A Fully Connected Deep Neural Network with Multi-Head Self-Attention Mechanisms Based on the Multi-Objective Ant-Lion Optimization Algorithm for Low-Carbon Economic Dispatch

As global climate change intensifies, decarbonizing the power system is key to achieving carbon-reduction targets. Inspired by the multi-objective ant-lion optimization (MALO) method and deep neural networks (DNNs), this study proposes an innovative approach to address the complex carbon reduction challenges faced by energy-consuming enterprises. The proposed fully connected deep neural networks with multi-head self-attention mechanisms based on the multi-objective ant-lion optimization algorithm (FCDNN-MHSAM-MALO, FMM) integrate the advantages of MALO and DNN. MALO plays a key role in optimizing the decision variables in economic scheduling. By contrast, DNN predicts the search direction of the optimal solution by learning historical optimization results. This reduces the number of iterations, improves computational efficiency, and speeds up the solution process. The multi-head self-attention mechanism calculates the importance of various input features, enabling the model to focus on factors that significantly impact the scheduling solution. This attention-driven approach improves prediction accuracy and enables MALO to optimize from an earlier starting point, thus achieving global convergence more efficiently. Compared with several state-of-the-art algorithms on IEEE 118- and IEEE 300-bus systems, the simulation results show that (1) both carbon dioxide emissions and costs can be reduced: carbon emissions are reduced by at least 1.02% and the cost is lowered by at least 0.64% when the FMM algorithm is adopted; (2) better real-time performance: an at least 17.11% reduction in computation time is achieved using FMM; and (3) better stability performance: the curves obtained by FMM for the two cases are closer to the Pareto frontier.