Urban fire station site selection prediction method based on GMM clustering and machine learning

Abstract

The invention discloses an urban fire station site selection prediction method based on GMM clustering and machine learning, and the method comprises the steps: firstly, determining a clustering center of a research region through employing a GMM clustering algorithm based on POI data, building a Voronoi diagram in consideration of the boundary features of the region and county on this basis, and achieving the reasonable division of the research region; secondly, based on consideration of urban fire risk distribution, road accessibility and the like, multi-source spatial data such as POI, road density, population density and the like are fused, and a comprehensive feature matrix is constructed; and finally, training and evaluating by respectively adopting five machine learning models, namely a random forest (RF), a gradient boosting decision tree (GBDT), a support vector machine (SVM), an extreme gradient boosting (XGBoost) and logistic regression (LR), and selecting a GBDT model with optimal F1 score and optimal AUC value to predict and analyze the site selection of the fire station. According to the method, site selection prediction is performed on the basis of region division based on a GMM model and a Voronoi diagram, so that the defect that direct gridding of a research region neglects significant differences of different regions in aspects of population density, urban functions, terrain and the like can be effectively avoided, and the subjectivity of determining the size of a grid unit can be overcome; and rationality and practicability of site selection of the fire station are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of public service facility planning and management, and more specifically to a method for predicting the location of urban fire stations based on GMM (Gaussian Mixture Model, GMM) clustering and machine learning technology. Background Art

[0002] In urban spatial decision analysis, traditional facility location models mainly include P-median models, P-center models, and set covering models. These models are usually NP-hard problems and rely on deterministic optimization methods to determine the location of facilities by maximizing the coverage of demand points or minimizing distances. These models have a strong theoretical basis in facility location problems and have been widely used in certain specific scenarios, but they have limitations when dealing with complex urban environments. They usually assume that the distribution of demand points is uniform and it is difficult to fully consider the heterogeneity within urban areas. In addition, these models have limited ability to handle the uncertainty and dynamics of data, do not fully combine the diverse and complex demand characteristics in actual cities, and are difficult to adapt to the rapidly changing needs of modern cities.

[0003] In recent years, with the improvement of data science and computing power, machine learning algorithms have also been applied to site selection problems. Compared with traditional models, machine learning algorithms can handle large-scale and complex spatial data, showing stronger adaptability and predictive capabilities. However, current site selection studies based on machine learning usually directly divide the study area into grids of equal size. Although this method is simple and easy to operate, some grids may span multiple functional areas, resulting in mixed features, and to a certain extent ignore the differences between urban areas. It is difficult to capture the unique needs of different areas in the city, which in turn affects the rationality of the site selection results; in addition, the boundaries of districts and counties in traditional grid divisions are usually fuzzy, and the determination of grid unit size is highly subjective.

[0004] At present, the research on fire station site selection prediction based on machine learning and multi-source spatial data is still relatively limited. Based on the above reasons, the present invention takes fire stations as the research object and proposes a site selection prediction method that takes into account the heterogeneous characteristics within the city. Summary of the invention

[0005] The present invention provides a method for predicting the site selection of urban fire stations based on GMM clustering and machine learning. As a clustering method based on probability density estimation, the GMM model can effectively identify the internal differences of urban areas through a flexible soft partitioning method. Combined with the Voronoi diagram, a more reasonable spatial partition can be performed according to the characteristics of different areas. Figure 1This is the overall process of the urban fire station site selection prediction method based on GMM clustering and machine learning in the present invention, including four parts: data collection and processing, regional division and feature construction, model training and evaluation, and analysis of optimal model prediction results.

[0006] The present invention first uses the GMM model to cluster the study area, fully reflecting the regional differences. On this basis, the Voronoi diagram is constructed taking into account the characteristics of district and county boundaries to achieve a reasonable division of the study area and improve the scientific nature of the site selection decision. Afterwards, based on the consideration of urban fire risk distribution and road accessibility, multi-source spatial data such as POI, road density, and population density are integrated to provide rich input information for model training. Finally, a variety of machine learning models are established to predict the optimal location of fire stations, and the performance of the models is evaluated and compared. The prediction results and driving factors are deeply analyzed to provide a scientific basis for the optimal layout of urban fire stations.

