Abstract: To address the challenge of simultaneously capturing linear trends and nonlinear fluctuations in the security situation prediction of 5G private networks within the Industrial Internet of Intelligence, we propose an ARIMA-LSTM-based prediction architecture for semantic separation between business and security features. Focusing on the 5G kite private network deployed in key industrial scenarios, we target two types of operational conditions: static stability scenarios and dynamic elasticity scenarios. Key indicators, such as the proportion of abnormal signaling at the N4 interface and the session management function (SMF) session failure rate, are selected. The model extracts periodic linear features using the autoregressive integrated moving average (ARIMA) model and combines them with the long short-term memory (LSTM) network's ability to learn nonlinear patterns from sudden disturbances and random residuals, thereby constructing a collaborative prediction mechanism for communication data and security situational awareness. Experimental results based on real-world data show that in static scenarios, the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the proposed model are as low as 0.82 and 4.2%, respectively. These figures represent reductions of 81.4% and 66.2% in RMSE, and 78.9% and 71.6% in MAPE, compared to standalone ARIMA and LSTM models, respectively. In dynamic scenarios involving network element handovers and signaling attacks, the average increase in RMSE is only 19.5%. Furthermore, the degradation rate for 60 min long-term prediction is only 48.2% of that observed in the LSTM model. The proposed model maintains optimal performance across 6 types of security indicators, verifying its stability and robustness. This research provides technical support for elastic security protection and intelligent operation and maintenance in the Industrial Internet of Intelligence.