From 2010, ERDF (the French electricity distributor) started the "Linky" project, which aims at implementing 35 million residential smart meters in France. These smart meters record consumption of electric energy every half hour and communicate this information to the remote data center for monitoring and billing purposes. The detailed electricity consumption information enables more accurate load models. In this context, two objectives are defined in the dissertation: the designs of load forecasting models for operation need and load estimation models for planning need. Two forecasting models are developed: one explores the mathematical formalism of time series and the other is based on the neural network mechanism. Both the two models forecast the electricity power consumptions of "D+1" and "D+2" with the prior knowledge of the information till day "D". A nonparametric load estimation model is proposed for the planning need. The model is compared to the "BAGHEERA" model, which is actually used by the ERDF for the same purpose. Completely driven by the individual consumption data, the model is presented with three variants: three nonparametric regressors (Nadaraya Watson, Local Linear and Adapted Local Linear). Validation scenarios by comparing with "BAGHEERA" model on the same data base showed that the nonparametric model is more accurate and adapts to the smart grid paradigm. All models presented in this dissertation are tested and validated with the real load data collected in the French distribution network.
