A surveillance and alerting mechanism for video aggregation platforms based on an ensemble time series forecasting model was proposed, in order to mitigate the risks of copyright infringement and content security brought by deep linking video aggregation platforms, as well as to facilitate the prompt detection and notification of network users who engaged with such platforms through illicit means. Initially, the network behavioral log data from multiple video aggregation platforms were leveraged. The network behavior characteristics of users were then extracted with IP address as the user dimension and day as the time dimension, on both the platform side and the channel side. Subsequently, long- and short-term time-series networks (LSTNet), recurrent neural networks (RNN) and multilayer perceptron (MLP) were harnessed as foundational models to construct a Stacking ensemble learning model for predicting user access behavior by learning features from base model. Ultimately, empirical validation was conducted through comparative and backtesting experiments. Results showed that the proposed method achieved a notable decrease of 0.9724 in mean squared error (MSE), a significant reduction of 0.5443 in mean absolute error (MAE), and a moderate improvement of 0.20 in balanced accuracy (BAC). The proposed method could effectively forecast access patterns to video aggregation platforms and provide early warnings for high-risk user behavior.