【作者】 吴涛；

【导师】 赖菲；

【作者基本信息】 西安热工研究院有限公司 ， 动力工程及工程热物理， 2019， 硕士

【摘要】 深度学习是近些年来的研究热点,在图像、音频和数据处理等方面均取得了良好的应用,本文将深度学习算法与光伏电站中的光伏功率预测和太阳能电池板缺陷检测结合展开研究。太阳能电池板作为光伏电站将太阳能转化为电能过程的核心设备,其质量好坏将直接影响太阳能转化效率和自身的可用寿命,对保障光伏电站安全稳定运行及收益的最大化也至关重要。因此对太阳能电池板进行缺陷检测具有十分重要的现实意义。本文提出首先将电站光伏阵列分成若干区域范围,基于深度学习算法对每个区域的光伏功率做预测,将该时刻预测值与对应时刻的实时光伏功率对比,如果偏差超出了设定误差范围,说明此范围内的太阳能电池板可能出现缺陷,然后在该区域范围内采集电池板电致发光(EL)图像,利用在深度学习领域中图像分类性能良好的卷积神经网络对其进行自动识别判断,找出有缺陷的电池板,以及确定缺陷的类型,对该缺陷电池进行相应处理以保障电站安全高效持续运行。本文的主要研究内容如下:(1)对获取的太阳能光伏功率及天气信息原始数据的不足进行了分析,利用箱型图确定数据中存在的缺失值和异常值的位置,基于KNN算法对缺失值和异常值进行了填充和替换,利用归一化对其进行数据预处理,建立太阳能光伏功率预测数据集。(2)对Dropout机制和Sigmoid、tanh、Re LU、ELU、Softplus五种激活函数数学原理做了分析,利用实验验证比较了各自在深度学习模型不同层和Dropout值不同大小对光伏功率预测结果精确度的影响,并指出了在实际应用中容易发生的问题。发现将Dropout设置在循环层和输出层可以提升网络预测准确率,隐含层放置Sigmoid函数或在输出层放置ELU函数时效果最好,各项误差最小,Re LU函数使得网络在训练时容易出现神经“坏死”的现象。改变门控循环单元网络的网络层数及节点数,探讨了网络容量对预测结果精确度和训练耗费时间的影响。经过反复实验发现:增加网络的深度要比增加单层节点数的收益更大。通过比较循环神经网络、长短期记忆网络、双向长短期记忆网络和门控循环单元在太阳能光伏功率预测方面的综合性能,发现当数据量庞大且精度要求不是很高的情况下,门控循环单元网络是最佳选择,耗时短的同时准确率较高且稳定,当数据量较小但精度要求较高时,双向长短期记忆网络更加适合。(3)将经典的卷积神经网络(Le Net-5)应用于太阳能电池板缺陷检测,发现分类效果不佳后,利用Tensorboard的可视化对其网络结构和超参数进行改进,对比分析了经典Le Net-5模型和改进后的Le Net-5以及支持向量机在太阳能电池板缺陷检测上应用的效果。(4)对本文主要研究内容进行了总结和概括,对本文研究工作的的不足做了分析,对后续的发展方向和思路进行了展望。更多还原

【Abstract】 Deep learning is a research hotspot in recent years,and has been well applied in image,audio and data processing.This paper combines the deep learning algorithm with photovoltaic power forecasting and solar panel defect detection in photovoltaic power plant.The quality of solar panels which as the core equipment in the process of converting solar energy into electricity in photovoltaic power station will directly affect the solar energy conversion efficiency and their own service life,which is also crucial to ensure the safe,stable operation and maximize the benefits of photovoltaic power station.Therefore,it is of great practical significance to detect the defects on solar panels.This paper proposes to divide the power plant array into several ranges,and predict the photovoltaic power of each region based on the deep learning algorithm.Compare the predicted values with the real-time photovoltaic power at the corresponding time.If the deviation exceeds the setting error range,the solar panels in this range may be defective,and then the panel electroluminescence(EL)images are collected in the region,which are automatically identified and classified by using a convolutional neural network with good image classification performance in the depth learning field.Identify the defective panels and determine the type of defects,then handle the defective battery to ensure safe,efficient and continuous operation of the plant.The research contents of this paper are as follows:(1)The shortages of raw data of solar photovoltaic power and weather information are analyzed,the box plot is used to determine the location of missing values and outliers in the data,and use KNN algorithm to fill and replace missing values and outliers.Use data normalization to process the new data,and establish a solar photovoltaic power prediction data set.(2)The Dropout mechanism and mathematical principles of five activation functions,such as Sigmoid,tanh,Re LU,ELU and Softplus are analyzed.The effects of different Dropout values in different layers on the accuracy of photovoltaic power prediction results are compared by experiments,and the problems likely to occur in practical applications are pointed out.It is found that the Dropout setting in the loop layer and output layer could improve the accuracy of network prediction.The Sigmoid function in the hidden layer or the ELU function in the output layer could improve the network prediction accuracy.The Re LU function makes the network prone to neuron ’necrosis’ during model training.By changing the numbers of network layers and nodes in the GRU network,the effects of network capacity on the accuracy of prediction results and the time consumption of model training are discussed.After repeated experiments,it is found that increasing the depth of the network is more profitable than only increasing the number of nodes in every layer.After comparing the comprehensive performance of Simple RNN,LSTM,Bi LSTM and GRU in solar photovoltaic power prediction,it is found that when the data volume is very large and the precision requirement is not very strict,the GRU is the best choice for its prediction results are accurate,stable and fast-convergence.When the amount of data is relatively small but the precision requirement is strict,Bi LSTM is more suitable.(3)The classical convolutional neural network(Le Net-5)is applied to the defect detection of solar panels,then get unsatisfactory performance.Therefore,by using the visualization of the Tensorboard,the network structure and super parameters are improved,and the application effects of the classical Le Net-5 model,the improved Le Net-5 model and the support vector machine in the defect detection of solar panels are compared and analyzed.(4)The main research contents of this paper are summarized.The shortcomings of the main research work are analyzed,and the future development direction and ideas are prospected.更多还原