The Proportional Fair (PF) scheduler currently implemented in cellular networks is optimal when the channel conditions are stationary. Using measurements, a recent work shows that the conditions for moving cars can be non stationary and vary along the route. Based on these observations, the authors of [8] devise an algorithm called (PF) 2 S that exploits Signal-to-Noise Ratio (SNR) maps and rate predictions to improve the utility over the standard PF algorithm. We propose an algorithm which gives a better prediction of the future rate allocation and has a better utility compared to both the PF and (PF)2S algorithms. The proposed algorithm employs projected gradient on a relaxed version of the problem to predict the future allocations. Simulation results show that non negligible gains in utility over (PF)2S can be achieved by this algorithm.