| At present,China’s urbanization construction is at high speed development stage;however,the urban fire risk is high.Moreover,the city fire protection planning in China is relatively lagging behind the rapid development of urbanization,especially for the future development of cities and the development of fire risk considering comprehensive enough.Therefore,we should consider more about the space and time under the influence of dynamic distribution of fire risk,as well as the layout research of fire resources.In this paper,by introducing the related mathematical model to comprehensively consider the effect of temporal variation of urban fire risk,and fully considering the spatial and temporal variations of fire risk,we can get a dynamic fire resource layout optimization model,which can provide theoretical support for improving the urban fire protection planning on the basis of the development of a compatible with the urban development.The main work is as follows:1.Through the use of machine learning(ML)algorithms and the spatial econometric models(SE models)at the city scale for the spatial modeling of infrastructure fires,we can make prediction or explanation for the spatial distribution of fire risk.In general,for the applicability of the model,we find that the random forest algorithm can show stronger prediction ability and stronger robustness,especially for the time when independent variables are changed and the dependent variable can be predicted more accurately.As to the SE models,they are able to find some useful regularity hidden in a particular data set(training set)about its spatial distribution and clustering features by using spatial weight matrix,thus we can accurately depict the space regularity of fire occurrence and the influence of the independent variables at the city scale.The final results show that SE models are more adequate to explain the occurrence of fire,and SE models are generally superior to the random forest algorithm.According to the results of resampling from the training sets,random forest has the highest prediction precision among all the ML models.At the same time,the distance to the nearest fire stations,the population distribution,enterprise density,road density,the distribution of temperature as well as the elevation are selected in RF.The result shows that these variables contribute the most to fire risk.So we can also infer that,with the passage of time,and when the spatial distribution of population,POI and roads are changed,there will be major changes in the spatial distribution of fir risk at the city scale.Through the analysis of the spatial econometric models,the statistical results show that the spatial autocorrelation model(SAC)has the minimum value of AIC,which is the best fitting SE model.In addition,through analyzing the regression coefficients of five SAC sub models,the results show that the road density has positive correlation with fire density,whereas the temperature,the altitude and distance to fire stations have a negative correlation with the fire density.The result shows that human activities have a quite important role on fire occurrence,and this variable exactly has a relatively large contribution to the spatial regression model.At the same time,the SAC model which performed well only among five training sets,but not among the testing sets.Reasons mainly comes from that for SAC,the parameter estimation is more dependent on the existing spatial weight matrix and the lack of parameter adjustment ability for the change of spatial structure.This suggests that the SE models are more suitable to explain the fire risk.Based on the prediction results extracted from SAC and RF on the entire data set,the results show that the SAC in general is better than the RF.Further through the analysis of residuas of the model and data visualization,the results show that SAC is better than the RF and SAC can explain the spatial structure of fire occurrence well.2.By using wavelet transform for infrastructure fires and the related meteorological factors in time series,the result can discover the potential regularity in time series after using the classical autoregressive moving average model and further by using markov chain monte carlo(MCMC)to fit the time sequence of fire frequency.Thus,we can get the most suitable mathematics model for predict fire time series and obtain the trend of fire frequency from the time dimension.By using wavelet analysis for the time series of fire frequency,there are four characteristic time scale including 4,18,34 and 56 months espectively,and the biggest cyclical fluctuation of fire frequency is 56 months(4 to 5 years),which shows that,we should better study fire at large interannual scale periodic.We can also speculate from the qualitative results that ater the period of 2010-2011,the total fire frequency in Hefei city will decline,and then the fire frequency will continue to rise.We studied the Meteorological factors of time sequence by using wavelet analysis and finally found that there have different characteristic time scales for the climate variables.They all contain 18 months of characteristic time scales,which suggests that the time series of meteorological factors and fire frequency has the same average cycle of 12 months.By using the classical time series model to model fire frequency and after several comparisons of different sets of parameters,we finally choose ARIMA(0,1,1).On the other hand,we used MCMC model which conbined the steady state markov chain with bayesian equation of the combination of parameter estimation to estimate the maximum likelihood function for time series modeling.And the result shows that the given model equation has effective convergence and the kernel density estimation function is smoothy,further evidenced by the stable convergence iterative history.In addition,the results show that coefficient value of the sunshine duration is the largest,which means that it has the biggest impact on fire.At the same time,the sunshine duration and the relative humidity of the coefficient value is positive,said positively related to the frequency of fire;The coefficient value of the average temperature and precipitation is negative is negatively related to the frequency of fire.In addition,the MCMC model with the removal of meteorological variables and the MCMC model containing the meteorological variables were compared.The results show that the MCMC model with meteorological factors fit better,which can better draw the outline of volatility in time series.Finally,as regards to the correlation coefficient between predicted values and actual values,the time sequence model of the classic examples in this chapter shows that the fitting effect is slightly better than MCMC,but the difference is not significant.Moreover,we can not get the relationship between the various meteorological factors and fire frequency or make the rationality of the quantitative evaluation.3.Firstly,we performed a spatial cross validation for a linear regression model and compared its results with a stochastic cross validation.The contribution to fire risk by variables varied in different sub training sets and we infer that this kind of nonstationary situation also existed across space and that SCV could reduce the prediction error.The results also showed that the variables LINE and ENTERPRISE were the most important variables for modeling fire risk,although their effects on fire risk were opposite to each other.The results showed that constant coefficient models like LM did not predict fire risk accurately and could not reveal the spatiotemporal heterogeneity.The statistical results highlighted the weakness of the LM considering the low R-squared value.With regard to GWR-based methods,the statistical performance was improved when compared to OLS and the GTWR was the best model.With regard to the spatial distribution of the residuals,the GTWR model showed lower values of semivariance than the GWR and the LM,as well as a flat semivariance line for the entire distance.The results showed the strong ability of GTWR to explain the spatial structure.For the validation process,GTWR is proved to be more robust because the model has the lowest RSS value,which indicates that for our dataset,the GTWR is the best of the fitted regression models.4.This chapter primarily used the predicted spatial distribution of fire risk for fire station location analysis and further comparison between different models.By using genetic algorithm for the location model simulation,the results shows that the model based on matching degree and multiple parts covering location model can reflect better for the model during the planning idea of actual fire "rank corresponding",namely different levels of risk should be corresponded with the levels of preventive measures.The higher the risk,the smaller the response time for fire stations should be,while the relief path should be shorter.Therefore,we have established a new model compared with other classical location model(including P-value,maximum coverage),and the new model shows cbetter performqnce for its rationality and practical significance.In addition,we use the path optimization algorithm based on capacity constrained routine planning to simulate the situation when concurrency fire occurs.According to the results of the shortest path optimization algorithm based on CCRP,CCRP can solve the shortest path selection problem for the fire engines based on the urban road network,but the current allocation of fire engines can cause a waste of resources.But if we change some settings of the uncertain environment,do not ignore the shortest time limit and considering the actual situation with more constraints,the results will be more reasonable.At the same time,it can also be considered as choosing the shortest path problem when a crowd of people on the node of the space for group evacuation. |
PDF Full Text Request