Research And Implementation Of Direct Normal Irradiance Prediction Method For Tower Photothermal Power Station

The tower photothermal power station uses the heliostat field to focus the sunlight on the heat sink for power generation.The energy source is Direct Normal Irradiance(DNI),and the movement of the cloud is the main factor affecting DNI.When the clouds in the sky block or leave the sun,sunlight reaching the heliostat will suddenly change,causing DNI fluctuations which will impact the heat sink and affect the stable operation of the photothermal power station.Therefore,predicting the DNI curve can optimize the steam turbine operating load curve,improve the power generation efficiency,and provide safety for the stable operation of the photothermal power station.This paper analyzes the influencing factors of DNI and the requirements of the prediction,and designs a project that can realize the ultra-short-term prediction of DNI.The project uses the full sky imaging and radiometer to collect data.Firstly,cloud detection and cloud speed calculation are performed on the full sky image to obtain the occlusion information of the future sun,and then the fusion model algorithm is designed to predict the real-time DNI.The main research contents of this paper are as follows:(1)Based on the imaging method of the full sky imager and the imaging characteristics of the sun,a method for correcting the distortion of the full sky image and the compensation of the vertical axis deviation using the sun position is designed.Considering that threshold recognition and cloud detection based on clear-sky background fitting can't solve the problem of cloud cluster recognition in the halo region,this paper proposes to extract the RGB three-channel gray value of the full sky image and the RGB three-channel gray value of the clear sky background as features,training random forest models to detect clouds.(2)The pixel template matching cloud speed algorithm and the SURF(Speeded Up Robust Features)feature point matching based cloud speed algorithm are designed and compared.The template matching algorithm is selected to calculate the cloud speed.Then the cloud speed is corrected based on Kalman filtering,and the trajectory of the sun relative to the cloud is predicted.(3)The image information that can represent the cloud are researched.Based on this,occlusion information is extracted as features.SVR,GBRT and Xgboost are selected as underlying models,and model fusion algorithm based on stacking is designed to predict DNIAfter verification,the article can correctly realize cloud detection and cloud speed calculation.The article can predict the DNI value in the next half hour.Compared with the actual DNI value,the predicted DNI value has a small error and can meet the expected design goals and performance requirements.