The parking guidance system (PGS) can alleviate the disordered parking problem in the peak time, and reduce the time to find parking space. But existing PGS techniques are quite dependent to the realtime data and historical data, its performance will be greatly degraded when the data is insufficient. To solve this problem, a data repairing based PGS method was presented. First, by mining the spatial data around the parking lots, a spatial similarity metric of parking lot was proposed. Then, the possibility of parking data similarity was calculated when the parking lots had spatial similarity. If the conditional probability was large enough, the known data of parking lots would be used as the learning samples. Finally, the reparative data could be generated by recurrent generative adversarial networks. Experimental results show that when the parking lots have high spatial similarity, the data of parking lots have high similarity probability also. The data generated by the recurrent generative adversarial networks also have the same distribution with real data. By the proposed method, a large number of reasonable data can be generated efficiently, and the PGS performance can be improved while only few parking data is available.