOv11n architecture and optimized it for edge deployment on the NVIDIA Jetson Orin Nano using TensorRT FP16 precision and the DeepStream SDK. To ensure methodological rigor, a two-level partitioning protocol was implemented to prevent data leakage during evaluation. The multi-dataset trained model achieved an 89.15% mAP@0.5 on the combined validation set and demonstrated robust cross-domain performance, significantly outperforming single-dataset baselines which experienced catastrophic performance degradation (as low as 18.06% mAP) when tested on unseen environmental conditions. Specifically, our model maintained high accuracy across all domains: 96.68% on FCMI, 80.32% on D-Fire, and 93.11% on Roboflow. Hardware evaluation showed that the optimized edge deployment achieved a real-time throughput of 74.2 FPS with a power consumption of only 6.8W, resulting in a unified energy efficiency score of 972.8. This work establishes a practical baseline for deploying robust, lightweight object detection models in power-constrained environments for industrial and IoT fire monitoring applications.