毕业生知识库

随着经济的快速发展和社会的不断进步，人类对能源的需求与日俱增。在现今全球能源日益紧张的形势下，节能管理受到了世界各国的高度关注，尤其是建筑节能。目前，许多国家已建立起建筑能耗管理系统并开展能耗的分项计量，从而能够获得越来越多的建筑能耗数据，然而这些数据尚未能够为建筑节能提供充分有用的信息。

作为计算智能的重要组成部分，神经网络能够对大量数据进行学习，实现函数近似、系统辨识、模式分类等功能。回声状态网是一种新型的递归神经网络，训练过程得到了简化，能够避免传统神经网络训练算法复杂、易陷入局部极小等问题。自适应动态规划是一种基于神经网络的智能控制方法，具有较强的学习和优化能力，能够良好地解决复杂非线性系统的最优控制问题。因此，本文旨在研究采用以上两种基于神经网络的方法实现建筑能耗的分析与优化。

本文的主要工作和贡献体现在以下几个方面：

1、基于建筑能耗数据，提出基于回声状态网的办公建筑能耗预测方法，对建筑各房间插座、照明和空调三种能耗进行预测。针对回声状态网储备池拓扑结构的优化，提出几种简化的储备池拓扑结构，并将具有不同储备池拓扑结构的回声状态网应用于办公建筑的能耗预测中，比较不同拓扑结构的预测性能。实验分析表明所提出的储备池拓扑结构具有与传统回声状态网接近的良好预测性能，同时拓扑结构的参数灵敏度分析表明，各储备池拓扑结构均具有较强的参数鲁棒性。

2、根据建筑房间能耗模型，提出基于回声状态网的办公建筑房间分类方法，对建筑房间进行分类。该方法首先采用三个回声状态网分别建立办公建筑房间插座、照明和空调三种能耗的模型，然后基于所建立的能耗模型，采用第四个回声状态网，根据不同类型房间的能耗特征，将房间分为办公室、机房、储藏室和会议室等四种类型。理论分析证明了分类算法的收敛性。实验分析表明，所提出的方法能够取得较高的分类准确度，并且与几种传统的分类算法相比，所提出方法的分类效果更好。

3、在获取房间分类结果的基础上，在房间引入储能设备蓄电池作为控制变量，提出基于自适应动态规划的办公建筑能耗管理方法，根据实时电价和用电需求信息，获得最优的电池充放电控制策略，对不同类型房间的能耗进行控制和优化，从而实现办公建筑房间的能耗管理。实验分析表明，所提出的方法对于不同类型的房间均能够有效实现相应的最优电池充放电控制，从而显著减少各房间的用电花费，同时所提出的方法能够取得优于其他优化算法的节电效果。

4、引入太阳能作为可再生能源，提出基于Q学习的办公建筑可再生能源分配与电池控制方法，进一步优化房间能耗，减少用电花费。在所提出的方法中，引入两个迭代以分别近似自适应动态规划算法中的最优性能指标函数以及最优控制。实验分析围绕办公建筑中的办公室展开，结果表明，所提出的方法能够获得办公室不同季节下的最优可再生能源分配与电池控制方案，进一步减少了用电花费。此外，对比分析表明，所提出方法的节电效果优于其他几种优化算法。

With rapid economic development and continuous social progress, human demands for energy resources have been constantly growing. Due to increasingly serious global energy shortage, countries throughout the world are paying great attention to the issue of energy management and especially building energy saving. At present, many countries have established building energy consumption management systems and carried out energy consumption subentry measurements, which have produced an increasing amount of energy consumption data, however, little useful information has been extracted from these data to facilitate building energy saving.

As an important part of computational intelligence, neural networks (NNs) can learn from a large amount of data, and realize function approximation, system identification, pattern classification, etc. Echo state networks (ESNs) are a new type of recurrent neural networks (RNNs) with a concise training process, which can avoid complex training algorithms, local minima and other problems of traditional neural networks. Adaptive dynamic programming (ADP) is a neural network based intelligent control method, which has strong ability of learning and optimization, and can perfectly solve the optimal control problem of complex nonlinear systems. Therefore, the two neural network based methods above are investigated in this paper to conduct analysis and optimization of building energy consumption.

The main works and contributions of this paper are concluded in the following aspects:

1. Based on the data of building energy consumption, an ESN based energy consumption prediction method for office buildings is developed to predict energy consumptions from sockets, lights and air-conditioners in office buildings. As regards the optimization of reservoir topology in ESNs, several simplified reservoir topologies are developed, and ESNs with the different reservoir topologies are applied to the prediction of office building energy consumption, where the performance of different topologies is compared. Experimental analysis shows that the developed reservoir topologies can achieve comparable prediction performance to traditional ESNs, and parameter sensitivity analysis on the topologies demonstrates that all the developed topologies present fairly strong robustness.

2. In accordance with the energy consumption models of rooms in office buildings, an ESN based room classification method for office buildings is developed to classify rooms into different categories. In the developed method, first the energy consumptions from sockets, lights and air-conditioners are modeled by three ESNs, respectively, and then based on the models, rooms are classified into office rooms, computer rooms, storage rooms and meeting rooms by a fourth ESN according to the energy consumption characteristics of different rooms. The convergence of the developed classification algorithm is proved. Experimental analysis demonstrates that the developed method can achieve high classification accuracies, and compared with several traditional classification algorithms, the developed method has better classification capability.

3. Given the results of room classification, energy storage equipments, such as storage batteries, are introduced into the rooms, and an ADP based energy consumption management method for office buildings is developed to obtain the optimal battery control strategies according to real-time electricity rate and energy demand, so as to control and optimize the energy consumption in different categories of rooms, and thereby achieving energy management of office buildings. It is illustrated by experimental analysis that the developed method can obtain optimal battery control strategies for different categories of rooms, thus reducing the total energy cost. In addition, the developed method can achieve better effects of energy saving than other optimization algorithms.

4. With solar energy introduced as the source of renewable energy, a Q-learning based renewable energy scheduling and battery control method for office buildings is developed to further optimize the energy consumption and reduce the energy cost. Two iterations are introduced in the developed method to approximate the optimal performance index function and optimal control in ADP algorithms, respectively. Experimental analysis focuses on offices in office buildings, and by implementing the developed method, the optimal renewable energy scheduling and battery control schemes of an office in different seasons are obtained, and the energy costs are further reduced. Moreover, the developed method can also save more energy compared with several other optimization algorithms.