[0007] The urban fire station site selection prediction method based on GMM clustering and machine learning proposed in the present invention is easy to implement, simple and efficient, improves the accuracy and representativeness of regional division, can flexibly adapt to the complex functional layout of the city, and provides a more accurate reference basis for the scientific site selection of fire stations.

[0008] Explanation of Figures and Tables

[0009] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings and tables required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are merely schematic diagrams and tables of the present invention. For ordinary technicians in this field, other drawings and tables can be obtained based on the provided drawings without paying any creative work.

[0010] Figure 1 A flow chart of the method for predicting the location of urban fire stations provided by the present invention;

[0011] Figure 2 Schematic diagram of the solution process of the GMM algorithm;

[0012] Figure 3 This is a schematic diagram of the regional division results based on GMM;

[0013] Figure 4 Fit ROC curves and AUC area result graphs for 5 machine learning models;

[0014] Figure 5 Prediction results map for site selection in the study area;

[0015] Figure 6 Schematic diagram of the importance measurement of each indicator feature. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The present invention selects the central urban area of ​​a city as the test set area, and uses other districts and counties of the city (except the central urban area) as the training set and verification set areas for example verification.

[0018] The following steps are the region partition part based on GMM:

[0019] Step 1: The crawler obtains POI data reflecting important functional points in the city. These data include residential areas, commercial areas, industrial areas, hospitals, schools, etc., covering areas where people’s activities are concentrated and areas with high fire protection needs, and can more comprehensively characterize the characteristics of fire protection needs in the study area.

[0020] Step 2: Use GMM to perform cluster analysis on POI data. Different from traditional hard clustering algorithms (such as K-Means), GMM is a soft clustering algorithm that can assign each data point to different clusters in the form of probability based on its characteristics. This flexible allocation method is particularly effective when dealing with POI data with overlapping distributions, and can more accurately identify areas in the city with similar fire protection needs. The GMM algorithm solution process is as follows: Figure 2 shown.

[0021] The Bayesian Information Criterion (BIC) is used to select the number of clusters for the GMM algorithm. BIC is a commonly used model selection criterion that can strike a balance between model fitting accuracy and model complexity, and avoids over-complexity of the model by introducing a penalty term, thereby preventing overfitting.

[0022] Step 3: Construct a Voronoi diagram based on the cluster centers and divide the study area into several irregular polygonal areas for further data processing and analysis. In addition, the boundary characteristics of each district and county are taken into account when dividing the area, avoiding the errors that may be caused by the fuzzy regional boundaries in traditional grid division. The regional division results are shown in Figure 3. Figure 3 shown.

[0023] The following steps are part of the feature matrix construction:

[0024] Step 4: Feature selection. Based on a comprehensive consideration of firefighting needs, key factors such as PO1 distribution, road accessibility, population density, and economic level in the region were included in the indicator selection, and the urban fire station site selection indicator system was constructed as shown in Table 1.

[0025] Table 1. Selection of urban fire station site selection prediction features

[0026]

[0027]

[0028] Step 5: Feature standardization. Since the dimensions and scales of the features may differ greatly, the features need to be further standardized. The present invention uses the Z-score standardization method to eliminate the proportional differences between the data without changing the spatial distribution of the original data.

[0029] The following steps are for the site selection prediction part of the city fire station:

[0030] Step 6: Model training. Different models have their own advantages and disadvantages when processing complex spatial data. Therefore, the present invention selects five widely used machine learning models for performance comparison to determine the optimal model for the final site selection prediction. The five models are: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost) and Logistic Regression (LR). The model is built based on Python and its built-in "scikit-learn" toolkit. Due to the imbalance of positive and negative samples, the SMOTE (Synthetic Minority Over-sampling Technique) oversampling method is used to balance the data set to improve the model's ability to classify positive and negative samples.

