Optimal scheduling of multiple multi-energy supply microgrids considering future prediction impacts based on model predictive control

Renewable energy based multi-energy supply microgrids not only can cover different types of demands (such as, electricity/heat/gas), but also can interconnect with different utility grid networks (electricity/heat/gas). When there are large numbers of grid-connected microgrids, how to operate these multiple microgrids in real-time is a problem. In this paper, day-ahead stochastic optimization scheduling and real-time sliding window model predictive control are used to control the operation of microgrids. In order to consider the influence of future prediction on the current optimal decision results, different prediction methods are adopted to predict the load demands and renewable energy output. For example, online learning Markov chain prediction, and support vector machine are used to predict the future values. As for comparison, robust prediction and bilevel optimization are adopted to describe the future prediction uncertainty. The real-time operation of microgrids aims to follow the day-ahead exchanged energy with utility grids, which can minimize the impact of the microgrid on the utility grids. The supply network is an IEEE30 + gas20+heat14 network. The results show that: 1) when the sliding window number is smaller, the total operation cost is larger, but the calculation time is smaller, the trade-off between sliding window numbers and calculation time should be considered; 2) the accuracy of the prediction impacts the 2-norm error of the operation cost, when we decrease by “1” unit of 2-norm prediction error of the whole system, the 2-norm operation cost will decrease by “0.15” unit; 3) from the view of the post-event analysis (total operation cost), for the Markov chain prediction method, the relative error is about 0.32%, is better than the support vector machine method; 4) in the robust cases, the larger the conservative value, the higher the stored hydrogen energy. At last, the results of real-time sliding window model predictive control problem are influenced by the future prediction methods and the window numbers.