| Solar photovoltaic(PV)power generation is one of the main means to achieve"carbon neutrality"and"carbon peaking",and the short-term solar radiation forecast(<6 h)ensures the PV power generation schedule.The existing forecast models all suffer from low forecast accuracy under cloudy conditions.The Smart persistence model is advanced compared to the Simple model,which is based on the persistence of clear-sky index with high accuracy under clear days and is often used to evaluate the performance of other models,but the forecast accuracy is still low in the presence of clouds.Improving the forecast accuracy of persistence models will be of great significance for enhancing the baseline of short-term solar radiation forecasting.This study simplifies the equations deriving cloud properties by relating surface-measured irradiances to cloud properties based on radiative transfer framework,and four new models with clear physics are developed that embodies the idea of the persistence of cloud properties are developed.The long-term observations(1998-2014)of the global horizontal irradiance GHI,direct normal irradiance DNI,and diffuse horizontal irradiance DHI with 15min resolution at the Southern Great Plains(SGP)in the USA are used for model evaluation.The performance of the models in a cloudy case,under the condition of eight different types of clouds and the whole study period(including clear,cloudy,overcast,rainy,and various weather combinations)are examined separately.The reasons for the differences in forecast accuracy among different persistence models are analyzed,the generalizability of the new models is verified,the defects of the models are pointed out as well.Finally,based on the derivation of the cloud properties,these predictors are introduced into data-driven models(including two machine learning models LSTM,XGBoost and a statistical model ARIMA)step by step to make the machine learning models explicable by the differences in predictors and model performances among different levels.Meanwhile,the advantages and disadvantages of various models and their applicability are clarified by the comparison in the performance of the cloud-property-based models and data-driven model in solar radiation forecasting,and the potential factors influencing model performance are discussed.The main results are as follows:1)Based on the radiative transfer equation,the theoretical relationship between cloud properties and observed irradiances is investigated.By use of the simplified equation,cloud relative radiative forcing RCRF,cloud relative radiative forcing ratio R,cloud albedo a_r and cloud fraction f can be easily derived from the observed GHI and DNI.The estimated cloud fraction and cloud albedo agree well with the measurement.The cloud parameters are classified into four layers regarding the progressive relationship between those cloud parameters and irradiances,and four new models associated with the hierarchy are constructed enlightened by the idea of persistence.The models based on RCRF and R(RCRF-PM and R-PM)are at the second and third levels,respectively,and the cloud fraction and cloud albedo based models(CA-PM and CF-PMR)are at the fourth level.The newly constructed persistence model can predict GHI,DNI and DHI being superior to the conventional Smart model which can only predict GHI.However,the RCRF-PM can be regarded as the extended Smart model in DNI and DHI forecast.2)The performance of the models is evaluated by an individual cloudy case,290,000observations under eight different cloud types as well as all the measurements at the SGP Central Facility site.The results show that the higher the level of the newly constructed models,the overall better performance with longer extension on forecasting lead time compared to the Simple and Smart model.The lead time in GHI is extended by 2?5 h,DNI is extended by0.75?1 h,and DHI is extended by 0.25 h at most relative to the Simple model.The extension from the new models ranges from 0.25 h to 0.75 h compared to the Smart model(RCRF-PM)for all radiative components.The improvements on the model performance can at most reach to 68%,47%and 29%for GHI,DNI and DHI by comparing the new model with the Simple model,the improvements over Smart model are 30%,16%and 23%,respectively.The new models have the greatest improvement in cumulus and cirrus clouds,followed by stratus clouds,and the strong convective clouds have the smallest improvements.3)The main reason for the higher-level models to have better forecast performance is that the higher the level,the better persistence in the corresponding cloud parameter.Although the fourth-level model has the overall highest forecast accuracy,CF-PM and CA-PM have the best performance in forecasting GHI and DNI,DHI,respectively.Further analysis shows that model performance is not only related to the persistence of the cloud covariates but also associated with transferability from the predictor error to the forecast error.Better persistence as well as weak transferability would lead to higher accuracy.The evaluation results from the other 31sites at SGP draw a consistent conclusion that model forecast accuracy is improved with the increase of the levels,despite of the differences in the observed solar irradiances at different sites,confirming the good generalizability of the newly constructed models.However,the new models also have defects that cloud fraction and cloud albedo can not be estimated in some cases which limits the application of the model to some extent.4)To overcome the shortcomings of the persistent model,more advanced data-driven models are introduced.The higher the level,the more predictors are employed in the data-driven model.Comparison between the two types of models show that the statistical model ARIMA only has better forecasting performance than the Simple model,and is inferior to the newly developed models,but the machine learning models LSTM and XGBoost have the best performance over all models.The more cloud predictors considered,the better the forecasting performance and the longer extension in the lead time.In general,the forecast factor itself contributes most to the forecast accuracy,the introduction of cloud fraction and cloud albedo also significantly improves the forecast performance,followed by aerosol quantities,but clear sky irradiance and the meteorological factors having smaller impacts on the forecast performance.Introducing cloud factors into the machine learning model step by step makes the machine learning models physical explainable.The simplification of the forecasting technique and the breakthrough makes it easy to achieve high accuracy solar nowcasting for PV plants that lack long-term observations.The new models also have improved generalizability to different weather and climate regimes and set a higher standard as benchmark models challenging the development of other physical models and machine learning models in solar radiation forecasting. |
PDF Full Text Request