Short-term forecasting of multienergy loads is of paramount significance for integrated energy systems operation. The central forecasting framework is confronted with the privacy disclosure issue. Besides, the intricate interdependencies among diverse energy loads present an opportunity to improve prediction accuracy. To this end, a privacy-preserving spatial-temporal adaptive personalized federated learning model is proposed in this article. Specifically, the proposed federated learning-based decentralized framework enables the sharing of local model weights while ensuring the confidentiality of raw measurement data. Besides, the spatial-temporal transformer leverages the self-attention mechanism to synchronously capture the complex dynamic dependencies among different types of energy load demand. Furthermore, the adaptive local aggregation mechanism is proposed to personalize the local model to address the data heterogeneity and subsequently improve forecasting accuracy. The proposed model is applied to a publicly available dataset. The results show that the proposed model can achieve highly efficient and effective forecasting accuracy.