基于神经网络风电场测风数据插补与长年代风速推算方法研究

发布时间：2018-03-28 22:24

本文选题：测风数据插补　切入点：长年代风速推算　出处：《武汉大学》2017年硕士论文

【摘要】：修建一个风电场首要的一步资源评估。评估的主要目的是获得一套能反映风电场运行寿命内其场址风资源水平的数据,以此作为在风电的建设可行性的重要参考。在实际工程中,现场的实测资料是评估的重要依据。然而,受恶劣天气、仪器故障等因素影响,测风设备的实测资料可能会出现缺测或数据失效等现象。因此,首先应找到适合的方法插补测风设备实测数据。因为测风资料存在年际变化,所以实测的短期数据很难反映出拟建风电场场址处的长年代风能资源水平。国内常用代表年法订正得到能体现风场多年平均风速分布及变化的测风序列。而"测量-相关-预测"(Measure Correlate Predict,MCP)方法则是国外通常采用的方法来推算测风长期序列。由此可见,测风数据的插补和长期序列推算是风能资源评估工作的基础,其精度将影响风电场后期的经济效益。本文针对国内风电场测风数据插补与订正方法的不足,基于神经网络提出了测风塔数据插补与长年代的测风序列推算的MCP方法,提高了测风数据的插补精度,消除测风序列在订正中产生的误差,取得了以下成果:以神经网络对多值输入及非线性关系拟合优势为出发点,通过算例比选,采用了 BP(Back-Propagation Network)和极限学习机(Exttreme Learning Machin,ELM)建立了神经网络方法的测风数据插补模型。比较BP与ELM插补模式和线性模式在的测风高度、地形条件、插补形式以及插补算法等因素不同时,对插补结果的影响,发现神经网络插补方法在各个因素下都较线性方法有优势。对于平原测风塔:同塔插补,BP或ELM较线性方法的均方根误差(Root Mean Square Error,RMSE)降低了约30%;异塔插补,BP或ELM的RMSE可减小4%。山区测风塔同塔插补,BP或ELM的RMSE约减小40%;异塔插补,BP或ELM在30m测风的RMSE可减小27.4%,在70m的可减小13.7%。采用遗传算法(Genetic Algorithm,GA)优化了的BP的初始权阈值。用GA-BP以及ELM极限学习机构建了 MCP方法中的关联模型。实现了以多参考站数据同时输入为基础长年代风速推算。将GA-BP与ELM算法用于风速的长年代推算时,其精度较线性方法有较大提高。计算结果表明:平原测风塔,BP较线性方法的RMSE减小了 3.6%,ELM算法RMSE减小了 4%,GA-BP的RMSE减小了 5.4%;山区测风塔,BP较线性方法RMSE减小了 7.9%,ELM的RMSE减小了 11.3%,GA-BP的 RMSE 减小了 12.8%。本文建立的基于神经网络的测风塔数据插补精度较高,在测风高度、插补方式、地形等因素影响下,误差都小于线性方法;基于神经网络的长年代推算模型实现了逐小时或逐月的长年代风速推算,为测风塔数据的处理提供了新方法。
[Abstract]:The primary step in the construction of a wind farm is resource assessment. The main purpose of the assessment is to obtain a set of data that can reflect the level of wind resources at the site during the operational life of a wind farm. It is an important reference for the feasibility of wind power construction. In the actual project, the field measured data is an important basis for evaluation. However, due to bad weather, instrument failure and other factors, Therefore, a suitable method should be found to interpolate the measured data of wind measuring equipment, because there is annual variation of wind data. Therefore, the measured short-term data can hardly reflect the long-term wind energy resource level at the site of the proposed wind farm. The domestic representative annual method is used to revise the wind series, which can reflect the annual average wind velocity distribution and variation of the wind field. The measure Correlate PredictMCPmethod is usually used in foreign countries to calculate the long-term wind series. The interpolation of wind data and the calculation of long term series are the basis of wind energy resource evaluation, and its precision will affect the economic benefit of wind farm in the later stage. This paper aims at the deficiency of interpolation and correction methods of wind data in domestic wind farm. Based on neural network, the MCP method of wind tower data interpolation and long term wind measurement sequence calculation is proposed, which improves the interpolation accuracy of wind measurement data and eliminates the errors produced in the correction of wind measurement data. The following results have been achieved: based on the advantages of neural network in multi-value input and nonlinear relation fitting, a numerical example is given. The wind data interpolation model of the neural network method is established by using BP(Back-Propagation Network and extreme learning Learning machine. The wind height, topographic condition, interpolation form and interpolation algorithm of BP and ELM interpolation mode are compared with each other. Effects on interpolation results, It is found that the neural network interpolation method is superior to the linear method in every factor. For the plain wind tower, the root Mean Square error of BP or ELM of the same tower interpolation is lower than that of the linear method, and the RMSE of BP or ELM is reduced. The RMSE of BP or ELM of wind measuring tower in mountain area can be reduced by about 40%, the RMSE of different tower interpolation BP or ELM can be reduced by 27.4% at 30 m and 13.7m at 70 m. The initial weight threshold of BP optimized by genetic algorithm is optimized by genetic algorithm (GA). The initial weight threshold of BP is optimized by GA-BP and ELM. The correlation model of MCP method is constructed by ELM extreme learning machine. Based on the simultaneous input of multi-reference station data, the long-age wind speed estimation is realized. When the GA-BP and ELM algorithms are used to calculate the long-age wind speed, The calculation results show that the RMSE of the wind tower in the plain is less than that of the linear method, and the RMSE of the GA-BP algorithm decreases by 5.40.The BP of the wind tower in the mountainous area is smaller than that of the linear method RMSE. The RMSE of 7. 9 / ELM reduces the RMSE of GA-BP by 11. 3 and decreases by 12. 8. The interpolation accuracy of wind tower data based on neural network is high. Under the influence of wind height, interpolation method, terrain and other factors, the error is smaller than linear method. A new method is provided for the processing of wind tower data.
【学位授予单位】：武汉大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TM614

【相似文献】