The short- and long-term forecasting of the grid electrical energy of a province or region is important for the management of the conventional electricity transmission and distribution network. Nowadays, the amount of electrical energy, which each residential building has taken from the grid, has gained importance within the scope of smart grids. Residential buildings that take electricity from the smart grid can generate electricity with alternative energy sources such as solar energy. Besides generating its energy these buildings can also take electricity from the electricity grid. In other words, they can use both the energy that they generate and the energy that they receive from the electricity grid. This bi-directional energy flow increases system complexity, and the grid system become more dynamic. Therefore, estimating the electrical load of such a system is also more difficult than the conventional grid system. In this study, the recurrent linear regression method (R-LR), which is based on the linear regression, was proposed. In order to test and validate the proposed approach, the Sundance dataset, which is shared in the U mass trace repository according to smart project was used. To confirm the success of the proposed method, the linear regression (LR), and the extreme learning machine (ELM) methods were used in each of the 59 different datasets. Obtained results, which were applied to 59 different residential buildings smart meter datasets, showed that lower root means square error and corrected symmetric mean absolute percentage error, which was proposed in this paper in order to eliminate zero dividing errors, values were achieved by R-LR compared to LR and ELM. It was found that the proposed R-LR provides better results in modeling dynamic systems and gives good results in forecasting analysis with a time-series dataset.