为应对化石能源耗尽与温室效应两大危机，近年来，世界风能与太阳能发电装机及并网容量都在快速提升，当大规模风光发电功率并入电网的同时，其随机波动性会给电网发电计划的制定带来新的问题和挑战。其中，研究高效、高精度的风光发电功率预测技术对缓解电网调峰调频压力以及提高电网对风光功率的消纳能力有着重要的现实意义。本文以风光发电功率为研究对象，通过综合利用深度学习神经网络技术、优化理论并结合所研究领域的专业知识，对基于深度学习神经网络的风光发电功率预测技术进行了细致的研究和探讨。主要研究成果如下：针对风光功率预测数据中的异常情况，实现了一种基于孤立森林算法的异常值检测方法。此外，针对光伏功率预测因子的选择过程，提出并实现了一种基于詹森-香农散度(Jensen-Shannon Divergence)的预测因子筛选方法。针对风功率确定性预测方式，采用卷积神经网络(Convolutional Neural Network, CNN)进行功率预测，提出了一种功率预测因子的组织方式，并对预测误差的经验分布进行简要分析。此外，与高斯过程回归预测、支持向量机预测及决策树预测的结果进行对比。针对光伏功率的确定性预测方式，提出了基于自适应进化规划算法的长效短时记忆(Long Short Time Memory, LSTM)神经网络参数寻优方法以降低预测误差。其中，对LSTM的离散网络参数进行了优化，在优化过程中还考虑了不同种类神经网络类型的综合网络设计。通过利用实际光伏发电功率数据对所提出的自适应进化规划参数优化方法进行了有效性验证，并进一步分析了神经网络层数、神经元数目、LSTM数量以及预测因子维度对光伏发电功率预测结果的影响。针对光伏功率的概率预测方式，提出了基于梯度训练算法的上下限估计(Lower Upper Bound Estimation, LUBE)区间预测方法。其中，以内点障碍函数优化方法对LUBE方法进行改进，分析了所提出梯度方法存在的第一类与第二类误差，结合实际光伏发电功率数据对所提出的基于梯度的LUBE区间预测方法进行了有效性验证。此外，还提出了一种基于预测区间边界均方误差的预训练方法，对内点障碍函数优化方法中的初始可行解进行搜索，同时分析了多类梯度算法在所提出的区间预测方法中的应用效果。
In recent years, owing to the exhaustion of fossil fuels and greenhouse effect, the installed and integrated capacities of wind power and solar power have increased rapidly. Meanwhile, the large-scale grid connected wind and solar power uncertainties have brought new problems and challenges for setting out power generation plans. Hence, it is worth to research on highly efficient and highly accurate wind/photovoltaic(PV) power forecasting methods, which are important to alleviate the pressure of peak load and frequency regulations and to enhance the capability wind/PV power accommodation. Basing on the wind/PV power generation, the thesis aims to study and discuss wind power and PV generation forecasting techniques based on deep learning neural networks through appling deep learning neural network technology, optimization theory and the expertise of the research field. The major achievements are as follows:The isolated forest outlier detection algorithm is implemented to detect the abnormal data in the wind power forecasting data. In addition, a feature selection method based on Jensen-Shannon Divergence is proposed and applied for the selection process of PV power predictors.For the deterministic wind power forecasting technique, convolutional neural network(CNN) is used for wind power forecasting, and an arrangement method of predictors is proposed for CNN. The empirical distribution of forecasting errors is briefly analyzed. In addition, Gaussian process regression, support vector machine and decision tree are used as benchmark forecasting techniques to conduct the comparative analysis.For the deterministic PV power forecasting technique, based on the long short-term memory(LSTM) neural network, a self-adaptive evolutionary programming algorithm is applied to optimize the discrete network parameters of LSTM in order to reduce the forecasting error, and the integrated network design of different neural network types is considered during optimization. In addition, the proposed method is verified by the actual PV power generation data, and the effects of neural network layer, number of neurons, LSTM layer, and dimension of forecasting predictors on forecasting accuracy are analyzed elaborately.For the probabilistic PV power forecasting technique, a lower upper bound estimation(LUBE) interval forecasting method based on gradient training algorithm is proposed. Meanwhile, the LUBE method is enhanced by the interior-point optimization method. The type I and type II errors existed in the proposed gradient method are analyzed, and the proposed method is verified by the actual PV power data. In addition, a pre-training method based on prediction interval boundary mean square error is proposed to search the initial feasible solution pertinent to the interior-point optimization method, and the results of applying multiple gradient algorithms are compared in the proposed interval forecasting method.
