Reliable GNSS positioning is a very challenging task in harsh urban environment. The main source of error is due to non-line-of-sight (NLOS) reception and multipath phenomena. The effect of assuming a direct path in a NLOS propagation environment leads to serious degradation in accuracy. Instead of discarding all measurements which are found to be in NLOS conditions, we propose to properly use these observations to improve the positioning accuracy and integrity in harsh environments. In this paper, we combine the sigma-ε variance model with a mean jump (i.e. NLOS bias) to model the pseudorange (PR) errors. First, we use a 3D model of the environment to detect the NLOS state of reception and to predict the NLOS bias related to the excess delay phenomenon. For reliable positioning, we use a C/No-based variance adjustment for the LOS PRs, and we subtract the bias from the NLOS PRs during the trilateration step of position computation. The performance of the proposed scheme is assessed using real data and compared to a standard Kalman filter without predicted information from the 3D simulator.