Ultra-Short-Term Photovoltaic Power Forecasting Based On Sky Images

Photovoltaic power generation has the advantages of low pollution,easy expansion,and low production cost.As a clean energy,it plays an important role in the process of replacing traditional fossil energy.In the process of photovoltaic grid-connected power generation,the stability of the output power directly affects the stability and safety of the overall grid operation.Therefore,accurate prediction of the output power is very important in the process of ensuring the stable operation of the grid.The change of cloud layer is the main factor affecting the fluctuation of ultra-short-term photovoltaic power generation.When the cloud cover changes the sunlight,the irradiance received by the ground in a short period of time changes accordingly,which in turn affects the output power of photovoltaic panels.In the past,the prediction algorithms using sky images could not make full use of the information in the images and ignored the cloud changes in multiple consecutive sky images.There is still room for improvement in ultra-short-term prediction.In this paper,under the premise of ensuring sufficient computing power for edge deployment,the RAFT optical flow network is used to extract the optical flow information of sky images in adjacent time intervals,and the irradiance is predicted by the twin neural network with Conv Ne Xt as the backbone network.Finally,the irradiance and The mapping relationship model between powers predicts the output power of photovoltaic panels.The main research content of this paper is as follows:(1)Aiming at the problems that traditional forecasting methods are not sensitive to cloud layer changes and insufficient utilization of continuous sky image information,a forecasting model combining Siamese neural network and RAFT optical flow network is proposed.The RAFT optical flow network extracts the optical flow information of the sky image in adjacent time intervals.The optical flow information contains richer cloud layer change details,which is richer than the feature information extracted by the backbone network from a single frame image.In the case of improving the prediction results,it takes into account the computing resource consumption and real-time requirements at runtime.(2)Replace the ResNet backbone network to extract features,use ConvNeXt as the backbone network’s twin neural network to predict irradiance,and improve the overall performance of the neural network.The Siamese neural network extracts multi-step interval information by comparing the change information between similar input and output features,and solves the problem of insufficient utilization of adjacent interval information by traditional prediction methods.In this paper,the performance of the model is measured based on the UCSD Folsom sky image dataset,and the RMSE index reaches 73.8W/m2,which is5.9% lower than the prediction based on the stacking method.(3)Use the Tensor RT-based deep learning reasoning solution to accelerate model quantization,shrink CUDA core usage,improve throughput,and reduce the dependence of edge computing end model reasoning on computing resources.The model inference time is compressed from 76 ms in Py Torch model to 46 ms in Float32 mode and 42 ms in Int8 mode,and the inference speed of the model has been improved.