Short-term Demand Prediction For Online Car-hailing Based On Deep Learning

At present,deep learning has been widely used in many fields and achieved good results.Nowadays,the rapid development of online car-hailing industry has brought favorable changes to people’s lives,but there are also problems such as unbalanced resource allocation and contradiction between supply and demand.Online car-hailing demand forecast has a certain positive impact on online car-hailing enterprises,drivers and passengers,which can alleviate the contradiction between supply and demand and optimize resource allocation.In this paper,the deep learning algorithm is used to predict the short-term demand of online car-hailing.The main work is as follows:1.Start with the real order data set of Didi Gaia plan,mine the travel law and demand changes of passengers according to the preprocessed historical orders.It is proposed that the travel demand of online car-hailing passengers has temporal and spatial characteristics.As for the temporal characteristics,three attributes are analyzed and put forward,which are proximity,daily periodicity and weekly periodicity,which are closely related to the work and living habits of local residents.The spatial characteristics are shown as high demand in the central region and low demand in the suburbs.From the center to the surrounding areas,the demand gradually shows a sparse trend.2.In view of the limitations of previous studies on the use of random forest to predict the demand of a single region,a method of multi-output random forest is proposed to predict the demand of online car-hailing in multiple regions.In addition,CNN and LSTM,which are good at data spatial feature extraction and good at data temporal feature extraction,are respectively used for prediction based on the spatial and temporal characteristics of online car-hailing demand data,and CNN and LSTM are used to verify the differences between fitting methods of proximity,daily periodicity,weekly periodicity and fusion of the three attributes.The results show that the method of fusing three attributes is more effective than the method of fitting a single attribute,and the RMSE of CNN is 27.88%lower than that of multi-output random forest and 15.74%lower than that of LSTM,indicating that CNN combines three temporal attributes while extracting spatial features,and the prediction effect is better.3.Based on CNN and LSTM,an improved short-term demand prediction algorithm for online car-hailing is proposed,which combines residual network and attention mechanism.Three components are designed according to the data’s proximity,daily periodicity and weekly periodicity,and the network structure of each component is the same.Firstly,since the shallow neural network can only extract local features,a network containing residual units is proposed to extract spatial features of data,so as to better depict the spatial correlation between any two regions in the city.Secondly,LSTM is used to extract temporal features.Considering the different impacts of different historical periods on the demand for future periods,the time processing module is improved and an attention mechanism is added to dynamically adjust the weight.Finally,the results of the three components are weighted and fused and output.The experimental results show that the RMSE of the combined model proposed in this paper is 22.28%lower than that of CNN,which shows the effectiveness and advantages of the combined model.