Real time small target fire detection in photovoltaic plants using attention enhanced multiscale fusion

Photovoltaic (PV) power plants concentrate flammable polymers and arc-prone electrical components, making early fire detection a safety-critical condition-monitoring task. Yet incipient fires appear as small, weakly contrasted smoke or flame, while specular reflections from glossy module surfaces fabricate fire-like cues, jointly degrading conventional detectors. We cast this as an applied computational-intelligence problem and propose PVFire-YOLO, a real-time neural detection framework. A patch-aware parallel attention block amplifies weak small-target responses against reflective backgrounds; a gather-and-distribute neck performs intelligent information fusion across four feature-pyramid levels to close the semantic gap of conventional necks; and a four-scale adaptive spatial feature-fusion head learns per-pixel scale weights, adding a high-resolution branch for sub-100-pixel targets. To enable training and evaluation, we built PV-Fire, the first annotated flame-and-smoke corpus with a PV-specific background distribution. Against fourteen baseline detectors, PVFire-YOLO attains 87.4% mean average precision at intersection-over-union 0.5 on PV-Fire and 84.4% on the public D-Fire benchmark, raises small-target average precision by 5.2 points over its baseline, and lowers the reflection-induced false-positive rate to 3.1% at higher recall. It runs at 79 and 21 frames per second on a desktop graphics card and an embedded edge device, supporting real-time deployment.