Detection of lost natural gas consumption with machine learning: Sakarya case

Energy is among the areas where countries around the world spend the most. The management and protection of supplied or produced energy resources is as important as accessing energy. Within the scope of this thesis, is was tried to determine the unaccounted consumption of natural gas, one of the world's most important energy sources. This situation is defined as the difference between the amount of gas sent to consumers from a distribution network and the amount of consumption reflected on the bills of consumers, and it causes serious economic losses to both the companies providing natural gas supply and the countries. Leaks in the transmission pipes, change in air temperature, inactive consumers, measurement errors and possibly defective meters can be shown as the cause of this situation. Intentional physical interventions and aging of the meters over time can be given as causes of gas meter failures. Interventions made by subscribers are mostly examined under "natural gas theft". This study focuses on the loss of natural gas consumption that may be caused by the meters. In this study, models estimating the consumptions of each natural gas subscriber were created and it was tried to determine the subscribers with abnormal consumptions based on the estimates made. As the study area, natural gas consumption data of 307.570 subscribers in the province of Sakarya, whose consumption times vary between 2011 and 2021, were used and the subscribers were divided into groups according to their consumption time sizes. A comprehensive data warehouse has been created by using the data set that indicates the meteorological conditions, which are important determinants in natural gas consumption, and the calendar events dataset that includes the information on weekends, public holidays, and holidays between the relevant consumption periods. This created data warehouse was used in model training and testing processes. A multi-layer artificial neural network model has been developed to predict the consumption of subscribers. The models of the subscribers were trained on the consumption data of each subscriber for one year, and the consumption of the remaining periods was estimated by the model. In the study, two different models were developed for each natural gas subscriber. These were realized by using the information, which is among the natural gas consumption data, whether the relevant subscriber's meter is renewed or not within each billing period. Using this information, the model in which the network is informed that the next consumption of the subscribers whose meters are renewed is realized with the renewed meter was created and it is called the "relative model". The other model is the one in which the meter renewal processes are specified pointwise, that is, the meter information is not given to the network, and it is called the "point model". Natural gas consumption estimations were made by performing training and testing processes for each subscriber with these two models. The performances of the predictions were evaluated using the MAPE (mean absolute percentile error) method. In the predictions made with the developed models, the mean MAPE value was 2.42 in the relative model and 2.12 in the point model. The estimation performances of the two models were evaluated and the effects of the meters used on natural gas consumption were determined. In this context, the distribution of MAPE values of all subscribers obtained from the two models was examined using statistical methods, and the limits of subscribers with normal consumption behaviors on the distribution were determined. Then, the subscribers outside the normal limits were divided into two subgroups. These are named intermediate-transition zone and abnormal zone. By using the obtained limit values, subscribers with abnormal consumption behaviors were determined and it was stated that the subscribers in this region had a high risk of having lost and unaccounted gas consumption.