Abstract: Central air conditioning chiller systems often operate with partial load to meet the cooling demand at building terminals, leading to high energy consumption. Load forecasting for central air conditioning chiller units is beneficial for energy-saving renovations to reach the optimal load conditions. A prediction model for the digital twin of chiller units is proposed based on convolutional neural networks (CNN), genetic algorithm (GA), and extreme gradient boosting (XGBoost) to cope with the complexities of interactions and multi-variables in a chilled water system. First, the CNN-GA-XGBoost model is trained with historical data. Subsequently, the trained model is integrated into the digital twin system through an application programming interface (API) for real-time predictions. Finally, the predicted results are visualized within the digital twin system. Results demonstrate a decision coefficient of evaluated indicators of 0.995 by the proposed method, with an mean absolute percentage error (MAPE) of 0.82 and a root mean square error (RMSE) of 2.22. The digital twin prediction model effectively bridges physical entities and data-driven approaches, enabling accurate predictions of building air conditioning loads. Furthermore, the proposed prediction method demonstrates higher accuracy and better generalization compared with other models.