【作者】 林星春；

【导师】 曹家枞；

【作者基本信息】 东华大学 ， 供热、供燃气、通风与空调工程， 2007， 硕士

【摘要】 随着“能源危机”的临近，今天世界各国都在竭尽全力地进行着节省能源和开发新能源的研究工作。太阳能被看成巨大而清洁的具有代表性的新能源，随着太阳能利用技术的进步，需要的太阳辐射资料很多，精度要求又很高，并且对未来时间的太阳辐射强度预测技术提出了更高的要求。太阳辐射预测可弥补我国辐射资料的不足，并为太阳能工程、新能源的利用提供重要参考工具；同时，一个合理的太阳辐射模型，是获得空调负荷精确预测的关键性前提。本文拟就太阳辐射的预测从理论上和方法上进行深入的探索与研究。由于太阳辐射具有高度非线性的特点，而神经网络处理非线性问题有其特殊的能力，因此本文利用神经网络的非线性函数逼近及自学习、自适应的特性，将神经网络预测方法作为主要的研究方法。考虑到神经网络激励函数Sigmoid函数的单调性，本文采用具有动态性的Morlet小波函数取代隐含层Sigmoid函数，结合了小波变换良好的时频局域化性质及神经网络的自学习功能，并且，引入了两个新的参变量，即伸缩因子和平移因子，使得网络具有更多的自由度，从而使其具有更灵活有效的函数逼近能力。针对于太阳辐射数据序列的非线性和瞬变性特点，本文在BP网络结构的基础上，在隐含层中增加一个反馈层，作为一步延时算子，达到记忆的目的，从而使系统具有适应时变特性的能力，这种自联方式使其对历史状态的数据具有敏感性，内部反馈网络的加入增加了网络本身处理动态信息的能力，能直接反映动态过程系统的特性，从而达到了动态建模的目的。从而，本文建立了太阳辐射的对角递归小波BP神经网络预测模型。考虑到一般神经网络存在的收敛速度慢，易陷入局部极小值的弊端，本文从网络的训练算法以及权阈值修改方案等方面提出了改进：对网络训练算法的各种改进进行综合比较，选择带动量的自适应变步长学习法训练网络，有效解决训练时间长，易陷入局部最优解的问题：将初始参数的确定与样本信息结合起来，避免随机赋值带来的不确定性，针对于训练样本的周期性，结合批量训练模式提出了批量平均权值法，保证每次网络训练都能得到最优初始值，训练获得最佳结果。另外，考虑到网络训练样本信息的重要性，对影响太阳辐射的因素进行细致分析，精简非独立影响因素，确定主要影响因素：一方面，对网络主要建模参数太阳辐射历史数据进行相关性分析，分析其内在的时间序列变化规律，选择与预测曰或预测时刻最相关的历史数据作为输入信息：另一方面，考虑到太阳辐射强度变化是具有动态性、时变性、多扰量性、不确定性等随机特性的典型非线性曲线，其中云量、大气状况、天气状况等随机性因素的影响至关重要。本文将关于天气阴晴状况的气象预报模糊化并修正后作为神经网络预测模型的重要输入信息进行训练，这样，网络能够充分地学习到太阳辐射和天气状况之间的关系，使得仿真预测结果能够主动地反映天气变化带来的显著影响，在很大程度上提高网络预测精度。基于以上研究方法，本文完成了太阳辐射复合预测的全套模型，包括太阳曰总辐射预测模型、太阳逐时总辐射预测模型、太阳日散射辐射预测模型、太阳逐时散射辐射预测模型。本文中分别以澳门1991～2000年和上海2001～2002年的相关数据作为预测模型的对象，对预测模型的网络性能参数、回溯预测和仿真预测的相对误差、绝对误差和均方根误差等进行了实例分析，为了更进一步验证预测模型的精度，本文将四种太阳辐射对角递归小波BP神经网络预测模型与几种典型的传统模型和普通神经网络模型进行了预测(计算)比较，其均方根误差和平均相对误差都有大幅度改善，决定系数都在0.93以上。更多还原

