毕业生知识库

为了满足人类社会可持续发展对能源和环境的要求，光伏发电作为一种利用太阳能这种清洁能源的重要方式，受到了越来越广泛的关注。由于受天气状态、云层移动、安装条件、系统老化等因素的影响，光伏发电功率具有很强的随机性、波动性、昼夜周期性和季节周期性等特点。另一方面，随着光伏发电装机容量不断增加，其大规模并网给电力系统的安全稳定运行带来了严峻的挑战。精准可靠的光伏发电功率预测是应对光伏发电功率不确定性，促进光伏发电消纳，减少“弃光现象”的重要手段。近年来，随着传感器技术的发展，积累了大量的光伏发电功率和天气变量监测数据，为采用统计机器学习方法实现光伏发电功率预测提供了数据支撑。然而，如何从光伏发电功率和天气变量数据中充分挖掘所蕴含的信息仍存在诸多困难与挑战，亟需研究新的预测方法以更加精准地预测光伏发电功率。

本文针对光伏并网给电网安全稳定运行带来的挑战，从预测结果的表现形式入手，采用统计机器学习方法对光伏发电确定性预测和概率预测方法开展研究，实现光伏发电功率精准可靠预测。具体地，为了优化发电机组组合，提升光伏发电渗透率，本文首先对光伏发电功率日前预测方法进行了研究。此外，为了给电能质量分析、电力系统稳定性分析等提供更加全面的信息，本文还分别对两种光伏发电功率即时概率预测方法进行了研究。本文的创新性研究成果主要包括以下几个方面：

（1）针对相似日选择方法中由于待预测样本无法准确划分为某一类别而造成预测误差大的问题，提出了一种融合多种功率模式的光伏发电功率确定性预测方法，实现更加精准的光伏发电功率日前预测。首先，该方法采用了基于光伏发电功率数据的聚类分析以实现相似日选择，进而得到不同的光伏发电功率模式，避免了使用天气变量进行聚类分析时天气变量选择困难等问题；然后，采用岭回归方法自动确定待预测日在每种功率模式下预测值所占最终预测值的权重，避免了人工设计权重；最后在公开数据集（GEFCom14）上对所提方法进行了实验验证，实验结果表明所提方法具有更好的预测性能。

（2）针对使用上下限估计方法（LUBE）对光伏发电功率进行即时概率区间预测所存在的问题，提出了一种基于堆叠自动编码器的光伏发电功率概率区间预测方法。所提方法采用堆叠自动编码器对光伏发电功率影响因素进行特征提取，得到其紧致表达并以此作为LUBE方法的输入，有效降低了LUBE方法的参数搜索空间。此外，本文借鉴了支持向量机最大化间隔的思想，在LUBE方法目标函数中增加预测区间的均方误差项，使得到的预测区间中心更加靠近实际光伏发电功率值，进而提升预测区间的鲁棒性。在实际光伏发电功率数据集上的实验表明，所提方法得到的概率区间在可靠性和锐度两方面均表现出优异的性能。

（3）针对融合确定性预测方法与核密度估计方法对光伏发电功率进行即时概率预测时，确定性预测方法存在预测精度较低的问题，提出了一种基于循环神经网络和核密度估计的光伏发电功率概率预测方法。首先，所提方法利用带有时间注意力机制的循环神经网络对光伏发电功率进行确定性预测，以提升确定性预测精度。然后，利用核密度估计方法对所得到的预测误差进行拟合，并将其与确定性预测结果进行融合以生成概率预测。此外，根据光伏发电功率的特点，所提方法对光伏发电功率进行了工况划分并在每种工况下分别构建预测模型，进一步提升了概率预测性能。在实际光伏发电功率数据集上开展实验研究，验证了所提方法的有效性。

