Research On Prediction Algorithm Of Free Parking Stalls Based On Genetic Algorithm And Ensemble Learning

For the past few years,with China's rapid economy development and urbanization,motor vehicles have been entered the ordinary people's family.By the end of 2017,China's ownership capacity of 217 million vehicles,but at the same time,the pace of development of parking stalls has lagged far behind,which makes the problem of parking more prominent,even in some areas have reached the “one difficult to ask” point.Parking problems caused by traffic congestion,environmental pollution and other issues on the existing production,life,environmental security has a great impact.Therefore,the real-time and effective prediction of the vacant parking stals is an important research direction to alleviate the above problems.Free parking stals effective prediction belongs to the research category of time series prediction,which can be divided into short time prediction and multi-step prediction according to different forecast demand.The short time prediction uses the historical data to predict the number of parking stals in a single point of time in the future.At present,and the exponential smoothing method and its derivation algorithm are widely used.The multi-step prediction is the use of historical data to predict the number of parking stals in the future for a period of time.Methods based on neural network,such as BP Neural Network,are widely used to solve this kind of prediction.In this paper,we first analyze the existing short time forecast and multistep prediction method of vacant parking stals.Secondly,studying their principles and proposing two improved algorithms for the limitations and accuracy of these methods in specific fields.With the improved algorithm is used to get rid of the prediction limits of short time prediction and multi-step of traditional parking stals prediction,has good compatibility and scalability,and also improves the accuracy and robustness.Because of the time series reflects the continuity and regularity of the development of objective things,it is the sequence of statistic indexes according to the sequence of their occurrence time,so the statistical values of the close time points has corresponding correlation,the closer the adjacent point correlation is stronger.There is a similar attribute of memory.In the field of artificial intel igence,the traditional neural network only has a short-term memory,and the Recurrent Neural Network(RNN)can remember its previous input.When it comes to continuous and context-related tasks,it has a greater advantage over other artificial neural networks and is the preferred neural network for dealing with time series problems.In order to solve the problem of low accuracy and complicated training process,one RNN algorithm based on GRU(Gated Recurrent Unit)structure is proposed firstly in the paper.Firstly,the time series data of parking stals are segmented into different input sequence for the input of network training,and then the optimal network structure is determined by implementing and comparing different structure RNN for the input sequential training effect.Finally,the optimal structure of RNN is compared with the existing widely parking stalls algorithm,and the experimental results show that the algorithm not only solves the problem of the traditional neural network in the gradient vanishing,but also shows good experimental results and achieves a good balance between accuracy and scalability in different prediction requirements such as short-term prediction and multi-step prediction.In order to further increase the accuracy of prediction and prevent over-fitting caused by neural network over-training,this paper uses the Ada Boost algorithm based on the first algorithm to improve the accuracy of prediction and to correct the loss caused by RNN prediction.The two algorithms are iteratively trained by genetic algorithm to find the optimal parameters.That is,the prediction algorithm of free parking stals based on genetic algorithm of the combination of RNN and Ada Boost.By setting a range of genetic algorithm parameters,the algorithm optimizes the structure type of RNN,different neural network layers,neuron numbers and the number of Ada Boost weak regression learners.At last,the best prediction results are obtained by means of the corresponding mean value treatment for each prediction result of two algorithms with minimizing the loss mode.