Digital twin-driven smart city energy management systems: an integrated approach combining big data and machine learning

Abstract

To address the issues of prediction divergence and severe lag in the execution of closed-loop control commands in complex energy distribution networks of smart cities when dealing with high-frequency nonlinear load disturbances, this paper proposes an end-to-end joint control framework driven by digital twin technology, combining deep feature extraction from big data with physical prior machine learning. This system eliminates hardware transient signal distortion by constructing a parallel digital space. Based on this, a joint topology prediction network using Long Short-Term Memory Graph Attention (LSTM-GAT) coupled with an energy Kirchhoff constrained functional is designed and trained. Furthermore, a null neural network (ZNN) control law with continuous-time derivative feedforward compensation is innovatively implanted into the execution layer, completely blocking the multi-level compound error accumulation chain of prediction drift to the physical execution layer. Rigorous comparative experimental data from offline and edge heterogeneous computing platforms show that, when facing extreme anti-mutation disturbances, the mean absolute percentage error (MAPE) of this fused scheduling network successfully converges to an extremely low threshold of 1.85%, and the system's single-time multi-dimensional perception command closed-loop issuance delay is compressed to an extreme physical boundary of 18.4ms. The quantitative mediation effect verification system confirms that the noise reduction physical smoothing mediation transmission efficiency of the architecture reaches 0.308, which greatly broadens the theoretical control extreme bandwidth of the urban wide-area multi-source heterogeneous equipment cluster without overshoot collaborative operation.