Research On The Prediction Model Of Available Parking Space Based On LSTM Neural Networks

With the rapid development of China's economy,the number of urban cars is increasing rapidly,and the demand for parking is increasing rapidly.The parking problem has become a serious traffic problem.The prediction of available parking space can be very important for drivers to choose the parking lot and route.This thesis is working as follows:(1)Based on the analysis of an underground parking lot in a hospital in Wuhan city,this thesis proposes an long short-term memory(LSTM)neural-networks model for available parking space based on single-step and multi-step prediction.The predicted results have been scrutinized against those of earlier models,so as to verify the the prediction accuracy and stability of the new model.The optimal parameter setting of the LSTM neural-networks model is found to be the method of grid searching.Moreover,the comparison of the iterative method and the direct method of the multistep prediction performance are under discussion.By analyzing the advantages and disadvantages of the above two methods,it reveals that the root reason for why earlier available parking space exists precision bottleneck,and the result provides an important theoretical support for solving the tough multi-step prediction problems.(2)The maximum Lyapunov exponents(LE)method,which is commonly used in the chaotic time series prediction,is introduced into the prediction of available parking spaces due to the chaos of the available parking space time-series data.By using the same data set for comparison experiments,the prediction performance of LE and LSTM neural networks models in step and multi-step prediction is compared.Experiment results show that in the early stage of the available parking space prediction period(prediction step length is less than 5 steps),LSTM neural networks model of prediction error is lower,but in the later step prediction period(prediction step length is more than 5 steps),LE method's performance is better.According to the results of comparative experiments,this paper puts forward a new kind of available parking space-a multi-step prediction model LSTM-LE.LSTM neural networks has a better shortterm forecasting learning ability during the early prediction,and LE method which perfectly reflect the chaotic characteristics of the late forecast period.By using parking data to conduct a virtual experiment,which is to predict the available parking space after 50 minutes,the results show the root mean square error of LE method can fell to 7.51 from 9.12 of LSTM neural networks model,while the root mean square error can be further decreased to 5.53 by using the new model proposed in this thesis.