The accumulation of ice on transmission lines actually represents a considerable threat to the safe and stable functioning of power grids. Traditional prediction methods usually find it quite difficult to capture the intricate spatiotemporal characteristics of meteorological data and are rather prone to getting stuck in local optima, which in turn restricts their prediction accuracy. In order to deal with these limitations, this paper puts forward a new prediction model that is based on the Particle Swarm Optimization (PSO) neural network for the ice accumulation on transmission lines. By combining meteorological time-series data together with terrain spatial features and dynamically optimizing the network weights and hyperparameters by means of the PSO algorithm, the proposed model can enhance the prediction accuracy to a significant extent. The experimental results show that the proposed model can achieve a 23% reduction in the Mean Absolute Error (MAE) when compared to the traditional CNN-LSTM model in predicting the ice thickness within 3–72 h, thus demonstrating its robustness and practicality under complex climate conditions. This study furnishes a dependable technical support for power grid anti-icing decision-making and also presents an intelligent solution for guaranteeing the safe operation of transmission lines in icy conditions.