Accelerating multi-energy system online optimization via integer state variable prediction with operation strategy learning

To address the low computational efficiency of solving mixed-integer linear programming (MILP) problems during the online operation optimization of multi-energy systems, this study proposes a state variable predictive framework based on operation strategy learning that decouples the original MILP problem: First, a dataset of integer decision variables (e.g., start/stop or charge/discharge states) under varying wind and solar generation and load profiles is obtained by solving offline optimization problems. Based on the dataset of optimized operational strategies, we develop a model leveraging a deep neural network that predicts the start/stop and charge/discharge states of energy devices using future wind/solar generation forecasts and load demand time series. The proposed model achieves an average prediction accuracy exceeding 94.1 % at 5-min time resolution in typical scenarios. Building on the predicted temporal sequences, two accelerated solution strategies are introduced: Integer variable elimination acceleration and optimized initial solution guided acceleration. The integer variable elimination method fixes future device states based on the prediction, converting the MILP into a linear programming (LP) problem. This approach achieves over 90 % reduction in computational time in typical scenarios. The optimized initial solution guided acceleration method incorporates the predicted sequences into MILP solver to curtail the cardinality of the integer decision space, achieving an average of 30.1 % reduction in computational time while preserving solution quality. Case study demonstrates that the proposed approach enables fast response within online optimization frameworks and improves load satisfaction and renewable energy utilization, thereby enhancing the overall operational performance of multi-energy systems.