基于RNN的发电机组排放预测及发电调度研究

发布时间：2019-04-08 12:11

【摘要】：近年来,对电力大数据进行挖掘、分析和应用已成为学术界和工业界的研究热点。其中发电机组排放预测问题是发电节能减排研究的重要部分。本文采用深度学习方法,提出一个新的发电机组排放预测模型,并以此为基础设计了一个具有节能减排效果的发电调度算法。目前研究者通过传统统计学方法、机器学习方法等手段在排放预测方面取得了一定的成果,但这些方法的预测效果较为依赖数据的特征质量。而深度学习方法具有表达能力强、不依赖高质量特征的优势。因此本文采用深度学习方法对发电机组排放预测问题进行研究,并通过对深度神经网络的研究与改进,提出一种精确的发电机组排放预测模型。随后,基于预测模型和智能优化算法,提出了一种可以有效节能减排的调度算法。具体而言,本文的研究工作主要有：(1)提出了一种基于递归神经网络(RNN)的发电机组排放预测模型。针对预测精度不足的问题,提出了两种深度神经网络激活函数的改进方案。并通过研究批规化(Batch Normalization)、抗过拟合、数据归一化、time steps选择等深度神经网络方法,提高了预测模型的精度和时间性能。最后与基于最小二乘法的多项式拟合、支持向量机回归(SVR)、隐马尔科夫模型(HMM)的方案进行对比,证明本文提出的基于深度学习的模型在发电机组排放预测问题上具有效性。(2)基于预测模型和蚁群智能优化算法,提出了一种能够有效节能减排的发电调度算法。针对蚁群算法易陷入局部最优解的问题,在转移规则、信息素更新、解集搜索等三个方面进行了改进,取得了更优的调度结果。通过在历史数据上的验证,本文提出的发电机组排放量预测模型和发电调度系统均达到研究目的,并比较传统方法取得显著的提高。
[Abstract]:In recent years, mining, analysis and application of electric power big data has become a hot spot in academia and industry. The problem of generating unit emission prediction is an important part of energy saving and emission reduction research. In this paper, a new model for generating unit emission prediction is proposed by means of in-depth learning method. Based on this model, an energy-saving and emission-reducing power generation scheduling algorithm is designed. At present, researchers have made some achievements in emission prediction by means of traditional statistical methods and machine learning methods. However, the prediction effects of these methods depend on the characteristic quality of the data. While the deep learning method has the advantage of strong expression ability and independent of high quality characteristics. Therefore, in this paper, the deep learning method is used to study the emission prediction of generating units, and through the research and improvement of the deep neural network, an accurate emission prediction model of generating units is proposed. Then, based on the prediction model and intelligent optimization algorithm, an efficient energy-saving and emission-reducing scheduling algorithm is proposed. Specifically, the main work of this paper is as follows: (1) A recursive neural network (RNN)-based emission prediction model for generating units is proposed. In order to solve the problem of insufficient prediction accuracy, two schemes for improving the activation function of deep neural network are proposed. The accuracy and time performance of the prediction model are improved by studying the batch (Batch Normalization), anti-overfitting and data normalization, time steps selection equal depth neural network method. Finally, the scheme of support vector machine regression (SVR), hidden Markov model (HMM) is compared with the polynomial fitting based on least square method. It is proved that the proposed model based on deep learning is effective in generating unit emission prediction. (2) based on the prediction model and ant colony intelligent optimization algorithm, an efficient energy-saving and emission-reducing generation scheduling algorithm is proposed. In order to solve the problem that ant colony algorithm is prone to fall into local optimal solution, the transfer rule, pheromone update and solution set search are improved, and better scheduling results are obtained. Through the verification of historical data, the emission prediction model and dispatching system of generator set proposed in this paper have achieved the purpose of research, and the traditional method has been improved significantly.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TM73

【相似文献】