Estimation and comparison of daily natural gas consumption with deep learning and statistical modeling methods

In this study, forecasts made for the next day's consumption of natural gas by using natural gas consumption data in Sakarya province between January 2014 and June 2018. Forecasts were made by the periodical autoregressive integrated moving average model with using only daily natural gas consumption values. However for the forecasts were made by simple neural networks and deep neural networks, has been used both natural gas consumption values and the meteorological values of Sakarya province. And success rates of those forecasts were compared each other. The 365-day forecast success of the Seasonal Autoregressive Integrated Moving Average model was 10.33 percent in terms of Mean Absolute Error Rate(MAPE) achieved. As a result of training of artificial neural networks by using data sets created with three different parameter selection methods, MAPE values of 6.68 with simple Artificial Neural Network and 7.23 with Deep Artificial Neural Network were reached. Using the data up to seven days in which these parameters are included, 7,61 with the Simple Recurrent Neural Network, 6,45 with the Long-Short-Term Memory Neural Networks and 6,38 with the Gated Recurrent Unit Neural Networks, Mean Absolute Percentage Error (MAPE) ratios has been reached. According to the findings obtained in this research, it was seen that the predictions made with the gated self-repeating unit neural network trained with the characteristics chosen by the expert opinion gave better results than other methods.