Research On The Prediction Of Building Energy Consumption Based On Value Iteration Algorithm

The buildings have always been the focus of the energy conservation because of the large coverage,large energy consumption and the complex energy consumption structure.Furthermore,the building is a complex and non-linear system with multi-systems and equipment.However,the building energy consumption is affected by many factors,these complex situations make it difficult to predict energy consumption.There are many methods to predict energy consumption in building energy conservation,such as engineering method,mathematical analysis method,artificial intelligence method etc.Among them,the artificial intelligence method is widely used now and in which the reinforcement learning(RL)method has attracted the attention of many scholars and is applied in many fields.The goal of the RL is to make the agent interact with the environment to achieve the optimal policy.The ability of self-learning and online-leaning make it an important branch of research on artificial intelligence field.This paper puts forward how to predict the energy consumption of buildings and how to predict the building energy consumption accurately and quickly.This paper estimates the energy consumption states with Deep Belief Networks and predict the energy consumption with value iteration(VI)algorithm in RL.The VI algorithm has slow convergence rate,poor stability and the problem such as “the curse of dimensionality”.In order to predict the building energy consumption rapidly and accurately,this paper proposed two improved algorithms with the methods of function approximation,automatic hierarchical option,reward shaping,etc.The main researches are concluded as follows:(1)With respect to the problem of unstable and slow convergence for traditional Value Iteration algorithm,we proposed an improved Residual Value Iteration Algorithm based on Function Approximation.The algorithm combines traditional Value Iteration algorithm and Value Iteration algorithm with Bellman residual,introduces weight factors and constructs new rules to update value function parameter vector.Theoretically,the new rule for updating value function parameter vector can guarantee the convergence of the algorithm and solve the unstable convergence problem of the traditional value iteration algorithm.(2)A heuristic value iteration algorithm with automatic hierarchical option was proposed.In the learning process,the proposed algorithm introduces the acyclic trajectories method to reduce the sample data and accelerate the recognition of the sub-goals as to improve the quality of the options set.In order to avoid the shortcomings of selecting some incorrect sub-goals,the algorithm introduces the mean bounding method.The method solves the problem that the states around the sub-goals are visited frequently,which makes the sub-goals selection more accurate.Moreover,the algorithm adopts the reward shaping method to construct more heuristic information to speed up the learning rate.After completing the construction of the option set,each option will be used as the input for the abstract state in value iteration algorithm to get the best policy.(3)A prediction of energy consumption based on value iteration method was proposed.On account of the curse of dimensionality,the value iteration algorithm fails in high dimensions.Furthermore,due to the fact that the traditional value iteration algorithm cannot deal with continuous states space,this paper contributes to improve value iteration algorithm by combining a Deep Belief Network(DBN)for continuous states estimation.The output of the DBN can be directly incorporated into the value iteration algorithm as the input states set as to complete energy consumption modeling and energy consumption prediction.The proposed method is experimentally evaluated with a dataset recorded by Baltimore Gas and Electric Company.The experimental results show that the accuracy of energy consumption prediction is improved remarkably with DBN.In addition,the two improved algorithms proposed in chapter three and chapter four are applied to the energy consumption prediction experiments as to further verify the performance of the algorithms.The experimental results show that the accuracy of the energy consumption prediction with two improved algorithms are both higher than the traditional value iteration algorithm.