基于云计算和智能算法的短期风电功率预测方法研究

发布时间：2019-03-21 09:47

【摘要】：随着化石能源的全面紧张和环境污染的日益加剧,世界各国普遍把开发、利用可再生能源作为其重要的能源发展战略。而在可再生能源中,风能是发展最快的清洁能源,风力发电方式也最具规模开发需求和商业发展前景。就目前情况来看,风力发电设备与技术已相对成熟,但是由于风力发电过程中的随机性、波动性和间歇性,造成风电输出功率稳定性较差,因此限电弃风问题严重制约风电并网。目前存在的主要问题是如何提高风电功率预测的精度,特别是未来24小时的预测精度。基于以上背景,本文主要就以下几个方面展开研究。(1)风力发电系统历史数据的真实可靠是进行风电功率预测的基础,而在风力发电系统运行或数据采集、测量、传输、转换等环节,尤其是人为的限电弃风,历史数据中不可避免的存在异常数据。本文在分析风电场异常数据特征的基础上,采用四分位数法对风电系统弃风数据进行前期预处理,提高历史数据的准确率。(2)相比于其他智能预测算法,人工神经网络在自学习性、自适应性、鲁棒性、容错性和推广能力方面性能表现突出。当前风电功率预测中使用较多的人工神经网络是静态神经网络,而利用静态型神经网络进行预测,造成风电功率序列丧失时变特性能力,因此预测精度不高。所以,本文选用能够更好反映风电功率动态特征的Elman神经网络,给出Elman神经网络风电功率预测算法。(3)Elman神经网络所采用的网络参数会影响网络性能,而当前在神经网络学习阶段,普遍采用固定梯度变化方向的梯度下降方法,采用这种方法会存在收敛速度慢、易陷入局部最优解等缺陷,这些缺陷限制了网络的寻优能力。因此,本文采用具有全局寻优性能的改进布谷鸟搜索算法优化Elman神经网络的权值和阈值,目的在于提高Elman神经网络的稳定性和泛化能力。(4)针对传统单机计算资源和存储资源不能很好的满足短期风电功率预测的实际需求,本文将改进型布谷鸟搜索算法和Elman神经网络进行并行化设计,在Spark云平台上对算法进行性能测试。实验分析表明,预测精度和实时性均优于传统单机功率预测算法。
[Abstract]:With the development of fossil energy and the increasing environmental pollution, renewable energy is regarded as an important strategy of energy development in the world. In renewable energy, wind energy is the fastest developing clean energy, and wind power generation has the most large-scale development demand and commercial development prospects. As far as the current situation is concerned, wind power generation equipment and technology have been relatively mature, but due to the randomness, volatility and intermittency of wind power generation process, the stability of wind power output is poor. As a result, the problem of limiting the wind power abandonment seriously restricts the wind power grid. At present, the main problem is how to improve the accuracy of wind power prediction, especially in the next 24 hours. Based on the above background, this paper mainly studies the following aspects: (1) the real and reliable historical data of wind power generation system is the basis of wind power prediction, and in the operation of wind power system or data acquisition, measurement, transmission, Conversion and other links, especially artificial power restriction wind, historical data inevitably exist abnormal data. On the basis of analyzing the characteristics of abnormal wind farm data, this paper uses the quartile method to pre-process the abandoned wind data of wind power system in order to improve the accuracy of historical data. (2) compared with other intelligent prediction algorithms, The performance of artificial neural network is outstanding in the aspects of self-learning, adaptability, robustness, fault tolerance and generalization ability. At present, more artificial neural networks are used in wind power prediction, but static neural network is used to forecast wind power series, which results in the loss of time-varying capability of wind power series, so the prediction accuracy is not high. Therefore, the Elman neural network which can better reflect the dynamic characteristics of wind power is chosen in this paper, and the wind power prediction algorithm based on Elman neural network is given. (3) the network parameters used by Elman neural network will affect the performance of the network. At present, in the learning phase of neural networks, the gradient descent method with fixed gradient change direction is generally used. This method will have some defects such as slow convergence rate, easy to fall into local optimal solution and so on, which limit the optimization ability of the network. Therefore, the improved cuckoo search algorithm with global optimization performance is used to optimize the weights and thresholds of the Elman neural network. The purpose of this paper is to improve the stability and generalization ability of Elman neural network. (4) the traditional single-machine computing resources and storage resources can not meet the actual demand of short-term wind power prediction. In this paper, the improved cuckoo search algorithm and Elman neural network are designed in parallel, and the performance of the algorithm is tested on the Spark cloud platform. Experimental analysis shows that the prediction accuracy and real-time performance are better than the traditional single-machine power prediction algorithm.
【学位授予单位】：华北电力大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;TM614

【参考文献】