The mgcv package proposes a flexible framework for fitting Generalized additive regression models.nHowever, classical fitting procedure can be computationally intensive. The bam procedure brings about substantial computational savings, by adapting standard fitting algorithms to provide scalability to "big" data sets [1].nIn particular, parallel approaches have been implemented to exploit multi-core architectures and to reduce memory footprint. We will present the results of joint work between the University of Bristol and one R&D team of EDF (the major French electrical utility). The new bam procedure has been used to model electrical load time series freely availablenfrom the NYC ISO. The new optimization algorithm (FREML) of bam allows the user to fit scalable additive models on data up to millions of observations and thousands of estimated parameters.