Demand forecasting of energy consumption with statistical and artificial intelligent techniques: Sakarya natural gas consumption application

Country governments are working on reducing economic and social losses by making energy demand forecasts with low error rates, reliable and accurate. In this thesis, the year-ahead monthly and the day ahead demand forecasting of natural gas which is one of the major sub-sectors of the energy sector, are completed using different methods to achieve the least errors. Suggested methods are univariate and do not need any data other than the consumption data. The originality of the thesis is that we attempt to detect the best model and approach that reduce penalties caused by inaccurate forecasting, and show that they are applicable in everyday life. In this study, two different cases are considered for natural gas demand forecasting. In the first case, the year ahead monthly whereas in the second case the day ahead demand forecasting is done. Firstly, based on the four-year natural gas consumption data, year ahead monthly demand forecasts are done on monthly and daily data volumes by applying the time series decomposition (TSD), Winters exponential smoothing (WES), autoregressive integrated moving average (ARIMA) and seasonal ARIMA (SARIMA) methods. It is observed that conversion of the daily forecasting into monthy consumptions decreases %12.9 MAPE to %11.9 MAPE in year ahead monthly forecasting by comparing monthly consumption based forecasting. In the daily demand forecasting, two different day ahead forecasts are done. In the first forecasting, TSD, WES, ARIMA and SARIMA model results are found %27 MAPE and 0,8 R2 for forecasting the year in one step. In the second forecasting, WES and artificial neural networks (ANN) are applied with the sliding window technique (SWT) for the day ahead forecasting. Here, the WES method is applied by computing α,β,γ parameters for each day of the first three years data. In the ANN method, backpropagation (BP) and artificial bee colony (ABC) algorithms are used to train the networks by using the same first three years daily data. Best results for SWT used WES and ANN-ABC models in the day ahead consumption forecasts are found %15, %14,9 MAPE and 0,94, 0,89 R2 value, respectively. To conclude, it is shown that the proposed forecasting methods are able to generate acceptable appropriate results on the natural gas consumption demand forecasting used frequently in the literature. With the proposed models, it is shown that forecasts, do not require any variables than historical data, is able to reduce the demand forecasting penalties paid in the real life.