A comparative study of long-term load forecasting techniques applied to Tunisian grid case

The rapid development of population, economy and technology currently has led to the fast increase in electric consumption. Therefore, efficient energy management and load forecasting have been flagged as a very important need for electric networks planning and operation. Recently, different techniques have been developed to deal with load forecasting. However, the stochastic nature and uncertainty characteristics of the electric load make demand forecasting a challenging task for electric utilities where no technique can be identified as the most efficient one. It becomes therefore relevant to identify the technique that best fits a specific dataset. Thus, in this paper we propose a comparative study of four techniques for long-term load forecasting, namely Box-Jenkins, multiple linear regressions, MLR and nonlinear autoregressive neuronal networks with and without exogenous variables, NARX and NAR. The study is conducted on the Tunisian electric consumption, and the performance of different techniques is evaluated as a function of historical data and forecasting period's variations. Quantitative and qualitative assessments of results are reported in this study in order to pinpoint the strengths and weaknesses of the different assessed forecasting techniques.