UC San Diego

In summertime mornings, marine boundary layer (MBL) stratocumulus clouds commonly cover the southern California coast. The formation and dissipation of MBL stratocumulus affect the photovoltaic (PV) power. Increasing rooftop solar PV generation over the coast necessitates accurate solar forecasts to facilitate the reliable and economical integration of solar PV into the electric grid. For forecast horizons of hours to days, numerical weather prediction (NWP) and machine learning techniques are considered as the most accurate methods and widely used. However, the comparisons of NWP irradiance forecasts with ground measurements show that NWP models consistently overestimate the solar irradiance at the surface due to both clear sky biases and cloud modelling issues. As for machine learning techniques, most researchers either deliberately ignore meteorological conditions (for endogenous forecast models) or lack meteorological expertise to select meteorological input variables that go beyond classical weather station measurements such as air temperature, wind speed, and relative humidity. In this study, several methods are proposed to improve both NWP and machine learning forecast accuracy. Firstly, the clear sky irradiance bias in the New Goddard Shortwave (SW) scheme of Weather Research and Forecasting (WRF) scheme are found to be missing absorption of water vapor continuum. Use of the new parameterization of water vapor including water vapor continuum reduced WRF’s clear sky biases. Secondly, we confirmed the positive correlation between temperature inversion base height (IBH) and inland extent of MBL stratocumulus, and postulated that WRF underprediction of cloud cover extent is due to underprediction of IBH. A thermodynamic method was developed to modify the boundary layer temperature and moisture profiles to better represent the boundary layer structure in WRF. Validation against satellite global horizontal irradiance (GHI) demonstrated that the best IBH ensemble improves GHI accuracy by 23% mean absolute error compared to the baseline WRF model and is similar to 24-hour persistence forecasts for coastal marine layer region. The spatial error maps showed deeper inland cloud cover. Thirdly, we focused on selecting appropriate meteorological input variables based on the characteristics of MBL clouds and studied how accurately support vector machine (SVM), random forest (RF), and gradient boosting (GB) machine learning models predict solar radiation. All three models significantly outperform physics-based NWP models and 24-hour persistence in predicting solar radiation, especially during cloudy periods in the morning. The most important meteorological variables are found to be liquid water path, IBH, and thickness between 1000 and 500 mb pressure levels.