The commercialization and deployment of China’s independently developed third-generation nuclear power technology HPR1000 is progressing steadily. HPR1000 is a pressurized water reactor, which is a type of light water reactor. As one of the most critical design-basis accidents in light water reactors, loss-of-coolant accidents (LOCAs) have been a focal point in nuclear safety research. However, existing studies on LOCA break size prediction, particularly for third-generation nuclear technologies like HPR1000, remain inadequate. Traditional machine learning methods exhibit significant limitations in real-time prediction, underscoring the need for more efficient and accurate models. This study proposes an attention-based convolutional neural network–long short-term memory model (ABCL model) for predicting LOCA break size in HPR1000. The model leverages an additive attention mechanism, enabling it to make highly accurate predictions using only the initial 12 to 15s of reactor data in the early stages of the incident, achieving a mean squared error (MSE) on the order of 10–4 and maintaining a relative error between 0.15% and 0.20%. Experimental results demonstrate that the introduction of the attention mechanism significantly enhances the predictive accuracy of the baseline model, improving MSE by two orders of magnitude from 0.0168. Furthermore, feature analysis reveals that early-stage data (the first 15s) are crucial for improving prediction accuracy, emphasizing the model’s superior performance with short time series. This study provides essential technical support for predicting LOCA break sizes in nuclear power plants and holds significant potential for broader applications.