Forecasting electricity prices in Turkey: A comparison of classical econometrics and machine learning techniques

Liberalization of electricity markets trends has been widening in the world since 1990s and the process is still continuing in many countries. As a result of liberalization, the private retailer and distribution companies role has increased in the markets. Like other countries, government enterprises was operating the Turkish Electricity market before 2000s. However, Turkish Government enacted a electricity market law for deregulation of the market in 2001. Accordingly, Turkish Electricity Trading and Contracting (TETAS) was established in 2001 with the aim of sustaining the energy trade and reach a competitive electricity market. Afterwards, Day-Ahead market and Intra-Day markets are established in the following years. As a result of the liberalization process, forecasting electricity prices has become an important requirement in energy companies' decision-making process to maximize profits by offering competitive bid and ask rates. On the contrary to other time series, electiricity prices have complex structures and dynamic determinants which makes difficult to forecast prices. For examle, one of the determinants of the electricity is demand and it is known that demand is highly related with temperatures. In summer, electricity demand expeced to increase due to cooling needs. However, it may change accordingly with tempetures. Futhermore, electiricty demand is also changing according to hour of the day or day of the week and so on. Another thing is that, generation source and type and the amount may also affect the prices. Consequently, electiricty prices are generally representing complex non-linear structures. There are many methods for forecasting electricity prices. However, artificial neural networks are known to be good forecasting method for covering non-linear effect of the prices. On the other hand, time series methods are easy to forecast and interpret the relationships. Lastly, quantile regressions are capturing the different conditional distributions and may be useful for capturing spkies. Therefore, artificial neural networks, time series and quantile regression is used. First, non-linear autoregressive network with exogenous inputs (NARX) is used with Levenberg-Marquardt learning algorithm. It is a two-layer feedforward network with sigmoid activation function. It is known that load is one of the most important factor affecting the prices, thus it is taken as inputs. Furthermore, to take seasonality in account, hour of the day, day of the week, season of the year, previous prices and other inputs are considered. The next day's prices are forecasted in out-of-sample data with multi-step ahead method and rolling window mechanism. Second, autoressive moving average (ARMA) models with different AR(p) and MA(q) orders are applied. To estimate a model and forecasts series in the model, series should be stationary and stable. Ng-Perron test is applied for testing whether series is stationary. Price series is found to be stationary. Next day's prices are forecasted in out-of-sample data with multi-step ahead method and rolling window mechanism. Furhermore, Optimal (p,q) are chosen after trial and error proces. Thirdly, ARMA models with exogenous variables are applied to see whether exogenous variables are affacting the performance of the autoregressive model. As in the ARMA model, series should be stationary. As in the NARX model, load is taken as inputs. Besides, hour of the day, day of the week and other variables are considered. Next day's prices are forecasted in out-of-sample data with multi-step ahead method and rolling window mechanism. Finally, quantile regression is applied. Different comibations of exogenous variables applied for different quantiles. Next day's prices are forecasted separetly for each quantile with multi-step ahead forecast methods and rolling windows mechanism. All methods with different combinations of the models are evaluted with mean absolute percentage error(MAPE) in out-of-sample data. According to findings, the NARX method outperformed the other models and forecast the electricity prices in TEM with lowest errors.