Experimental data and simulation results are presented on the separation of methane-nitrogen mixtures in a two-bed laboratory PSA installation packed with a carbon molecular sieve adsorbent. The effect is analysed of methane concentration in the feed gas, pressure adsorption and cycle length on the average methane concentration in the low-pressure product. Based on the experimental data a neural network model is proposed and a quantitative description of non-isothermal PSA processes is validated.