Super-short Term Forecast Of Solar Irradiance Based On Numerical And Ground-based Cloud Images

As a kind of abundant,clean and safe renewable energy,solar energy is widely concerned.At present,the development and utilization of photoelectricity has become an important direction of its development.In recent years,with the increase of global photovoltaic power installation,grid-connected photovoltaic power generation is also growing rapidly.However,the output power of photovoltaic power generation is fluctuating and intermittent.When the grid-connected photovoltaic power generation exceeds a certain proportion,the fluctuating photovoltaic output will bring serious impact on the stability,safe operation and power quality of the power system.The main factor that causes the randomness and fluctuation of photovoltaic output power is the uncertainty of solar irradiance.Therefore,it is an urgent task to accurately predict the solar radiation received by photovoltaic power station,which is the premise of the safety and stability of photovoltaic grid-connected system,and it is of great significance to the power grid dispatching decision.However,due to the influence of various meteorological factors such as cloud cover,aerosol thickness and humidity,especially the shielding caused by cloud movement,solar radiation has high volatility and strong randomness,which makes it difficult for traditional time series prediction methods or machine learning models to obtain good prediction results.In order to obtain accurate ultra-short-term solar irradiance,this paper builds a variety of prediction models based on three different monitoring data,namely,single value,multi-value and cloud image,and discusses the role and prediction performance of different inputs and the constructed models in predicting ultra-short-term solar radiation.Specific research contents are as follows:(1)Firstly,in view of the problem that the direct solar irradiance(DNI)changes constantly with the distance between the sun and the observation station in clear sky,the basic clear sky model is used to convert the predicted target DNI into clear sky coefficient,so as to eliminate the prediction error caused by the change of the sun’s position.Then to clear sky coefficient of a single variable as the research object,constructs the support vector regression,multilayer perceptron,extreme learning machine and long short-term memory network forecast model,the prediction model was established according to the rule of BIC information input dimension,with America’s national renewable energy laboratory site open data training and test the prediction model.The experimental results show that the long-short-term memory network model is relatively optimal in predicting the ultra-short-term DNI,which lays a foundation for improving the model and improving the prediction accuracy in the future.(2)A variety of meteorological factors affecting the variation of the DNI were discussed from the solar irradiation transmission process,and nine numerical variables including aerosol optical thickness,atmospheric mass,relative humidity,temperature,air pressure,cloud cover,wind speed,albedo and clear sky coefficient were selected as the input of the model.Meanwhile,based on the above optimal model,a convolution-bidirectional long short-term memory network model(CNN-BI-LSTM)is proposed.The model uses one-dimensional convolutional neural network to carry out data spatial dimension fusion and feature extraction on the above 9 different types of impact factor data sequences,and the feature extraction of time dimension is completed through the BI-LSTM layer,thus improving the prediction accuracy of the DNI in the next 10 minutes.(3)Considering that cloud,the meteorological factor,is the key variable that causes the mutation of DNI,and the single numerical cloud data cannot fully reflect the cloud information,the siamese convolutional neural networked model(Siamese CNN-LSTM)is proposed based on the numerical input and the foundation cloud image to complement the meteorological cloud information.Firstly,the Siamese CNN part is used to extract the features of multiple continuous moment cloud images,then the numerical data and cloud image data are fused through the fusion layer.Finally,the LSTM part extract the features of the time dimension of the fusion features and predict the DNI in the next 10 minutes.The experimental results show that the proposed model has better performance than other models in predicting the DNI in the next 10 minutes,especially in cloudy and rainy days.