Research On Deep Forest And Reinforcement Learning Method For Building Energy Saving

For a long time,with the rapid development of China’s economy,it is not only the continuous improvement of people’s living standards,but also the increasingly serious energy consumption problems.Among them,the total energy consumed by buildings is the total energy consumption of society The proportion in China is constantly increasing.According to the "2018 China Building Energy Research Report",in 2016,China’s total energy consumption for the year was 899 million tons of standard coal,accounting for 20.62% of the country’s total energy consumption,of which public building energy consumption accounted for building energy.The total energy consumption is 38.53%,urban residential building energy consumption accounts for 37.71%,and rural building energy consumption accounts for 23.76%.Related data is still rising.The increasingly prominent problem of high energy consumption in buildings has made building energy saving a research focus in the field of smart buildings.The prerequisite for effective energy saving in buildings is to be able to establish a stable and effective prediction model for building energy consumption,based on accurate predictions.Do related energy saving work.Building energy consumption prediction is a complex nonlinear problem,and its data is often multi-dimensional.When processing multi-dimensional data,the decision tree is based on its layer-by-layer division and node-by-node growth processing methods.It has clear logic and visible structure.Advantages,so this study combines deep forests based on decision trees and reinforcement learning,using reinforcement learning as a whole structure,m GBDT as an agent for data processing,and in the process of interacting with the environment,predicting the next time period Energy consumption.It mainly includes the following three parts:(1)Aiming at the disadvantage that the traditional deep neural network belongs to "black box" and the decision-making process is not visible,a new Double-m GBDT-based Q-Learning algorithm is proposed.The algorithm uses m GBDT instead of DNN,which introduces m GBDT as a function approximator.In the learning process,based on the state obtained through interaction with the environment,the Bellman equation is used to construct the target value,and on this basis,the m GBDT is trained in an online manner.Like DQN,this paper also uses two m GBDT frameworks to solve the problem of easy overestimation.In order to verify the performance,the proposed algorithm and DQN and m GBDT are applied to the Cart Pole and Mountain Car problems in Open AI Gym.The results show that the algorithm can converge to the best strategy,and prove that compared with DQN,the stability of the algorithm after convergence is better.(2)The algorithm proposed in the first part is used in the field of building energy consumption prediction.Since the energy consumption of air conditioning accounts for a large proportion of building energy consumption,this part mainly uses the algorithm of the first part for the energy consumption of air conditioning.prediction.From the analysis of the influencing factors of air conditioning energy consumption,the chiller water inlet and outlet temperature of the chiller,the inlet and outlet temperature of the cooling water,the chilled water flow,the cooling water flow and the cooling capacity are selected as input variables.Due to the occurrence of unexpected conditions in the process of data collection,the data will be abnormal.Therefore,the pre-processing of the collected raw data should be performed.In order to avoid the adverse effects of abnormal values on the model,the K-means algorithm is used.To exclude outliers,secondly,in order to enable the model to learn more general laws within the data,and also to improve the accuracy of the model,normalize the data.Use the processed data to train the model.Finally,using the data collected by an environmental college to conduct experiments,the experimental results show that the Double-m GBDT-based Q-Learning algorithm can predict the overall trend of air conditioning energy consumption,and can provide guidance for the overall energy consumption prediction of the building.(3)Based on the problems found in the second part of the experimental process of air conditioning energy consumption prediction,make corresponding improvements in the data preprocessing stage.Apply the algorithm of the first part to the energy consumption prediction of the whole building.Select indoor temperature,humidity,outdoor temperature,humidity,weather and time as input variables,and then preprocess the input variables to make a training set.Secondly,combined with the framework of reinforcement learning,construct an overall training process,use the training set for model training,and finally,verify the feasibility of the prediction model through experiments.