基于北京市火灾统计数据的时间序列分析

Time series consequential analysis on the statistics of Beijing fire-disaster data

采用Box-Jenkins法分析了城市火灾时间序列上的趋势性规律,通过数据预处理和模型的识别与诊断,最终建立了城市火灾预测模型。对北京市2000—2006年的火灾统计数据进行了平稳性处理,通过城市火灾预测模型的建模方法对北京市火灾发生次数的时间序列进行了ARIMA建模。根据得到的ARIMA(1,1,0)(1,1,1)12模型对北京市2007年的火灾发生次数进行了预测,并将该模型和BP神经网络模型的预测结果与实际情况进行对比分析。结果表明,ARIMA模型特别适用于随机性较大的火灾数据的趋势预测,并且方法简单,算法经济。

The paper intends to introduce the advantages and proba- bility in prediction-making problem, the theoretical basis and struc- tural form of the ARIMA model （ auto regressive integrated moving av- erage model）. First of all, it is necessary to introduce the structural procedure of the city fire incidence prediction model, whose main steps are data preprocessing, model identification, order confirmation and parameter estimation, model diagnosis and examination, and the prediction analysis to be done by using the model that can withstand the testing. Secondly, the paper has made a tendency analysis of the city fire on time series by the Box-Jenkins method. Thirdly, a city fire forecasting model has been built through the data stationarity pre- treating, model identification and diagnosis. Having analyzed the fire- disaster statistical data that took place in Beijing from 2000 to 2006, it can be found that the data stability fail to satisfy the requirements of the data stationarity test. Then we have applied the logarithmic differ- ence operations on the time series data, which are qualified to pass the stationarity test. In so doing, we have determined seasonal ARI- MA model parameters according to the model identification and evalu- ation, and gained acceptable results. In the end, we have established a city fire prediction model for predicting the city fire incidence time series in Beijing. We have made our prediction of the urban fire dis- asters in 2007 in Beijing municipality sphere according to the ARIMA （1, 1, 0） （1, 1, 1）12 model and the BP neural network model, and gained forecasting results and compared our forecasting results by us- ing the ARIMA model with the actual numbers and time series that took place in the city and found both of the results very close. The average relative error of the predictive value and the actual value is only 0.251 5%, which indicates that the forecast data results are much higher than those offered by the BP neural network model. Therefore, it can be said that the ARIMA model is feasible for the random-predictable data, such as the fire forecasting trend. The method we have developed is relatively simple, with the use more e- conomical algorithm. And, therefore, its application prospect is bril- liant and promising.