Improved method for solar resource assessment using simulations from the numerical weather prediction model Meso-NH and artificial neural networks

Abstract

Numerical Weather Prediction (NWP) models tend to show significant errors to what solar radiation forecast is concerned, mainly due to difficulties with the correct parameterization of clouds, the usage of monthly-mean aerosol climatology, and the high computational effort associated to the detailed modeling of radiative transfer in the atmosphere. This is especially evident in the case of direct normal (DNI) and diffuse horizontal irradiation (DHI) while global horizontal irradiation (GHI) predictions show a smaller bias compared with ground-based measurements. Although GHI is sufficient to obtain a rough estimation of energy output in solar photovoltaic systems, knowledge of both DNI and DHI components of solar radiation, or DNI and GHI, allows obtaining an improved and more detailed prediction of power output of such systems. This work presents a method for the assessment of the solar resource (DNI and GHI) obtained by developing a site adaptation model based on artificial neural networks (ANN) and using as input the results of Meso-NH NWP model simulations and aerosol data from Copernicus Atmosphere Monitoring Service (CAMS). A typical meteorological year (TMY) was determined for Évora, a city centered in the region of the south of Portugal, using sixteen years of ground-based measurements of nine meteorological variables related to solar radiation, air temperature, relative humidity, and wind speed. The weather simulations for this TMY were obtained using the NWP model Meso-NH coupled with the ecRad radiation scheme which generates atmospheric data for the desired region (south of Portugal) with a resolution of 1.25 km. The simulation results showed an overestimation of the solar resource (higher for DNI) that can be caused due to difficulties in correctly simulating clouds and the usage of monthly-mean aerosol climatologies. Thus, a method for site adaptation using an optimized ANN model was developed, which is based on the testing of different parameters, including the addition of aerosol data from CAMS as inputs, and choosing the variables that result in higher improvements in terms of mean squared error. The ANN models are obtained for the same location used for the generation of the TMY, in this case Évora, and then applied to the whole region simulated by the NWP model (south of Portugal). The resulting annual relative mean bias errors for GHI and DNI at Évora and typical TMY are of 0.55 % and 0.98 %, respectively, while the values for the NWP simulations are of 8.24 % for GHI and 31.71 % for DNI. This method is then validated using ground-based data from a network of eight solar radiation measuring stations scattered throughout the south of Portugal, showing relative mean bias errors of 2.34 % for GHI and 3.41 % for DNI, while the NWP simulations presents relative mean bias errors of 8.50 % and 29.54 %, respectively, for the same locations. The method presented is reproducible and can be used to any region of the world, however, it should be noted that the area simulated in this work is of approximately 46 000 km2 and the generation of larger maps and/or for locations with very different climates will probably decrease the improvement brought by the application of an ANN model developed for a specific site in that region. The understanding of these implications can be a topic for future research. The high resolution GHI and DNI maps generated by this method are extremely important for solar radiation applications, such as planning, design and dimensioning of solar energy systems. Thus, the usage of artificial neural networks incorporating aerosol data as site adaptation method of NWP models output allows for a more accurate assessment of the solar resource.