Energy Demand Forecasting: Avoiding Multi-collinearity

Abstract

As having one of the major economies and rising population, Turkey's energy demand is increasing substantially. The main objective of this research was to apply ridge regression to estimate Turkey's primary energy consumption. Gross domestic product, population, automobile ownership, export and import rates, manufacturing and electricity consumption values of the country were involved in the forecasting model as independent variables. Although regression models end up closer estimates to the real values, having multi-collinearity between variables makes those models unreliable. Therefore, other techniques such as time series, artificial neural networks and genetic algorithms have been tried and performed better than regression models. Ridge regression, a rarely applied and underappreciated model in the literature, is used to overcome the multi-collinearity problem which means high correlation among independent variables. In this study, the ridge regression technique was compared with time series methods and artificial neural networks. The principal results showed that ridge regression is better to estimate energy demand and gave lower mean squared error than other techniques (16.51 for ridge regression followed by 19.00 for neural network). Moreover, estimated values were also found closer to the real energy demand than official projections of the Ministry (only 5% deviation with the proposed model, while official projections occurred by 20% error). Since the accurate forecasting of energy demand is significant for the proper policy design, the best methodology should be opted for and ridge regression seems one of those alternative techniques. In addition, the easiness of the ridge regression makes it applicable to several forecasting methods.