Forecasting the PV Panel Power Based on Image Processing and Historical Outputs

The installation of solar photovoltaic(PV) systems is rapidly expanding worldwide, but the intermittency of solar power generation hinders its deeper integration with the grid. Short-term PV fluctuations are partially caused by sudden weather changes, such as variations in cloud cover, which can significantly affect the output of PV panels on a minute-scale time frame. Sky images can provide information about current and upcoming cloud cover, thereby improving the accuracy of solar power forecasts. This study uses a convolutional neural network(CNN) to link the power output of solar panels with current sky images and utilizes the historical output data of PV panels to further enhance the prediction accuracy. Additionally, the sensitivity of the model to machine learning configuration parameters(such as the number of neurons and network width) was evaluated, along with the uncertainty of the stochastic methods used and the impact of different input and output arrangements on performance metrics. The proposed model achieved a root mean square error(RMSE) of 2.37 kW on the actual sampled dataset.