Research On Semi-supervised Spatial Clustering Method And Its Application In Urban Public Facility Location Planning

Urban public facilities include social facilities and civil engineering facilities, which are the basis for maintaining the normal functions of a city. The benefit of public facilities is directly affected by the reasonable location planning of public facilities. Effective space configuration to maximize the efficiency and equity benefits of public facilities is the fundamental characteristic to differentiate between public facilities and non-public facilities. The reasonable layout of urban public facilities is not only an important factor for the spatial structure of a city and life quality of residents, but also an important part of urban planning. At present, most related researches on public facility location planning are based on the classic facility location models, which take relatively little account for the effect of obstacles. Spatial clustering is one of the most important methods for analysis of the agglomeration degree and distribution patterns of public facilities. The spatial clustering method using domain knowledge can more effectively guide the spatial clustering process.The semi-supervised clustering method is studied in this dissertation, and a semi-supervised clustering algorithm based on the gravitation model with pairwise constraints is proposed. Next, path searching methods based on vector data model and raster data model are presented respectively. And it is applied to the public transportation network optimization problem of Wuhu. Finally, a spatial clustering method with obstacle constraint is proposed, which is applied to the public facility location planning problem of Wuhu.The main work and innovations are as follows:1. Based on the semi-supervised clustering method, aiming at the problem that affinity matrix constructed by traditional distance functions can not adequately represent similarity among data points, the affinity matrix is constructed based on the gravitation model. Considering each sample in a data set as an object in the feature space, one object is allowed to move toward another one under the influence of universal gravitation. A gravitational inspired constrained spectral clustering algorithm based on the adjusted affinity matrix is presented. The experimental results indicate that the average Accuracy and CRI of our algorithm present a trend of increasing, as the constraints rates increasing. And our algorithm outperforms all the other compared methods in most case. Even more important is that the proposed algorithm seems more stable over different data sets.2. Taking public transportation network optimization as application background, path searching methods based on vector data model and raster data model are presented. The optimal collision-free path can be planned rapidly in the workspace with multiple obstacles. And the path planning method based on the improved ant colony optimization algorithm is applied to the public transportation network optimization of Wuhu. We analyze the situation of Wuhu City public transportation network. Experimental results show the effectiveness of the optimal path in the public transportation network.3. Considering the impact of obstacle entities and facilitator entities, the method of calculating the obstacle distance between two points in the space is presented. Taking obstacle distance as similarity metric, based on the clonal selection and memory mechanism of the artificial immune system, the spatial clustering method with obtacle constraint is proposed subsequently. Our method is also applied to the public facility location problem of Wuhu in order to establish its practical applicability. In order to solve the problems of the traditional clustering algorithm in sensitivity to the initial value and the tendency to fall into local optimum, a novel spatial clustering algorithm with obstacle constraint is proposed based on artificial immune mechanism. A comparative analysis of our method and the classical clustering method demonstrates that the method has a significant effect on improving the quality of clustering.