Intelligent Dispatch Of Integrated Energy System Based On Machine Learning

At present,the construction of the integrated energy system is in its infancy,and the key technology of how to provide an intelligent dispatching strategy for the integrated energy system has not yet been clarified.The traditional method is based on the numerical optimization model to provide dispatching strategy,but this method cannot handle the high-dimensional,non-linear and multi-agent characteristics of the new integrated energy system.Considering the de-modeling characteristics of machine learning technology,the optimal solution can be explored in a complex environment to help solve the problem of intelligent dispatching of integrated energy systems.Therefore,this article focuses on the intelligent control scheme of the integrated energy system based on machine learning.The main contents are as follows:(1)For the demand-side users of the integrated energy system,our research intends to study the decision-making method for end users of the integrated energy system under the market environment.First,a smart energy community model with a local energy pool is proposed.The model maintains the price of the local energy pool within a reasonable range through the proposed pricing mechanism to promote users to purchase electricity from the local energy pool.At the same time,a reinforcement learning algorithm that can provide users with real-time decisions is presented which could help users in the energy-sharing community realize the transition from consumer to producer so that they could participate more in energy market transactions.Moreover,they could save energy costs with the assistance of intelligent storage and control equipment.Finally,the feasibility of the proposed model and algorithm is verified in the provided examples.(2)For integrated energy system generators,our research focus on how to apply reinforcement learning algorithms to solve the problem of economic dispatch for integrated energy system.First,the economic dispatch problem for integrated energy system is modeled as an infinite Markov decision process,and corresponding constraints are set to simulate the actual environment.Next,a distributed training architecture is designed for the Proximal Policy Optimization algorithm in deep reinforcement learning,thereby increasing the speed of data collection,improving the speed and quality of training,and adapting to the economic dispatch problem of the integrated energy system.Finally,this is verified on two test cases and compared with the results based on numerical optimization algorithms.(3)For the bidding decision-making problems in the integrated energy market,the solution method based on game theory requires the assumption that the information between market participants is fully shared which does not conform to the competitive environment in the actual market.While,for agent-based method,the assumption that each generator does not share information with each other,which is inconsistent with the coexistence of "cooperation and competition" in the actual market environment.To solve the mentioned problems,this thesis proposes a local information interaction layer based on graph neural network,so that neighbor information can be effectively communicated,with the help of which,final decision-making effect of market participants can be improved.Finally,the performance and scalability of the proposed method is verified in a 14-node system and a 35-node system,and the results are compared with game-theory-based model.