اراﺋﻪ ﯾﮏ ﻣﺪل دوﻣﺮﺣﻠﻪ اي ﺟﻬﺖ ﺗﺸﺨﯿﺺ ﺗﻘﻠﺐ در ﺷﺒﮑﻪ ﺗﻮزﯾﻊ ﺑﻪ وﺳﯿﻠﻪ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ

A Two-Stage Model to Detect Electricity Fraud in The Distribution Network Using Deep Learning

ﭼﮑﯿﺪه: ﺷﺮﮐﺖ ﻫﺎي ﺑﺮق از دﯾﺮﺑﺎز ﺑﻪ دﻧﺒﺎل ﺷﻨﺎﺳﺎﯾﯽ و ﮐﺎﻫﺶ ﻣﻮارد ﺑﺮق دزدي ﺑﻪ ﻋﻨﻮان اﺻﻠﯽ ﺗﺮﯾﻦ ﺑﺨﺶ ﺗﻠﻔﺎت ﻏﯿﺮ ﻓﻨﯽ ﺑﻮده اﻧﺪ.از ﻃﺮﻓﯽ ﺷﻨﺎﺳﺎﯾﯽ اﯾﻦ ﻣﻮارد ﻟﺰوﻣﺎً از ﻃﺮﯾﻖ ﺑﺎزرﺳﯽ ﻣﺸﺘﺮﮐﯿﻦ ﻣﻤﮑﻦ اﺳﺖ ﮐﻪ ﺷﺮﮐﺖ ﻫﺎي ﺑﺮق ﺑﻪ دﻻﯾﻠﯽ ﻧﻈﯿﺮ ﻫﺰﯾﻨﻪ ﺑﺎﻻ، ﺗﻌﺪاد ﻣﺸﺘﺮﮐﯿﻦ و ... ﺑﻪ دﻧﺒﺎل ﮐﺎﻫﺶ ﻣﺤﺪوده ﺑﺎزرﺳﯽ ﺑﻪ ﻣﻮارد ﺑﺎ اﺣﺘﻤﺎل ﺑﺮق دزدي ﺑﯿﺸﺘﺮ ﻫﺴﺘﻨﺪ. ﯾﮑﯽ از راﻫﮑﺎرﻫﺎي ﮐﺎﻫﺶ ﻣﺤﺪوده ﺑﺎزرﺳﯽ، اﺳﺘﻔﺎده از روش ﻫﺎي ﻫﻮش ﻣﺼﻨﻮﻋﯽ اﺳﺖ، اﻣﺎ ﭼﺎﻟﺶ ﻣﻬﻤﯽ ﮐﻪ در اﯾﻦ ﺣﻮزه وﺟﻮد دارد ﻋﺪم ﺗﻌﺎدل در ﻧﺴﺒﺖ ﻣﺸﺮﮐﯿﻦ ﻣﺸﮑﻮك ﺑﻪ ﻣﺸﺘﺮﮐﯿﻦ ﻋﺎدي اﺳﺖ ﮐﻪ ﻣﻨﺠﺮ ﺑﻪ ﻋﻤﻠﮑﺮد ﺿﻌﯿﻒ اﻟﮕﻮرﯾﺘﻢ ﻫﺎ ﻣﯽ ﺷﻮد. در اﯾﻦ ﻣﻘﺎﻟﻪ ﺑﺎﻫﺪف ﻏﻠﺒﻪ ﺑﺮ اﯾﻦ ﭼﺎﻟﺶ ﺑﺎ ﻓﺮض اﯾﻨﮑﻪ ﺑﺘﻮان رﻓﺘﺎر ﻣﺸﺘﺮك ﻣﺸﮑﻮك را ﺑﻪ ﺻﻮرت ﺗﺎﺑﻊ رﯾﺎﺿﯽ از رﻓﺘﺎر ﻣﺸﺘﺮك ﻋﺎدي ﺑﯿﺎن ﮐﺮد، در ﻣﺮﺣﻠﻪ اول اﻟﮕﻮي ﻣﺼﺮف ﻣﺸﺘﺮﮐﯿﻦ ﻋﺎدي و ﻣﺸﮑﻮك دﺳﺘﻪ ﺑﻨﺪي ﺷﺪه اﺳﺖ؛ ﺳﭙﺲ ﯾﮏ ﺷﺒﮑﻪ ﻋﻤﯿﻖ اوﻟﯿﻪ ﺟﻬﺖ ﻣﺪل ﺳﺎزي رﻓﺘﺎر ﻣﺸﺘﺮﮐﯿﻦ ﻣﺸﮑﻮك آﻣﻮزش داده ﺷﺪه اﺳﺖ. در اداﻣﻪ ﺑﻪ ﮐﻤﮏ ﺷﺒﮑﻪ آﻣﻮزش داده ﺷﺪه اوﻟﯿﻪ، ﺳﻨﺎرﯾﻮﻫﺎي ﻣﺤﺘﻤﻞ ﺑﺮق دزدي ﺑﻪ ازاي ﻣﺸﺘﺮﮐﯿﻦ ﻋﺎدي ﭘﯿﺶ ﺑﯿﻨﯽ ﺷﺪه اﺳﺖ. درﻧﻬﺎﯾﺖ ﯾﮏ ﺷﺒﮑﻪ ﻋﻤﯿﻖ ﺛﺎﻧﻮﯾﻪ ﺟﻬﺖ ﺗﻔﮑﯿﮏ ﻣﺸﺘﺮﮐﯿﻦ ﻋﺎدي و ﻣﺸﮑﻮك آﻣﻮزش داده ﺷﺪه اﺳﺖ. ﺑﺮرﺳﯽ ﻣﺪل ﭘﯿﺸﻨﻬﺎدي ﺑﻪ ازاي ﺳﻨﺎرﯾﻮﻫﺎي ﻣﺨﺘﻠﻒ و ﻣﻘﺎﯾﺴﻪ ﺑﺎ ﺗﺤﻘﯿﻘﺎت ﭘﯿﺸﯿﻦ ﺑﺮ روي ﻣﺠﻤﻮﻋﻪ داده واﻗﻌﯽ ﺑﺎ ﺑﯿﺶ از 6000 ﻣﺸﺘﺮك ﻋﻤﻠﮑﺮد ﺑﺎﻻي آن را ﻧﺸﺎن ﻣﯽ دﻫﺪ.

Electricity utility have long sought to identify and reduce energy theft, which represents significant part of non-technical losses. On the other hand, once a fraudulent customer is detected, on-site inspection is necessary for final verification. Since inspecting all customers is expensive, utilities seek to reduce the range of inspection to cases with a higher probability of theft. One way to reduce the scope of inspection is to use artificial intelligence-based methods. An essential challenge here is data imbalance in terms of the ratio of normal to fraudulent customers, which leads to the poor performance of algorithms. In This paper in order to overcome this challenge, assuming that suspicious behavior can be expressed as a mathematical function of normal behavior, in the first stage, the consumption pattern of normal and suspicious customers is categorized using clustering algorithms. Then a deep neural network is trained to model suspicious customers. Next, using trained network, possible theft scenarios for normal costumers are predicted. Finally, a secondary deep neural network is trained to separate the normal and suspicious customers. Assessment of the proposed algorithm for different scenarios on a real data-set with more than 6000 customers and comparison with previous research shows its high performance.