Abstract: With continuous development of photovoltaic industry, fire monitoring and alarming in photovoltaic system has become an important issue. In this paper, a feature-enhanced Faster-RCNN algorithm is proposed for smoke detection based on video surveillance under the background of photovoltaic power station project. Considering that existing Faster-RCNN algorithm based on region proposals has high detection accuracy for single-scale images, but it has such shortcomings as too large parameter scale, poor recognition effect for multi-scale smoke images and poor real-time performance. In order to adapt to the complexity of photovoltaic power station scenarios and smoke dispersion modality and fire alarming system's accuracy and real-time performance, Faster-RCNN algorithm is chosen as baseline. With the introduction of residual network architecture ResNet-50 as the backbone, and a lightweight feature pyramid network into Faster-RCNN, the feature extraction capability of this model is enhanced. In experiments, we use smoke generators spraying at different locations to simulate fires, with multi-angle surveillance cameras recording video images. Four surveillance videos containing smoke samples are used for testing. Results show that detection rate and error rate of our proposed method is 98.25% and 3.13% respectively. The speed of our detection system reaches 0.0721s per frame, which verifies effectiveness of our method.