Integrated Energy System Operation Optimization Based On Deep Reinforcement Learning

As the total energy consumption continues to grow,environmental issues and the balance of energy supply and demand become increasingly serious.A regionally integrated energy system containing distributed energy sources such as photovoltaic,wind power,and gas turbines can reduce the system’s carbon emissions while meeting energy demand.An efficient and accurate scheduling strategy for integrated energy systems can reduce the impact of the uncertainty of renewable energy sources such as wind and light on the stability of system operation and enhance the flexible scheduling capability among system devices.Meanwhile,accurate load prediction and wind and light output prediction are crucial to the planning,design,operation,and scheduling of integrated energy systems.Therefore,this paper researches the optimal scheduling of integrated energy systems based on wind and light output forecasts and load forecasting data.With the rapid development of artificial intelligence technology,machine learning algorithms are widely used in integrated energy system optimization and data prediction.Based on the optimization of integrated energy system operation for machine learning algorithms,the research content of this paper includes.(1)Research on multivariate load prediction of integrated energy systems.A load of an integrated energy system has random solid fluctuation,and each influencing factor has different degrees of influence on the load.Pearson correlation analysis is used to screen the load influencing factors of the integrated energy system,and a long and short-term memory neural network based on the attention mechanism is established for multivariate load prediction.The prediction model uses the attention mechanism to improve the training efficiency of the prediction model and enhance the load prediction accuracy while overcoming the gradient disappearance and gradient explosion problems.(2)Research on wind and light output prediction of the integrated energy system.This part starts from wind speed and solar radiation and adopts the indirect method to realize an integrated energy system’s wind power and photovoltaic output prediction.The APSO-CNN-LSTM algorithm is used to establish the wind speed and solar radiation prediction model.The APSO algorithm mainly selects the hyperparameters for the CNN-LSTM hybrid neural network.The CNN network extracts the features of wind speed and solar radiation,and the LSTM network performs feature learning.The APSO-CNN-LSTM prediction results are combined with wind power and photovoltaic processing mathematical models to obtain wind speed and solar radiation prediction results,which provide data support for capacity optimization and operation scheduling of integrated energy systems.(3)Capacity optimization study of the integrated energy system.Based on the load forecast data and wind speed and solar radiation data,a two-layer dynamic optimization model is established from the system planning layer and the scheduling layer to realize capacity optimization of the integrated energy system.The system planning layer takes the system investment cost and operation and maintenance cost as the objective function.In contrast,the dispatching layer takes the energy consumption cost,power generation feed-in revenue,and the power storage device’s peak and valley reduction revenue as the objective function.It uses the improved particle swarm algorithm to solve the optimization model and obtain the optimal capacity of the integrated energy system.(4)Operation optimization study of the integrated energy system.Based on the capacity optimization results,the integrated energy system optimization model is established with energy consumption cost,operation and maintenance cost,system operation benefit,and system carbon emission benefit as the objective functions and combined with system operation constraints and physical constraints.By transforming the system optimization model into a Markovian decision process,the integrated energy system is optimally solved using a deep reinforcement learning algorithm.