Voltage sag is a critical power quality disturbance that poses a significant threat to sensitive industrial loads, causing equipment trips and substantial economic losses. Traditional assessment methods are often impractical, as they rely on prohibited on-site testing or hard-to-obtain equipment parameters for simulation. To overcome these challenges, this paper proposes a novel assessment method for voltage sag-induced load losses using a Deep Neural Network (DNN). The method leverages non-intrusive monitoring data from the Point of Common Coupling (PCC) to establish voltage-time (U-T) and power-time (P-T) trajectories. By applying a mutual information technique, the model automatically detects mutation points in these trajectories to extract eight key features, including sag duration, voltage drop amplitude, and power variation characteristics. These features are then used to train a DNN model to predict the load loss rate (LLR). A case study on a simulated IEEE 33-node system demonstrates the method’s superiority. Compared to a traditional Random Forest model, the DNN achieved a 96.3% reduction in Mean Squared Error and an R2 value of 0.99938, offering a highly accurate and robust solution for load loss assessment.