Digital Twin-Enabled Distributed Robust Scheduling for Park-Level Integrated Energy Systems

With the deepening of multi-energy coupling and the integration of high proportions of renewable energy, the Park Integrated Energy System (PIES) 1demonstrates enhanced energy utilization flexibility. However, the random fluctuations in photovoltaic (PV) output also pose new challenges for system dispatch. Existing distributed robust scheduling approaches largely rely on offline predictive models and therefore lack dynamic correction mechanisms that incorporate real-time operational data. Moreover, the initial probability distribution of PV output is often difficult to obtain accurately, which further degrades scheduling performance. To address these limitations, this paper develops a PV digital twin model capable of providing more accurate and continuously updated initial probability distributions of PV output for distributed robust scheduling in PIESs. Building upon this foundation, this paper proposes a distributed robust scheduling method for the PIES based on digital twins. This approach aims to maximize the flexibility of energy utilization in PIESs and overcome the challenges posed by random fluctuations in PV output to PIES operational scheduling. First, a PIES model is established after investigating a park-level practical integrated energy system. To describe the uncertainty of PV output, a PV digital twin model that incorporates historical data and temporal features is developed. The long short-term memory (LSTM) neural network is employed for output prediction, and real-time data are integrated for dynamic correction. On this basis, error perturbations are introduced, and PV scenario generation and reduction are carried out using Latin hypercube sampling and k-means clustering. To achieve multi-energy cascade utilization, the objective of optimization is defined as the minimization of the sum of system operating cost and curtailment cost. To this end, a two-stage distributed robust optimization model is constructed. The optimal scheduling scheme was obtained by solving the problem using the column-and-constraint generation (CCG) algorithm. The proposed method was finally validated through a case study involving an actual industrial park. The findings indicate that the constructed digital twin model achieves a significant improvement in prediction accuracy compared to traditional models, with the root mean square error and mean absolute error reduced by 13.3% and 10.81%, respectively. Furthermore, the proposed distributed robust scheduling strategy significantly enhances the operational economics of PIESs while maintaining system robustness, compared to conventional methods, thereby demonstrating its practical application value in PIES scheduling.