Design and optimal scheduling of forecasting-based campus multi-energy complementary energy system

This study presents a complete campus multi-energy complementary energy system (MCES), including an accurate forecasting model, efficient MCES model, and effective multi-objective optimal scheduling strategy to better utilize renewable energy. A hybrid forecasting model, including multi-scale mathematical morphological decomposition, a bidirectional long short-term memory network, and subsequences to the original sequence (S2O) manner based on the rolling approach (RA), is utilized to forecast renewable energy variations. RA continuously updates input datasets to improve forecasting accuracy. Decomposition and forecasting modules are employed in an S2O manner to reduce the number of required modules and forecasting cost. The volatility of renewable energy is mitigated by supplementing energy sources with storage. During operation, the conversion times of different energies are reduced by reasonably planning the energy supply sequence based on different loads on the demand side, increasing the energy utilization rate. The proposed multi-objective optimal scheduling strategy includes a stacked multilevel-denoising autoencoder, non-dominated sorting genetic algorithm-II, and deep reinforcement learning (DRL) for surrogate-model building, Pareto frontier establishment, and optimal solution selection. This is the first study to use DRL to select the final optimal solution. A performance comparison confirms the proposed model effectively decreases costs and pollution while increasing thermal comfort.