Learning-based data fusion joint optimization of energy cost and resource allocation in modern energy systems

To reduce user and grid costs and realize smart grid management, this paper proposes a data fusion smart distribution scheme based on deep reinforcement learning (DRL), which contains a charging decision and joint resource allocation decision. The joint optimization problem is decomposed into two subproblems to simplify the optimization decision process by building a data fusion decision model for centralized and distributed energy storage. Since the charging and discharging decision problem for user energy storage devices is NP-hard, the complexity of the decision grows exponentially as more users join the grid. In order to obtain the optimal solution to the unloading decision problem, this paper introduces the improved Rainbow DQN algorithm. Compared with other DRL algorithms, the Rainbow DQN algorithm shows superior performance in an experimental setting. Considering the periodic fluctuation of power consumption in the grid, this paper integrates the Long Short-Term Memory Network (LSTM), Course Learning and DRL to replace the traditional Multi-Layer Perceptron (MLP) in order to capture the temporal characteristics of grid power. Experiments show that the algorithm proposed in this paper has superior decision making performance compared to baseline methods in modern energy systems.