Özet:
Turkey requires reliable forecast values and assessment of its electricity imbalance to realize energy plans since it is import dependent on energy sector with very low energy intensity. Back propagation (BP), recurrent (R) and radial basis (RB) neural networks (NN) are compared with a measure of MSE applying five-fold cross validation after composing conceptual framework for energy forecasting. We propose aggregate and sector based models including input predictions with most promising networks. The essential point is that rates for all socioeconomic variables are preferred to derive a benefit from neural networks nonlinear capability and to pay attention to yearly demand behavior not settling with trend. As a result of performance improvement trials, resilient propagation and its own parameters in favorable recurrent network have a remarkable decreasing effect on MSE which resulted as 0.006 for training, 0.009 for test sets. As detailed analysis, simulations with the preferred model are realized composing 11 folds. Most of the average values for mean absolute errors for demand rate are below 3 per cent. Another major difference of this study is its focus on electricity imbalance following the total energy consumption prediction, thus, avoiding drawbacks due to inconsistent data, immature market and uncertain climate. Linearly increasing electricity percentage share in total energy consumption and load factor changes are reflected to future with simple but strongly significant regression models. Looking forward to 2020, Turkish energy import dependence will increase also with a growing gap of electricity imbalance although demand rates will reflect a little downturn after 2012.
