Buildings operate under a wide range of conditions (e.g., weekdays and weekends), each associated with different energy usage patterns. The amount of data across different conditions may vary significantly, resulting in a data imbalance issue. This issue can lead to biased neural network models that perform poorly when predicting energy demand under data-scarce conditions. To address this issue, this paper proposes a building energy demand prediction method based on multi-task neural networks. This method first identifies operating conditions using the clustering decision tree algorithm. Then, a multi-task neural network model is trained. This model incorporates shared and task-specific parameters. The shared parameters are used to capture common patterns across different conditions, while the specific parameters enable the model to adapt to the unique characteristics of each condition. This method is evaluated using one-year cooling load data from a public building. Three baseline models are selected for performance evaluation, including single-model neural networks, single-model neural networks with data augmentation, and multi-model neural networks. According to the results, the multi-task neural network model shows significantly higher accuracy than the three baseline models. The R2 and mean absolute error of the multi-task neural network model are improved by 5.2% and 24.7%, respectively, under all conditions compared to those of the single-model neural networks.