基于时间序列分解与全连接神经网络的警情长周期时间序列预测

传统的警情时间序列预测以实际的发案数量为目标,且仅能实现短期的预测,但由于警情时间序列本身固有的强随机性使预测很难达到理想的效果。根据警情时间序列数据的特点,从公安工作的实际需求出发,提出了一种基于时间序列分解与全连接神经网络的(STL-FNN)预测模型,该模型以预测警情的单日发案的风险等级为主要目标,能够实现警情风险等级的长周期预测。利用该模型对B市侵财类警情数据进行了时间序列长周期预测的实证分析,结果表明:STL-FNN模型能够实现一年的警情单日发案风险的预测,平均准确率为0.624 7,预测性能优于Holt-Winters、LSTM、Prophet和ARIMA等模型。

Long-term Time Series Forecasting of Crime with STL-FNN

Traditional crime time series prediction aims at the actual number of incidents, and can only achieve short-term forecasting. But, it is difficult to achieve the ideal effect for long-term forecasting due to the inherent strong randomness of the crime data. Based on the characteristics of crime data and the urgent needs for policing work, this paper puts forward a prediction model of STL-FNN composed of a time series decomposition method and fully connected neural network, which takes the prediction of the risk level of a single day as the main target and can predict the crime risk level for a long period in the future. The empirical prediction case of property crime in City B shows that the model could predict the risk level of every day for a year with an average accuracy of 0.6247. In other words, the model could correctly predict the crime risk level of 228 days in a year. The performance is better than the common models such as Holt-Winters, LSTM, Prophet and ARIMA.