Integrated Energy Agent State Awareness Method Based on Multi-Agent Deep Learning Algorithm
Abstract
As the global energy system accelerates its transformation to clean, low-carbon, and intelligent systems, the integrated energy system, as a key carrier supporting multi-energy synergy and complementarity, has significantly increased its operational complexity and dynamic uncertainty. Traditional centralized state awareness methods struggle to cope with the problems of nonlinear interference and incomplete information caused by multi-energy flow coupling, making it urgent to break through the limitations of local identification for a single agent. This paper proposes a comprehensive energy agent state awareness method based on a multi-agent deep learning algorithm, which enables cross-level mapping of local observations and global state inference by constructing a distributed agent network. This method utilizes dual upsampling fusion technology, combined with deep and shallow feature extraction modules. The deep feature extraction unit integrates the ShuffleNet v2 efficient network, resulting in a parameter count approximately half that of an ordinary network. It also introduces the SE (Squeeze-and-Excitation) channel attention mechanism to enhance the capture ability of key information. Experiments based on the I-BLEND dataset demonstrate that the model achieves an accuracy of 97.21% after 52 months of power data training, which is significantly better than the comparative models' accuracies of 90.10% and 93.50%. The state awareness results show that the external state prediction error of the agent effectively converges within the steady-state covariance ellipse region. The position error decreases to approximately 4 when the prediction time step is 2, and the minimum update frame rate increases to approximately 38. Multi-index joint analysis reveals that the power forecasting range spans-20 to 20, the price forecasting error fluctuates within the range of 0.05 to 0.20, the load level forecasting deviation is controlled within the range of 0 to 8, and the introduction of an uncertainty margin enables dynamic adaptation to changes in forecasting difficulty. The relative rate distribution of agents under different parameter combinations shows significant differences. For example, K1 agents aggregate in the relative rate interval of 80-100, while K3 agents have a state perception score of more than 0.8. Through principal component analysis optimization, the dimensionality of the model is reduced to 9 dimensions when the cumulative variance contribution rate reaches 85%, and the F ratio is as high as 4.70, which verifies the robustness and real-time advantages of the algorithm in multi-energy flow coupling scenarios.