The Large Synoptic Survey Telescope (LSST) ranks among the next generation of wide field imagers designed to probe Dark Energy. It will cover half of the visible sky from Cerro Pachon (Chili) in six optical photometric bands and be able to identify millions of galaxies, up to redshift 6. The huge amount of observed objects makes impossible to ensure a spectroscopic follow-up of the complete photometric survey. High accuracy reconstruction of photometric redshift is needed in order to fully exploit the Dark Energy probes (supernovae, week lensing, Baryon Acoustic Oscillations, etc.) from LSST observations. A new method for photometric redshift reconstruction has been developed by the French LSST collaboration. The algorithm implements template fitting method and a statistical test to remove outliers. It has been developed using the expected apparent magnitudes computed for LSST, from galaxies of a mock catalog based on GOOD survey luminosity function. In order to removed outliers galaxies, we test two different techniques : Likelihood ratio statistical test (LR) and Boost Decision Tree (BDT). Both methods are based on 16 discriminate parameters : the 5 color terms and 11 parameters from posterior probability density function of the fit. By applying some quality cuts on LR or BDT variable, we manage to reject the most important part of outliers (from 6% to 0.15% at redshift 1, with an LR cut) and so, improved the accuracy of photometric reconstruction. We are able to estimates photometric redshift up to redshift 2.5, respecting the LSST requirements. Those two discriminate methods give comparable results on photometric reconstruction and precise determination of the cut parameter will be driven by physics (more galaxies, higher redshift versus better resolution) and will need a dedicated study. In addition, we show that the best compromise between redshift reconstruction quality and loss of statistic is obtained using galaxies well observed (S/N>5) in at least 5 of LSST pass-bands.I will first introduce the LSST project. Then I will go through the method we used, to finish by presenting you some of ours results.