最后，设计和实现了光伏发电功率预测原型验证系统，并通过实验验证了所设计原型系统的有效性和可行性。

In order to meet the requirements of energy and environment for sustainable development of human society, photovoltaic (PV) power generation has become an important way to use clean and recyclable solar energy, which has attracted more and more attention. However, due to the influence of weather conditions, cloud movement, installation conditions, system aging and other factors, photovoltaic power generation has strong randomness, fluctuation, day-night periodicity and seasonal periodicity. In addition, with the increasing installed capacity, large-scale PV power grid connected to the power system has brought severe challenges to the safe and stable operation. Accurate and reliable prediction of PV power generation plays a significant role in dealing with the uncertainty of PV power generation, promoting the dissipation of PV power generation, and reducing the abandonment phenomenon. In recent years, benefiting from the development of sensor technology, accumulated monitoring data about PV power generation and weather variables provide data support for PV power prediction using statistical machine learning methods. However, it is still facing many problems and challenges on how to fully exploit the potential information from PV power and weather variables. Thus new prediction methods are in urgent need of research to forecast PV power more accurately.

Aiming at the challenges brought by grid-connected PV power generation to the safe and stable operation of the power system, this paper proceeds from the manifestation of prediction results, and studies on the deterministic and probabilistic forecasting methods of PV power generation using statistical machine learning methods. As a result, a more accurate and reliable prediction of PV power generation was obtained. Specifically, in order to optimize the combination of generating units and improve the penetration rate of PV power generation, this paper first studies the day-ahead deterministic prediction method of PV power generation. In addition, in order to provide more comprehensive information for power quality analysis and power system stability analysis, this paper also studies two real-time probability prediction methods of PV power generation. The innovative research results of this paper mainly include the following aspects:

(1) Aiming at the problem of large prediction error in the similar day selection method because the sample to be predicted cannot be accurately divided into a certain category, a deterministic prediction method for PV power generation combining multiple power models is proposed to achieve more accurate day-ahead PV power prediction. First, cluster analysis based on PV power data is used to realize similar day selection, and different PV power modes are obtained, which avoids the difficulty of selecting weather variables when using weather variables for clustering analysis. Then, the weight of the forecast value to the final forecast value in each power model is automatically determined by ridge regression method, avoiding the artificial design weight. Finally, the proposed method is validated on the PV power data set provided by the GEFCom14 (Global Energy Forecasting Competitions of 2014), and the experimental results show that the proposed method has the better prediction performance.

(2) Aiming at the problem of using the lower and upper bound estimation method (LUBE) to predict the real-time probability interval of PV power generation, a method based on a stacked autoencoder is proposed to predict the probability interval of PV power generation. The proposed method uses stacked autoencoder to extract features from factors influencing PV power, and obtain its compact expression which would be used as the input for the LUBE, thus effectively reducing the parameter search space of the LUBE. In addition, this paper uses the idea of support vector machine to maximize the interval, and add the mean square error term of prediction interval in the objective function of the LUBE method. After that, the center of the obtained prediction interval is closer to the actual PV power, with a result of improving the robustness of the prediction interval. The experiments on the actual PV power data set show that the probability interval obtained by the proposed method shows excellent performance in both reliability and sharpness.

(3) Aiming at the low prediction accuracy of the deterministic prediction method when combing it and the kernel density estimation method for the real-time probability prediction of PV power generation, a new probability prediction method based on recurrent neural network and kernel density estimation is proposed. First, the proposed method uses a recurrent neural network with a time attention mechanism to improve the accuracy of deterministic predictions of PV power generation. Then, a kernel density estimation method is used to fit the obtained prediction error which would be fused with results of the deterministic prediction to generate a probability prediction. In addition, the proposed method divides the operating conditions of PV power generation according to its characteristics, followed by separately constructing prediction models under each operating condition, thus further improving the performance of probability prediction. Experiments on the actual PV power data set verify the effectiveness of the proposed method.

Finally, the prototype verification system of photovoltaic power prediction is designed and implemented, and the effectiveness and feasibility of the prototype system are verified by experiments.