Very-Short-Term Solar Prediction Based On Heterogeneous Data Fusion

The high variability of renewable energy is a major obstacle to the integration of a large number of renewable energy sources in the power system.While wind energy prediction techniques are pretty mature,very-short-term solar forecasting is more challenging and there are few satisfactory solutions in the literature so far.Because of the strong influence of clouds,there are severe and irregular fluctuations in solar productions.Solar prediction with high precision and fine time resolution can alleviate the negative impact caused by solar energy grid integration,and thus increase the penetration level of solar energy into the grid.This work aims to accurately predict solar radiation values a few minutes in advance,especially to enable early prediction of ramp events.However,traditional solar prediction methods based on a single data source cannot capture and predict ramp events well.Therefore,this paper proposes a solar energy prediction method based on heterogeneous data fusion using deep learning algorithms.The heterogeneous data adopted in our prediction algorithms include all-sky images,meteorological data and solar geometry data.The main contents of our work are as follows:Firstly,in view of the strong influence of cloud motion on solar energy variations,this paper uses Convolutional Neural Networks and Transfer Learning algorithms to extract sky image features.Firstly,we build a CNN classification network to identify whether the sun is shaded by the clouds or not,and then migrate the features learned by the network to another regression network and verify the validity of image features.The experimental results show that R2 is 0.8185 and the total energy error percentage is only 1.91%.Compared with other high-complexity methods for predicting solar energy using sky images,the feature extraction system proposed in this paper can capture the key information with much less complexity and lower dimensionality.Furthermore,the features can easily be used for other solar prediction algorithms.Secondly,to address the challenges of the integration of multi-source data and the sequential dependence among heterogeneous data,this paper uses Long-Short Term Memory algorithm for heterogeneous data fusion and solar time series prediction.This paper predicts the average global horizontal irradiance(GHI)in 5 minutes and 10 minutes by combining three data from the past 30 minutes and compares it with the naive predictor and four other solar prediction algorithms.The experimental results show that predicting solar radiation in 5 minutes is better than that in 10 minutes with NRMSE as low as 0.089 and energy error as low as 2.71%.Compared with the naive predictor,our proposed algorithm achieves a 40%performance improvement.Compared with other algorithms,the system prediction error is reduced by 15%to 44%.Our research shows that the proposed prediction system is superior to other methods in the literature in terms of accuracy and robustness.This work has the significant contribution of reducing the impact of slope events on the power systems.