基于地基云图和深度学习的光伏功率超短期预测研究

Large scale photovoltaic (PV) grid connecting is an effective way to sufficiently use the solar energy. However, PV power is intermittent and fluctuated. Accurate prediction of PV power output will directly affect the stability of the grid, the quality of power output and economic benefits for power stations as well. Among the various PV power forecasting methods, the method based on ground-based cloud images is able to integrate the information of real-time cloud images, which makes it more applicable than the method based on numerical data, especially in super short-term PV power forecast within 4 hours. However, the performance of present forecasting models using ground-based cloud images depends on the quality of manually extracted features, which is usually not satisfactory. Deep learning is capable of extracting the features automatically under the super powerful computing platform, GPU, which is potentially a new powerful method to improve the quality of extracted features. In this project, a model based on Time Series Analysis will be constructed firstly for the super short-term irradiance forecast using the sufficiently available ground-based cloud images and solar irradiance data, which will be used as the base model of PV power forecasting. The ground-based cloud images will then be used to further improve the performance of the base model. The convolutional neural network (CNN) will be applied to automatically extract features from ground-based cloud images, for the purpose of detecting and classifying clouds. Furthermore, the recycled neural network (RNN) will be employed to predict super short-term PV power with historical solar radiation data. Previous work of our research group has basically proved that the forecasting model based on deep learning is potentially to achieve better forecast than traditional methods.

大规模光伏并网发电是充分利用太阳能的有效方式，但光伏出力具有随机性和波动性。能否实现光伏电站输出功率的准确预测，将直接影响电网安全、电能质量和电站经济效益。相比基于数值预报数据的预测方法，基于地基云图的预测方法由于引入了实时云图信息，更加适合0-4小时超短期预测。但是现有的基于地基云图的预测方法效果依赖于人为提取的图像特征，预测效果不能令人满意。深度学习能利用GPU超算平台的强大计算能力实现自动特征提取，为解决上述问题提供了新思路。本课题计划在充分采集地基云图和太阳辐照度数据的基础上，先建立基于时间序列的辐照度超短期预测模型，作为光伏预测的基本模型，然后利用后续的云图处理结果进一步优化预测结果。拟采用卷积神经网络（CNN）自动提取云图特征并进行云图的识别与分类，采用循环神经网络（RNN）实现辐照度超短期预测。初步测试表明基于深度学习的预测方法确实有可能取得比传统方法更好的预测结果。

为了达到碳中和，碳达峰目标，减少人类活动的碳足迹，作为清洁能源重要一员的光伏迅速发展。近些年来，光伏电站的装机容量逐年攀升，提供了大量的光伏发电量。但由于光伏发电受天气，光照条件，云等环境因素影响较大，导致发电不够稳定。能否实现光伏电站输出功率的准确预测，将直接影响电网安全、电能质量和电站经济效益。本项目主要研究内容就是利用深度学习模型以及地基云图图像数据对未来0-4小时的光伏功率进行预测。围绕此内容我们开展了诸多研究。晴空模型可以帮助我们计算晴空指数，帮助判断天气状况，针对历史晴空模型的误差较大的问题，提出了一种基于调整输入和误差估计的晴空直接辐射估计方法；针对图像和数值多种输入的问题，提出了一种混合预测框架。同时针对图像时序信息编码的问题，提出了一种基于3D-CNN网络的短期预测方法，提取云图动态特征融合入时序预测模型ANN当中；针对时序光伏时序信息特征提取问题，提出了一种混合非线性的时序预测方法以及一种基于高斯过程的时许预测方法；针对光伏电站中，电池片温度对发电效率有影响的现状，提出了一种基于多态动态热能模型为光伏组件温度进行建模。通过从辐照度到组件各个部分的精确建模，从而能够更好的对输出功率进行预测，在实际电站中也有着实际应用价值。