We investigate in this article a method based on proper orthogonal decompositions (POD) in order to reduce atmospheric chemical mechanisms. In a preliminary test, we apply this algorithm to the reduction of the CBM IV mechanism in a box model. The reduced basis computed with POD gives accurate results in its domain of validity while the CPU time needed for integration has been drastically lowered. The number of degrees of freedom of the local reduced model is 2 or 3. Moreover, we investigate some numerical strategies and a sensitivity analysis related to the choice of the reduced basis. This reducing technique is applied to three different scenarios for initial conditions: rural, urban, and polluted. Our tests indicate the deep relation with the slow-fast behavior of chemical kinetics.