Trajectory Hotspot Prediction Based On Convolutional Neural Network

With the rapid development of navigation and positioning,communication,sensing and other technologies,it has become practical to use sensors to collect the location information of mobile objects over time.As a kind of time series data,the mobile trajectory data records in detail the time and position change information of the object movement,as well as its own unique attributes,state and behavioral characteristics,which can reflect to a certain extent the interaction between the object and the various elements in the environment.A lot of meaningful information can be obtained by analyzing and mining the motion trajectories of mobile objects,and the results can effectively serve the fields of smart ocean and smart transportation.In order to explore the hidden information and knowledge from the massive spatio-temporal trajectory data and provide support for future decision making and management,this paper takes the fishing trajectory of ocean fishing vessels and the travel trajectory of urban online taxi as the research objects,and mines the massive trajectory data through data analysis and mining,artificial intelligence and other technologies to achieve the prediction of future fishing hotspot operation time and passenger travel hotspot ride demand.The main work of the paper is as follows.1.For the marine fishing boat trajectory data,firstly,we use the rulebased algorithm to divide the single fishing boat outbound trajectory,and use the DBSCAN algorithm plus K-Means algorithm for secondary clustering to obtain the fishing boat operation trajectory points.At the same time,for the characteristic that the fishing boats have a freer heading when conducting fishing,this paper uses a honeycomb grid to divide the study area,and then the operating hours of fishing boats are counted in time and space dimensions.In addition,an integration strategy based on the honeycomb grid is proposed in this paper,by using the average value of the regional prediction as the final prediction result.The implementation of the integration strategy resulted in the reduction of RMSE and MAE by 8.8%and 9.8%,respectively,which effectively improved the accuracy and robustness of fishing vessel fishing hotspot prediction.2.In terms of prediction algorithms,benefiting from the excellent results of Transformer in the fields of natural language processing and image recognition,this paper proposes a spatio-temporal Transformer prediction model.First,the location embedding layer encodes the spatio-temporal location information based on the input data,and then the encoded data enters the spatio-temporal Transformer module for spatially dependent and temporally dependent feature extraction.At the same time,the spatiotemporal Transformer module can be overlaid to extract deep spatiotemporal information.By comparing the classical prediction models,the algorithm in this paper reduces the RMSE and MAE by 23.8%,24.1%,26.6% and 30.8%,36.4%,37.7%,respectively,when predicting the fishing vessel operating hours data for the next 5,10 and 15 days compared with the best model.3.For urban online taxi trajectory data,data cleaning is performed first,and then the map matching algorithm is used to correct the trajectory points that deviate from the actual road network.Passenger boarding and alighting data extraction algorithm is used to obtain passenger boarding and alighting time and location data,while a square grid is used to divide the study area,and by projecting and aggregating passenger boarding and alighting time and location data in the spatio-temporal dimension,the OD matrix data characterizing passenger demand in each region at each time period is obtained.4.To verify the effectiveness and applicability of the Spatio-Temporal Transformer in the field of spatio-temporal data prediction,we apply the model to the OD matrix data obtained from the city’s online taxi trajectory data.By using the historical passenger on and off the net taxi data,we predict the passenger ride demand in the future time period of Chengdu city.Through model comparison tests,the model reduces the RMSE loss by29.5% and 11.7% in short-time prediction(5 steps)and long-time prediction(15 steps),respectively,which proves that the model can be applied to other spatio-temporal data prediction fields.