基于混沌时间序列的短期风电功率组合预测方法研究

发布时间：2018-03-17 09:37

本文选题：混沌理论　切入点：经验模态分解　出处：《湖南大学》2016年硕士论文　论文类型：学位论文

【摘要】：风电功率具有波动性、间歇性和随机性等特点,大规模风电的接入严重影响电网运行的安全性、稳定性和经济性。实践证明,高精度的风电功率预测能够有力缓解大规模风电并网产生的不利影响,尤其是短期风电功率的精确预测,对于电力调度部门安排调度计划,保证电能质量和电力系统安全经济运行有重要作用。首先,选择混沌理论来挖掘风电功率时间序列的特性,并利用相空间重构理论对其进行分析。将重构相空间相量作为风电功率时间序列预测模型的训练样本,为后面建立预测模型以实现对风电功率混沌系统相空间运动轨迹预测提供支持。然后,提出了一种基于混沌时间序列的改进型BP神经网络。该模型利用综合法对基本BP神经网络的不足进行改进：在最速下降法中同时引入自适应学习速率和动量项对其本身进行改进；利用遗传算法从全局开始搜索,找到权值系数最优解所在区域的某一解；再利用改进的最速下降法去调整权值系数,经过少量样本训练的训练,就能够得到权值系数的最优解。此外,将模型的输入维数取为相空间最小嵌入维数,从而减少了建模的盲目性。实例仿真表明,本文提出的改进型BP神经网络收敛速度快,预测精度高,不易陷入局部极小值点。其次,提出了一种基于混沌时间序列的自适应遗传算法-Volterra神经网络模型。该预测模型是基于自适应-Volterra级数和三层BP神经网络的等价性建立的,其截断项数取为重构相空间的最小嵌入维数,大大提高了预测模型的自适应性。自适应遗传算法-Volterra神经网络模型将自适应-Volterra级数精准建模能力、改进的BP算法训练学习能力、遗传算法全局搜索能力结合起来,实现了对风电功率混沌系统相空间运动轨迹的精确预测,应用于实际短期风电功率预测中,预测精度显著提高。最后,采用经验模态分解降低风电功率数据的非平稳性,分别建立了经验模态分解-遗传算法-Volterra神经网络和经验模态分解-改进型BP神经网络模型两种组合预测模型。将两种组合预测模型应用于实际短期风电功率预测的仿真实验中,结果表明,利用经验模态分解可以有效降低风电功率数据的非平稳性,更好地挖掘数据内部蕴含的规律,与单一预测方法相比,预测性能进一步提升。结果同时表明,本文提出的经验模态分解-遗传算法-Volterra神经网络组合预测模型,精度高,性能优越,应用于短期风电功率混沌时间序列的预测中可取得较好效果。
[Abstract]:Wind power has the characteristics of volatility, intermittence and randomness. The access of large-scale wind power seriously affects the security, stability and economy of power grid operation. High-precision wind power prediction can effectively mitigate the adverse effects of large-scale wind power grid connection, especially the accurate short-term wind power prediction. It is important to ensure the power quality and the safe and economical operation of the power system. Firstly, the chaotic theory is chosen to excavate the characteristics of the wind power time series. The phase space phasor is used as the training sample of wind power time series prediction model. A prediction model is established to predict the trajectory of phase space motion in wind power chaotic system. Then, In this paper, an improved BP neural network based on chaotic time series is proposed. The model uses synthesis method to improve the shortcomings of basic BP neural network: adaptive learning rate and momentum pair are introduced into the steepest descent method at the same time. Improve itself; The genetic algorithm is used to search the whole world to find a certain solution in the region where the optimal solution of the weight coefficient is located, and then to adjust the weight coefficient by using the improved steepest descent method, which is trained by a small number of samples. In addition, the input dimension of the model is taken as the minimum embedded dimension in phase space, thus reducing the blindness of modeling. The simulation results show that the proposed improved BP neural network converges quickly. The prediction accuracy is high and it is not easy to fall into the local minimum. Secondly, An adaptive genetic algorithm-Volterra neural network model based on chaotic time series is proposed, which is based on the equivalence of adaptive Volterra series and three-layer BP neural network. The truncation term is taken as the minimum embedded dimension of the reconstructed phase space, which greatly improves the self-adaptability of the prediction model. The adaptive genetic algorithm -Volterra neural network model can accurately model the adaptive -Volterra series, and the improved BP algorithm can train the learning ability. Combining the global searching ability of genetic algorithm, the accurate prediction of phase space motion trajectory of wind power chaotic system is realized, which is applied to the actual short-term wind power prediction, and the prediction accuracy is improved significantly. The empirical mode decomposition (EMD) is used to reduce the nonstationarity of wind power data. Two kinds of combined forecasting models, empirical mode decomposition-genetic algorithm-Volterra neural network and empirical mode decomposition-improved BP neural network, are established respectively. The two combined forecasting models are applied to the simulation experiment of actual short-term wind power prediction. The results show that the empirical mode decomposition can effectively reduce the non-stationarity of wind power data, and better mine the laws contained in the data. Compared with the single prediction method, the prediction can be further enhanced. The combined prediction model of empirical mode decomposition-genetic algorithm-Volterra neural network presented in this paper has high precision and superior performance. It can be applied to the prediction of short-term wind power chaotic time series with good results.
【学位授予单位】：湖南大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TM614

【相似文献】