Research On Building Power Energy Scheduling Method Based On Deep Reinforcement Learning

In the face of rapid urbanization and the attendant problems of high energy consumption and low energy efficiency,building energy consumption prediction and energy regulation are crucial solutions.In the context of increasingly intelligent infrastructure,improved data collection technology provides great convenience for subsequent data analysis and statistics,as well as the development of data-driven building energy consumption prediction and energy control technology.Energy consumption forecasting can evaluate operational plans and improve energy supply and management,which is the basis for building energy consumption scheduling.The building energy dispatching system can timely change the operation scheme based on energy consumption prediction,achieve peak shaving and valley regulation,and balance the building load curve.Currently,domestic and foreign research focuses on a single direction of energy consumption prediction and energy scheduling control,but there is still a lack of effective solutions to the joint problem of the two.In order to solve the above problems,this article uses deep reinforcement learning algorithms as the main tool to carry out research from the overall perspective of construction operations.The main contents are as follows:(1)Aiming at the existing problem of building energy consumption prediction,a dual network prediction algorithm based on second-order time difference error is proposed to further improve the accuracy of energy consumption prediction and lay the foundation for subsequent energy consumption scheduling.This algorithm is based on traditional time difference errors,proposes the concept of N-order time difference errors,and constructs a new value function update formula based on second-order time difference errors to improve the stability of value function estimation.At the same time,a dual network model based on second-order time difference errors is proposed to construct two isomorphic value function networks,which are used to represent the state information of the building environment at the previous and subsequent two moments,collaborate to update network parameters,reduce computational errors,and further improve the accuracy of energy consumption prediction.Applying this method to actual building energy consumption prediction problems,through comparative experiments,the algorithm proposed in this section has higher prediction accuracy compared to the classic deep reinforcement learning algorithm.(2)Aiming at the problem of existing building load imbalance,a strategy selection based building load balancing method is proposed to reduce the instability of building load from the demand side.Due to the temporal nature of building load,when modeling it as a continuous action space problem within reinforcement learning,incremental evaluation is proposed,and a new strategy selection method is proposed combining the soft maximization method and the depth deterministic strategy gradient algorithm.This method estimates the state action value through incremental evaluation,providing a value judgment basis for the decision-making choice of building load adjustment.On this basis,a soft maximization strategy selection method is used to optimize strategy calculation,facilitating the model to select the most valuable strategy and guiding the model to obtain the highest cumulative reward.Based on the Open AI Gym experimental platform,the proposed algorithm was validated through the City Learn module.The experimental results show that the proposed algorithm is more stable in controlling building load compared to the classic depth deterministic strategy gradient algorithm.(3)In order to solve the scheduling problem of building energy consumption as a whole,a new general framework for building power energy scheduling is proposed,which aims to balance building load and reduce energy use costs.This method is based on building energy consumption prediction,providing the prediction results to the load balance model as environmental feedback,achieving timely learning and adjustment of control strategies,and constructing a joint function of electricity prices and rewards to achieve the goal of reducing building energy use costs.The proposed framework is validated on a dataset collected from an intelligent town.The experimental results show that the universal framework can effectively schedule building electric energy and reduce the cost of using electric energy.