Abstract: To address the challenges faced by dispatchers in analyzing abnormal power grid states—caused by the massive volume of alarm information generated by power grid equipment under the integrated grid regulation and control mode—this study proposes an intelligent risk assessment methodology for temporal sequence data of power grid alerts. First, the power grid alarm information is pre-processed, and multi-group segmentation is performed based on temporal features. Next, the textual alarm information is structured, and text data features are extracted and clustered using term frequency-inverse document frequency and the K-means algorithm. Subsequently, an alarm information transition probability model based on the long short-term memory neural network is constructed, followed by the development of an alarm information risk assessment model to calculate the risk vector of alarm sequences for risk evaluation. Simulations and risk assessments are conducted using alarm information from substations at various voltage levels within a specific power grid. Compared to the risk assessment model based on a first-order Markov model, the risk assessment model based on the long short-term memory neural network increases the high-risk rate by 0.3594%, 0.0360%, and 0.0184% for 35 kV, 110 kV, and 220 kV substations, respectively.