【Abstract】 With the increasing problem of world "Energy Crisis", today, every country all over the world are emphasizing on the research work of energy conservation and new energy development, teeth and nail. Solar energy is a representative new energy which is abundant and clean. As the technology of utilization of solar energy develops, the need of solar irradiance data are increasing, the demand of precious are heightening, and the technique of forecast solar irradiance intensity in future are requiring further improving. Solar irradiance forecast can fill in the irradiance database and provide an important tool for solar energy application and new energy utilization. Meanwhile, an appropriate solar irradiance model is a key precondition of accurate load forecast for air conditioning systems. This paper aims to the deep research into the theories and methods of solar irradiance forecast.Solar irradiance is of non-linear characteristic. Neural networks are skilled in handling non-linear problems. Because of the neural networks’ capability of non-linear function approximation and self-study, self-adapt, it becomes the main research approach of this paper. Considering the monotony of Sigmoid function, which is the activation function of neural network, this paper replaced the Sigmoid function with Morlet wavelet function, namely, wavelet neural network. Wavelet neural network combines the prominent ability of wavelet functions in time-frequency domains multi-resolution analyses and the good performance of neural networks in approximating nonlinear functions. Two new variables were introduced in Wavelet neural network, viz. dilation factor and translation factor, which entitles the network more degree of freedom, then the network, can approximate the function more appropriately and effectively. Considering the characteristic of non-linear and time-varying of solar irradiance, a feedback layer was introduced in the hidden layer of back-propagation network. The feedback layer functions, as a one-step delay operator for memorizing, which, accordingly, entitles system the function of adapting time-varying characteristic. This self-connection method enables the network to have sensitivity to historical data, and inner feedback network enhances the network capability of handling dynamic information, which can describe the characteristic of dynamic system directly and achieve the target of dynamic modeling. According to the analysis above, a diagonal recurrent wavelet back propagation neural network model for solar irradiance forecast was built.Considering the advantages of neural networks, slow convergence and local minimum trend, the improvement of training algorithm and the updating method of the weights and biases of network were proposed: (1). Compared existing modified training algorithms and chose the adaptive variable-step back-propagation algorithm that has momentum terms to train network, which overcame the problems of slow convergence and local minimum trend effectively. (2). Contacted the selection of initial values with learning data sample, with the attempt to avoid the uncertainty which brought by random initializing. According to the periodicity of learning data sample, combining with batch training mode, a method of batch-average-weight was put forward. The batch-average-weight method can ensure the optimal initial values and excellent training result every time.Besides, considering the importance of the training sample data of the network, the influence factors of the solar irradiance should be analyzed carefully to eliminate dependent factors and determine independent factors. On one hand, implemented correlation analysis to historical solar irradiance data, the main modeling parameter, analyzed the inner changing rule of the time sequence and choose the most correlative historical data of the forecasting day or forecasting hour as the input information of the network. On the other hand, the fluctuation of solar irradiance intensity is a typical non-linear curve which is dynamic, time-varying and multi- interfering. Among the influence factor of solar irradiance, the stochastic factors, such as cloud coverage and atmosphere condition are very important. This paper introduced the weather forecast information about cloud to the neural network model. The weather forecast information were fuzzificated and modified as the important input of the network, thus, the network could learn the connection between solar irradiance and air condition adequately, the result of emulating forecast could reflect the sudden influence caused by weather change actively. In a word, the proposed model achieves higher forecasting accuracy.According to the analysis above, the achievements of this paper was a complete set of models of forecasting solar irradiance, including daily total solar irradiance forecast model, hourly total solar irradiance forecast model, daily diffuse solar irradiance forecast model, hourly diffuse solar irradiance forecast model. This paper took the correlative data of Macao from 1991 to 2000 and Shanghai from 2001 to 2002 as the investigate object of the forecast models. The performance parameters, relative errors and absolute errors of tracing forecast and emulating forecast were analyzed according to the example. In order to further assess the goodness of the forecast model, several traditional models are used to calculate or predict solar irradiance for comparing with the four diagonal recurrent wavelet back propagation neural network forecast models, the results demonstrate that the root mean square errors and the mean relative errors are improved significantly, and the coefficient of determinations are higher than 0.93.更多还原