Abstract: This paper aims at implementing a more accurate forecast of crime trend based on historical data. Taking advantage of the special property of recursive structure of recurrent neural network that can transfer the characteristics of the previous period to the current period, a method for forecasting crime situation was proposed based on improved LSTM network. First, the number of crime events within each time unit was counted as the value of a time node and multiple time nodes would form a time series. The time series were standardized through a mean square variance filter to access the training data of the network. Then, the proposed LSTM network was constructed that included the input layer, the hidden layer, the fully connected layer and the output layer. At the training stage, prediction of the last time unit was improved by relying on the actual values rather than on the forecast values. Data consisting of the criminal records of the Los Angeles area in 2016 was used in the experiments. Compared to the traditional method, the root mean square error (RMSE) of the predicting results dropped from 139.65 to 85.88 using the proposed method, which demonstrated the effectiveness of the proposed method by showing better prediction of the trend. Additionally, it shows that this method is superior to other existing methods in terms of both time performance and accuracy.