基于WRF的DBN风速预测与并行优化研究

发布时间：2018-10-09 10:35

【摘要】：高精度的风速预测对风电发展具有重要意义。目前针对风速预测研究大多集中于两方面:第一,采用中尺度数值模式进行风速预报,但其因计算量大具有计算时长和硬件设备的局限性,并且单一中尺度模式预报效果已不能满足当前预测精度需求,需要进一步引入学习模型对其预报风速进行订正和预测;第二,采用学习模型进行风速预测,现采用的学习方法大多为浅层机器学习方法,学习能力有限,预测精度还有待进一步提高,而深度学习模型具有更深层次的学习能力,能更好地描述目标物体。鉴于此,为了对中尺度WRF(Weather Research and Forecast,天气研究与预报)模式进行并行优化和提高风速预测准确率,本文主要做了以下三个工作:(1)WRF模式风速预报效果评估:首先采用修正海拔高度后的地形资料对WRF模式进行风力模拟,得出修正数据在山地等复杂地形下对WRF模式具有一定影响的结果。通过获取的简单地形下某70m高测风塔的实际测风数据,对WRF模式重新计算后输出的气压、气温、风速、风向四个气象要素预报结果进行了全面检测和分析,结果表明,WRF模式预报与实测数据具有一定的整体相关性且符合建立DBN(Deep Belief Nets,深度信念网络)风速预测模型的需求。(2)WRF模式的并行优化:分别搭建了多机多核集群和小型塔式工作站两种并行计算平台,采用三种并行方式对WRF模式进行模拟计算,获得了较优的并行效能;并利用并行效率、加速比、价格等参数分别对并行方式和计算平台进行对比分析,以便用户能根据计算需要合理选择更高效的并行方式和计算平台。(3)构建基于WRF的DBN风速预测模型:为提高WRF模式风速预报的准确率,引入深度学习模型DBN,该模型具有先无监督后有监督学习的优点,通过将WRF模式风速预报结果与实测数据作为输入对深度信念网络进行逐层训练,构建了基于WRF的DBN风速预测模型,并进行仿真实验。以上实验验证了本文对WRF模式并行优化的有效性和适用性,以及通过相关对比实验证明了本文构建的风速预测模型具有更深层次的学习能力,获得了更高的预测精度且更具应用性。
[Abstract]:High precision wind speed prediction is of great significance to wind power development. At present, the researches on wind speed prediction are mainly focused on two aspects: first, the mesoscale numerical model is used for wind speed prediction, but due to the large amount of calculation, it has the limitations of long calculation time and hardware equipment. And the prediction effect of single mesoscale model can no longer meet the demand of current forecast precision, so it is necessary to further introduce learning model to revise and forecast the predicted wind speed. Secondly, the learning model is used to predict the wind speed. Most of the current learning methods are shallow machine learning, with limited learning ability, and the prediction accuracy needs to be further improved, while the depth learning model has deeper learning ability and can better describe the target object. In view of this, in order to optimize the mesoscale WRF (Weather Research and Forecast, weather research and forecast model in parallel and improve the accuracy of wind speed prediction, The main work of this paper is as follows: (1) the wind speed prediction effect of WRF model is evaluated. Firstly, the wind model of WRF is simulated with the topographic data of modified altitude. The results show that the modified data have a certain influence on the WRF model under the complex terrain such as mountainous area. Based on the actual wind data obtained from a 70 m high wind tower under a simple terrain, the forecasted results of the four meteorological elements such as air pressure, air temperature, wind speed and wind direction after recalculating the WRF model are comprehensively detected and analyzed. The results show that the prediction of WRFmodel and the measured data have a certain overall correlation and accord with the demand of establishing DBN (Deep Belief Nets, depth belief network) wind speed prediction model. (2) parallel optimization of WRF model: the multi-machine multi-core cluster and the small cluster are built respectively. Two parallel computing platforms for tower workstations, Three parallel methods are used to simulate and calculate the WRF mode, and the parallel efficiency, speedup ratio, price and other parameters are used to compare and analyze the parallel mode and the computing platform, respectively, and the parallel efficiency, speedup ratio, price and other parameters are used to compare and analyze the parallel mode and the computing platform. So that users can reasonably select a more efficient parallel mode and computing platform according to the needs of computing. (3) build the DBN wind speed prediction model based on WRF: to improve the accuracy of WRF model wind speed prediction, The depth learning model (DBN,) is introduced. The model has the advantages of unsupervised learning and supervised learning. The depth belief network is trained layer by using the wind speed prediction results of WRF model and the measured data as input. A DBN wind speed prediction model based on WRF is constructed and simulated. The above experiments verify the validity and applicability of this paper to the parallel optimization of WRF model, and prove that the wind speed prediction model constructed in this paper has deeper learning ability. Higher prediction accuracy and more application are obtained.
【学位授予单位】：新疆大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM614

【参考文献】