Research On Usage Demand Prediction Problem For Resources With Multiple Fixed Stations

With the development of urbanization,people's lives are increasingly dependent on rich and diverse resources,such as transportation resources,land resources,energy and so on.Usually the total amount of these resources consists of the amount of resources contained in multiple fixed stations.People's production and lifestyle have caused the spatial distribution of these resources to change over time,resulting in the total amount of resources at some stations not meeting people's needs at some point.By predicting the resource usage requirements of multiple fixed stations,it is possible to find out the shortage of supply in advance,thus solving the imbalance between supply and demand.The demand for resources has two attributes of time and space,thus it belongs to the category of spatiotemporal data.This paper takes the demand of the bike sharing system as an example.Based on the historical information of the bike usage number and the weather feature of certain place,it improves the common methods in the spatiotemporal data prediction problem to forecast bicycle demand of certain stations at target time.The main work of the thesis is:1.An iterative station clustering algorithm based on geographic location information and usage pattern information is proposed.Different from the existing clustering methods that only use the geographical location information of stations,the clustering method in this paper also considers the difference between the usage patterns of different stations.In the clustering process,the geographic location information and the usage pattern information analyzed from the historical data are comprehensively considered,and the iterative structure is used to make the clustering result more stable.After that,the clustering result is corrected with the idea of the label propagation algorithm.Through the experimental design,the clustering algorithm is compared with other existing algorithms.The clustering algorithm used in this paper can obtain more uniform and well-defined clustering results,and the results also contribute to the accuracy of subsequent prediction algorithms.2.A hierarchical prediction structure for bike sharing demand at stations is proposed.When predicting the bicycle usage of the station,unlike the traditional method of directly predicting,the prediction structure used in this paper first uses the Gradient Boosting Regression Tree algorithm to predict the total usage of all stations.Then,according to the ratio of the bicycle usage of the different stations to the total usage in the historical data,the check-out amount of each station set at the target time is calculated.After that,the bike transfer probability is analyzed by the transport model of the shared bicycle between stations sets,and the check-in quantity of each station set at the target time is calculated.By comparing with the prediction methods in the existing literature,it is found that the hierarchical prediction structure can obtain more accurate prediction results in both the total demand prediction and the class-level demand prediction.3.Improve prediction accuracy by temporal feature filtering and spatial feature extraction.After analyzing the feature data,the idea of combined optimization is used to eliminate the redundant temporal feature,and the convolution operation is used to extract the implicit spatial feature information from the historical data.Finally,shared bicycle demand is predicted by combining the filtered temporal features with the extracted spatial features.Experiments show that after performing this series of feature operations,the prediction results are improved compared with the original prediction method.