Prediction of natural gas consumption in Ankara region using artifical neural networks

Energy has been an important factor in terms of development and social structure of societies throughout history. Today, energy demand is met by a variety of resources namely, petroleum, natural gas, coal, wind, solar and geothermal energy sources. As a result of the technological developments in renewable energy field, the share of renewable energy supply in overall sum has increased in the previous years. In spite of this fact, today approximately 86% of energy demand of the world is still met by fossil energy resources. Although the use of renewable energy sources have advantages especially in sustainability and environmental aspects, with today's technology they are not enough to met the whole demand of the world. The majority of the energy demand is met by fossil fuels all around the world. Petroleum is the most widespread energy resource in the world and coal and natural gas come after it. As the distribution of energy resources of petroleum, natural gas and coal cannot be found in the world in a balanced amount and the demand for them countinuously increases, they have a substantial amount of economical and strategical value. Therefore, conversion and sales operation of them attracts many countries as well as companies. Turkey has an increasing energy demand due to its increasing population and developing industral activities. Although geographical location of Turkey makes it closer to many energy suppliers, it does not have a notable amount of fossil energy resources. Therefore, it is foreign-dependant in terms of energy supplies. The energy storage efforts in the last years resulted in a partially secure supply of natural gas. Natural gas is the leading energy supply in Turkey. Most of it is imported from neighboring countries. The demand estimation of natural gas is crucial for economical and strategical reasons. Prediction of natural gas consumption in advance is required for various reasons. First of them is that storage of natural gas is possible only for limited amounts. Therefore, to prevent any problems that can be caused by deficiency of natural gas, the amount required must be known. Secondly, if the purchased amount of it is more than necessary, it would be a disadvantage for financial aspects. In case of a faulty demand prediction where more than the amount specified in the commercial agreement is required additional natural gas can be reached. According to the international contracts, supply of the additional natural gas can be purchased by penalized natural gas fees. These penalized costs are also applied to the natural gas wholesale trade companies. On the other hand, if the predicted demand is more than the actual required amounts, the companies are charged with redundant costs. Due to these reasons, successful demand prediction of natural gas is an important factor from all aspects of natural gas management processes. The distribution of natural gas is realized by the companies which are licensed by EMRA. During the licensing process, distribution regions for each company is also described by EMRA. In Turkish natural gas sector, two main natural gas consumer types exist. One of them is domestic customers and the other one is business organizations. In this work the data belong to the domestic customers were used. In this thesis study, natural gas demand of Ankara was estimated using Artificial Neural Networks. Natural gas demand estimation is also a required process for regional natural gas distribution companies for investment planning and prediction of operational processes. Artificial Neural Networks are widely used in estimation studies due to their learning capabilities, high performance in non-linear systems, and tolerance to missing and noisy data. Learning process of Artificial Neural Networks is realized by supplying former input and output data related to the target system. By deciding the inputs of the system and forming a suitable network structure, they become capable of forming a relation between inputs and outputs of the system. The trained network then is fed by test data which have not been used in the training process, and performance of network is evaluated. If the perfomance of the system meets the requred criteria, the network can be used in the future prediction processes. Amongst the types of Artificial Neural Neworks, Back Propagation Neural Networks with multilayer architectures were used in this reseach. In Back Propagatiion Neural Networks, the information flows from inputs to outputs whereas error propagates from outputs to inputs. One input, one hidden and one output layers were used to generate the Artificial Neural Network structures. To determine the parameters that have affect on output of the system is an important part of these kind of studies. In this study, the inputs of the system were determined as minimum temperature prediction, average temperature prediction, maximum temperature prediction, humidity, sensed minimum temperature prediciton, sensed maximum temperature prediciton, number of users and calorific values. The output of the system is natural gas consumption. Sensed temperature data in the input set was calculated from predicted temperatture data. Data used in the networks formed belong to the years between 2014-2017. Temperature and humidity values were taken from Turkish State Meteorological Service, number of users, calorific values and consumption data were taken from Baskent Natural Gas Distribution A.S. Data between 2014-2016 were used for training of the networks and data of 2017 were used for testing the network performances.To obtain a network model with high performance, various Artificial Neural Networks were generated. Different neuron numbers were tried in the hidden layer and the type of activation functions were also changed. Multiple trials with random initial weights were made to ensure that models do not stuck in local minimum values. The differences between the actual demands and the outputs of Neural Networks were calculated and based on the evaluation of these values, networks with the highest performances were detected. The performances of the networks were compared and the best network was selected based on evaluating the performances of all of the networks produced. Moreover, to compare the performance of Artificial Neural Networks with linear modeling approach, multilinear regression analysis was also realized in the scope of this research. The results show that, successful outputs for the prediction of natural gas demand were obtained using Artificial Neural Networks.