[0031] Step 7: Model reliability evaluation: The present invention combines the confusion matrix and the AUC area under the ROC curve to evaluate the model performance.

[0032] 1) Confusion matrix: Based on the confusion matrix, we select precision, recall, and F1 score to measure the classification performance of the model. Five machine learning models are evaluated, and the results are shown in Table 2.

[0033] Table 2 Performance evaluation results of five machine learning models

[0034]

[0035] 2) AUC area: By calculating the AUC area under the ROC curve, the model performance can be more comprehensively tested. The ROC curves and AUC values ​​of the five models are shown in Figure 2. Figure 4 shown.

[0036] Based on the performance evaluation results of the five machine learning models, the GBDT model performed best in precision, recall, F1 score and AUC value, with an AUC value of 0.96, showing strong classification ability and wide adaptability.

[0037] Step 8: Comparative verification. In order to verify the effectiveness of the regional division method used in this invention, a regional gridding comparison experiment was conducted. The study area was divided into grids with a resolution of 1km×1km, and corresponding data processing was performed. On this basis, 5 machine learning models were retrained and evaluated, and the average F1 score was 0.7981. This further proves the effectiveness of the regional division method used in improving the overall performance of the model.

[0038] Step 9: Site selection prediction analysis based on the GBDT model. Based on the evaluation of the performance of various models, the GBDT model was finally selected to predict the site selection of fire stations in the study area. The prediction results are as follows: Figure 5 There are 9 existing fire station locations in the study area, and the model predicts 12 suitable fire station locations, 8 of which fall into the same block as the existing fire station locations. The predicted locations and numbers of the 12 fire stations are highly consistent with the existing stations, which shows that the model has successfully identified the rationality of the existing fire station locations to a certain extent, and further verified the reliability of the prediction results.

[0039] Step 10: Driving factor analysis. The feature importance calculation of the GBDT model is determined by evaluating the contribution of each feature to the decision when building the tree model. According to the feature importance ranking returned by the GBDT model ( Figure 6 ), road density is the most important feature, accounting for nearly 30%, which shows that the accessibility of the urban transportation network plays an important constraining role in the site selection of fire stations, especially in areas with high-density transportation networks. The reasonable layout of fire stations can effectively shorten the response time and thus improve the efficiency of fire services.

[0040] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for predicting the location of urban fire stations based on GMM clustering and machine learning, characterized in that: include: S1: GMM clustering is performed based on POI data, and the number of cluster centers is determined using the BIC criterion; S2: Construct a Voronoi diagram based on the cluster centers to divide the study area; S3: Considering the key indicators such as POI distribution and road accessibility in the region, a site selection index system for urban fire stations is constructed; S4: feature standardization using the Z-score method; S5: Construction of model dataset; S6: Train and verify 5 machine learning models respectively; S7: Evaluate the reliability of the models and select the model with the best performance for subsequent analysis; S 8: Conduct fire station location prediction analysis based on the optimal model and compare with the number and location of existing fire stations; S9: Returns the feature importance ranking of the model and performs driving factor analysis.

2. The method for predicting the location of urban fire stations based on GMM clustering and machine learning according to claim 1 is characterized in that: In steps S1 to S2, the Voronoi diagram is used to divide the regions based on the GMM cluster centers and taking into account the boundary characteristics of each district and county.

3. A method for predicting the location of urban fire stations based on GMM clustering and machine learning according to claim 1 or 2, characterized in that: In steps S3 to S4, a feature matrix is ​​constructed from four aspects: fire risk distribution, road accessibility, population density, and economic level. The comparability and adaptability of the model are improved through feature standardization.

4. A method for predicting the location of urban fire stations based on GMM clustering and machine learning according to claim 1, 2 or 3, characterized in that: In steps S5 to S9, site selection prediction analysis is performed based on regional division based on GMM clustering combined with Voronoi diagram.

Images