Electric power systems are moving toward smarter and more sustainable systems. These trends result in several positive advantages such as active participation of customers in electricity markets. However, resulting demand side flexibilities cause high demand fluctuations and increase the difficulty to maintain the power balance and reliability of smart grids. To address this challenge, this paper proposes a self-partitioning local neuro fuzzy model, which is capable of performing a fast and accurate short-term load forecasting. The proposed model, not only maintains the linearity as well as learning-from-data property via their fuzzy inference systems of local linear neuro fuzzy, but also benefits from partitioning the input space into linear and nonlinear vectors and assigning them separately into different local models. The proposed model is trained with the hierarchical binary-tree learning algorithm and rule premises are calculated through sigmoid partitioning functions. These appealing properties make the model appropriate for a fast and accurate analysis of the load time series featuring both linear and nonlinear characteristics. The effectiveness of the proposed model is compared with recently published forecasting models in terms of statistical performance.