Deep Learning Based Parking Prediction on Cloud Platform

With the explosive growth of the urban population, an increasing number of cities face the parking problem. Studies shows that more than 30% of the urban traffic is due to these "search berth" traffic, which aggravate the traffic congestion and create more traffic pollutants. The prediction for parking availability plays more and more important role in the urban traffic management system. However, existing parking prediction methods have lower accuracy or require a large number of data and computation cost, which make it difficult for parking prediction to be widely used. In this paper, we propose a deep learning based parking prediction system architecture on cloud platform. The LSTM network is used to predict the availability of parking space, which performs well in time series prediction. In order to improve the prediction accuracy, we comprehensively take more factors into account such as time of day, weather and holiday. In addition, we also propose an economical workflow based on the elastic computing service (ECS) provided by cloud platform. In the optimized workflow, the model training and updating processes need not to be in running state all the time, which can reduce avoidable computation and cost significantly. We implement the proposed system and deploy it on the Ali Could platform. Experimental results show that our prediction model based on LSTM network outperforms BP neural network, and our system has good scalability to apply to the cloud platform.