A modern renewable energy forecasting system blends physical models with artificial intelligence to aid in system operation and grid integration. This paper describes such a system being developed for the Shagaya Renewable Energy Park, which is being developed by the State of Kuwait. The park contains wind turbines, photovoltaic panels, and concentrated solar renewable energy technologies with storage capabilities. The fully operational Kuwait Renewable Energy Prediction System (KREPS) employs artificial intelligence (AI) in multiple portions of the forecasting structure and processes, both for short-range forecasting (i.e., the next six hours) as well as for forecasts several days out. These AI methods work synergistically with the dynamical/physical models employed. This paper briefly describes the methodology used for each of the AI methods, how they are blended, and provides a preliminary assessment of their relative value to the prediction system. Each operational AI component adds value to the system. KREPS is an example of a fully integrated state-of-the-science forecasting system for renewable energy.

Wildland fire decision support systems require accurate predictions of wildland fire spread. Fuel moisture content (FMC) is one of the important parameters controlling the rate of spread of wildland fire. However, dead FMC measurements are provided by a relatively sparse network of remote automatic weather stations (RAWS), while live FMC is relatively infrequently measured manually. We developed a high resolution, gridded, real-time FMC data sets that did not previously exist for assimilation into operational wildland fire prediction systems based on ML. We used surface observations of live and dead FMC to train machine learning models to estimate FMC based on satellite observations. Moderate Resolution Imaging Spectrometer Terra and Aqua reflectances are used to predict the live and dead FMC measured by the Wildland Fire Assessment System and RAWS). We evaluate multiple machine learning methods including multiple linear regression, random forests (RFs), gradient boosted regression and artificial neural networks. The models are trained to learn the relationships between the satellite reflectances, surface weather and soil moisture observations and FMC. After training on data corresponding to the temporally and spatially nearest grid points to the irregularly spaced surface FMC observations, the machine learning models could be applied to all grid cells for a gridded product over the Conterminous United States (CONUS). The results show generally that the rule-based approaches have the lowest errors likely due to the sharp decision boundaries among the predictors, and the RF approach that utilizes bagging to avoid over-fitting has the lowest error on the test dataset. The errors are typically between 25%−33% the typical variability of the FMC data, which indicate the skill of the RF in estimating the FMC based on satellite data and surface characteristics. The FMC gridded product based on the RF runs operationally daily over CONUS and can be assimilated into WRF-Fire for more accurate wildland fire spread predictions.

This paper describes the development and testing of a cloud-regime-dependent short-range solar irradiance forecasting system for predictions of 15-min-average clearness index (global horizontal irradiance). This regime-dependent artificial neural network (RD-ANN) system classifies cloud regimes with a k-means algorithm on the basis of a combination of surface weather observations, irradiance observations, and GOES-East satellite data. The ANNs are then trained on each cloud regime to predict the clearness index. This RD-ANN system improves over the mean absolute error of the baseline clearness-index persistence predictions by 1.0%, 21.0%, 26.4%, and 27.4% at the 15-, 60-, 120-, and 180-min forecast lead times, respectively. In addition, a version of this method configured to predict the irradiance variability predicts irradiance variability more accurately than does a smart persistence technique.

In order for numerical weather prediction (NWP) models to correctly predict solar irradiance reaching the earth’s surface for more accurate solar power forecasting, it is important to initialize the NWP model with accurate cloud information. Knowing where the clouds are located is the first step. Using data from geostationary satellites is an attractive possibility given the low latencies and high spatio-temporal resolution provided nowadays. Here, we explore the potential of utilizing the random forest machine learning method to generate the cloud mask from GOES-16 radiances. We first perform a predictor selection process to determine the optimal predictor set for the random forest predictions of the horizontal cloud fraction and then determine the appropriate threshold to generate the cloud mask prediction. The results show that the random forest method performs as well as the GOES-16 level 2 clear sky mask product with the ability to customize the threshold for under or over predicting cloud cover. Further developments to enhance the cloud mask estimations for improved short-term solar irradiance and power forecasting with the MAD-WRF NWP model are discussed.

This work compares the solar power forecasting performance of tree-based methods that include implicit regime-based models to explicit regime separation methods that utilize both unsupervised and supervised machine learning techniques. Previous studies have shown an improvement utilizing a regime-based machine learning approach in a climate with diverse cloud conditions. This study compares the machine learning approaches for solar power prediction at the Shagaya Renewable Energy Park in Kuwait, which is in an arid desert climate characterized by abundant sunshine. The regime-dependent artificial neural network models undergo a comprehensive parameter and hyperparameter tuning analysis to minimize the prediction errors on a test dataset. The final results that compare the different methods are computed on an independent validation dataset. The results show that the tree-based methods, the regression model tree approach, performs better than the explicit regime-dependent approach. These results appear to be a function of the predominantly sunny conditions that limit the ability of an unsupervised technique to separate regimes for which the relationship between the predictors and the predictand would differ for the supervised learning technique.