Research on SARIMA-LSTM Crime Prediction Model Based on Nonlinear Combination of RBF Neural Network
Dawei Song · 2024
Addressing the limitations of existing crime prediction models in capturing the composite features of crime time series data and responding promptly to environmental changes, this paper designs a crime prediction model based on the non-linear combination of Radial Basis Function (RBF) neural network, SARIMA, and LSTM. In this model, the prediction results of crime quantities from SARIMA and LSTM undergo non-linear combination through an RBF neural network, utilizing backpropagation algorithm for weight learning. The weight matrices determined by each layer’s neurons function as the proportions of the two methods in the combined prediction. By synergizing the advantages of the SARIMA model in linear time series prediction and the LSTM network in non-linear feature exploration, the model aims to enhance predictive accuracy. Experimental comparisons with real crime data from Vancouver and San Francisco affirm that the SARIMA-LSTM model, grounded in the non-linear combination of RBF neural network, excels in capturing the composite features of crime time series data, exhibiting superior accuracy compared to other models.