The high penetration of renewable energy in user-side integrated energy systems (IES), combined with random consumption patterns, renders optimizations based on deterministic supply–demand scenarios inapplicable to dynamic real-world conditions, leading to a disconnect between design and operation. A bi-level optimization framework coupling Bayesian optimization with deep reinforcement learning is proposed. At upper level, a tree-structured Parzen estimator (TPE) searches for optimal equipment configurations; at lower level, a soft actor-critic (SAC) agent learns adaptive operational policies under uncertainty. By embedding equipment configuration information into the agent’s state space, the framework enables configuration-aware strategy learning. An agent reuse mechanism based on configuration similarity accelerates training by transferring knowledge across configurations, reducing the average training episodes from 184.7 to 68.0—a 63.2% improvement in computational efficiency. Validated on a multi-energy office building with 100 Monte Carlo scenarios, the framework achieves annual total cost savings of 6.2% over deterministic optimization and 2.0% over two-stage stochastic programming. Cooling penalty costs are reduced by 96.3% and heating penalty costs by 99.7% compared to deterministic baseline. Additional benchmarks against genetic algorithm-SAC and particle swarm optimization-SAC confirm that TPE outperforms population-based meta-heuristics by 1.3%–1.8% under identical evaluation budgets. Sensitivity analysis across four uncertainty levels (±10% to ±40%) and 30 extreme stress-test scenarios confirms the advantage holds across all tested levels. Overall, the framework offers a practical path toward uncertainty-aware IES co-optimization.

Accurate and efficient prediction of energy consumption in district building groups is a critical foundation for effective energy management and achieving building energy-saving and emission reduction goals. In recent years, deep learning methods have been widely applied to building energy consumption prediction. However, most existing approaches rely heavily on abundant historical data for model training. In practice, data scarcity, due to factors such as data privacy concerns or newly constructed buildings, poses a significant challenge. Additionally, many methods fail to account for spatial dependencies between buildings. To address these limitations, this paper proposes a novel prediction framework based on a spatiotemporal graph convolutional network integrated with adversarial domain adaptation, specifically designed for energy consumption prediction in district building groups under data-scarce conditions. The approach combines a graph convolutional network (GCN) based on an adaptive adjacency matrix with a graph attention network based on a dynamic time warping (DTW) correlation matrix to capture the spatiotemporal features of energy consumption, across the building group. The proposed model is first pre-trained on a data-rich source domain. Then, adversarial domain adaptation is employed to extract domain-invariant spatiotemporal features from both the source and target domains. This enables a "multi-buildings to multi-buildings" transfer of predictive knowledge from the source domain to the target domain. Extensive experiments were conducted using historical energy consumption data from two real-world district building groups to validate the effectiveness and robustness of the proposed method. Results demonstrate that adversarial domain adaptation can successfully leverage domain-invariant knowledge to improve prediction accuracy in data-scarce scenarios. Compared to models trained directly on the target domain without transfer learning, the proposed method reduces prediction error by 2.36% to 16.08%. Furthermore, when benchmarked against various mainstream transfer learning methods, the adversarial domain adaptation approach more effectively bridges the domain gap, yielding superior predictive performance. Additionally, to simulate real-world data collection scenarios, the impact of varying training data ratios on the proposed model performance was also evaluated.