Solar radiation influences many and diverse fields like energy generation, agriculture and building operation.Hence, simulation models in these fields often rely on precise information about solar radiation. Measureme...Solar radiation influences many and diverse fields like energy generation, agriculture and building operation.Hence, simulation models in these fields often rely on precise information about solar radiation. Measurementsare often restricted to global irradiance, whereby measurements of its single components, direct and diffuseirradiance, are sparse. However, information on both, the direct and diffuse irradiance, is necessary forsimulation models to work reliably. In this study, solar separation models are developed using 10-min trainingdata from two different locations in Austria. Direct horizontal irradiance is predicted via regressing the directfraction using several objective functions. The models are first trained on a data set including data from bothlocations, and evaluated regarding root mean squared deviation (RMSD), mean bias deviation (MBD), andcoefficient of determination (R2) on measured and estimated direct normal irradiance. The two best performing models are then selected for further analysis. This analysis comprises of an evaluation of the models per season,transferability of trained modes between two locations in Austria, a transferability and generalisability studyconducted for four more locations in Central Europe, and a comparison with the trusted Engerer model. Thesolar separation model with polynomial terms up to order three and Ridge regularisation outperforms thesecond model based a logistic term in combination with mixed quadratic terms as well as the Engerer model.展开更多
文摘Solar radiation influences many and diverse fields like energy generation, agriculture and building operation.Hence, simulation models in these fields often rely on precise information about solar radiation. Measurementsare often restricted to global irradiance, whereby measurements of its single components, direct and diffuseirradiance, are sparse. However, information on both, the direct and diffuse irradiance, is necessary forsimulation models to work reliably. In this study, solar separation models are developed using 10-min trainingdata from two different locations in Austria. Direct horizontal irradiance is predicted via regressing the directfraction using several objective functions. The models are first trained on a data set including data from bothlocations, and evaluated regarding root mean squared deviation (RMSD), mean bias deviation (MBD), andcoefficient of determination (R2) on measured and estimated direct normal irradiance. The two best performing models are then selected for further analysis. This analysis comprises of an evaluation of the models per season,transferability of trained modes between two locations in Austria, a transferability and generalisability studyconducted for four more locations in Central Europe, and a comparison with the trusted Engerer model. Thesolar separation model with polynomial terms up to order three and Ridge regularisation outperforms thesecond model based a logistic term in combination with mixed quadratic terms as well as the Engerer model.
