基于深度自编码器高斯混合模型的窃电行为检测

传统极轨卫星观测频次低和观测时间固定，难以对一个区域进行实时监测。因此，基于Himawari-8 (H-8)静止卫星提出了一种能持续监测输电线路山火的时空上下文火点自动识别算法。该算法首先根据中红外和热红外波段对火点的敏感度差异，采用固定阈值法快速实现绝对火点识别和潜在火点提取，然后在空间上利用中红外和热红外的通道值、平均值和标准差之间的比值关系对潜在火点进行细判。在时间上利用H-8的高观测频次并融合火点发生概率进行持续性火点识别。所提技术已应用于南方某省级电网输电线路山火监测，并且准确率达72.5%，漏报率为43.9%。监测结果明显优于固定阈值法与传统的上下文法。

Abnormal detection of electricity theft using a deep auto-encoder Gaussian mixture model

The traditional polar-orbit satellite with low-frequency observations and a fixed observation time is intractable to real-time monitoring for a specific region. Accordingly, a spatio-temporal contextual algorithm for continuous monitoring of wildfires near transmission lines is provided by the Himawari-8 (H-8) geostationary satellite. First, based on the difference in sensitivity of the mid-infrared and thermal infrared bands to the fire spots, the absolute and potential fire spots are determined by a fixed-threshold algorithm. Then, the flame pixels are identified by comparing the ratios between the channel value, mean, and standard deviation of the mid-infrared and thermal infrared from space. Also, persistent fire spots can be located by taking advantage of the high-frequency observation capability of H-8 in conjunction with the fire occurrence probability in terms of time. The proposed technique has been applied to the wildfire monitoring of transmission lines in a provincial power grid in southern China, boosting the accuracy rate to 72.5% and lowering the missing alarm rate to 43.9%, respectively. The current study results are significantly superior to the conventional contextual fire detection algorithm and fixed threshold algorithm. This work is supported by the National Natural Science Foundation of China (No. 52107017, No. 52177016, and No. 52130707).