Search

Consistency, in a narrow sense, denotes the alignment between the forecast-optimization strategy and the verification directive. The current recommended deterministic solar forecast verification practice is to report the skill score based on root mean square error (RMSE), which would violate the notion of consistency if the forecasts are optimized under another strategy such as minimizing the mean absolute error (MAE). This paper overcomes such difficulty by proposing a so-called "potential RMSE skill score," which depends only on (1) the cross-correlation between forecasts and observations and (2) the autocorrelation of observations. While greatly simplifying the calculation, the new skill score does not discriminate inconsistent forecasts as much, e.g., even MAE-optimized forecasts can attain a high RMSE skill score. [ABSTRACT FROM AUTHOR]

• Five data availability scenarios are compared for different physical model chains. • Nominal power and module orientation are the most critical design parameters. • Unknown tilt and azimuth angles and row spacing should be estimated. • Absolute error is more sensitive to missing design parameters than squared error. • Physical power calculation is encouraged even if some design parameters are unknown. Accurate photovoltaic (PV) power forecasts are essential for the grid integration of the technology. Most solar forecasting studies deal only with solar irradiance forecasting, which highlights the importance of the irradiance to power conversion methods to enable the practical utilization of the advanced irradiance forecasting techniques. The PV power output is commonly calculated with physical model chains that rely on the main design parameters of the PV plants. This study presents an analysis of how the unavailability or uncertain knowledge of the design parameters affects the physical power forecast accuracy. The analysis is performed by verifying the power forecasts created by the bests of 32,400 model chains for 16 PV plants in Hungary for five data availability scenarios based on numerical weather prediction data. The results show that even though the knowledge of all relevant design parameters ensures the lowers errors, the critical parameters are only the nameplate capacities and the module orientation. The increase in mean absolute error and root mean square error is only 1.3% and 0.5%, respectively, if only these crucial parameters are known compared to the baseline scenario. If necessary, the tilt and azimuth angles can also be inferred, where the resulting inaccuracy depends on the estimation errors. The results reveal that the physical model chains can be used with a decent accuracy for power forecast calculations even if the design parameters are unknown, which highly broadens the applicability of the physical method. [ABSTRACT FROM AUTHOR]

Under the two-step framework of photovoltaic (PV) power forecasting, that is, forecasting first the irradiance and then converting it to PV power, there are two chief ways in which one can account for the uncertainty embedded in the final PV power forecast. One of those is to produce probabilistic irradiance forecast through, for example, ensemble numerical weather prediction (NWP), and the other is to pass the irradiance forecast through a collection of different irradiance-to-power conversion sequences, which are known as model chains. This work investigates, for the first time, into the question: Whether pairing ensemble NWP with ensemble model chain is better than leveraging any individual method alone? Using data from 14 utility-scale ground-mounted PV plants in Hungary and the state-of-the-art global mesoscale NWP model of the European Centre for Medium-Range Weather Forecasts, it is herein demonstrated that the best probabilistic PV power forecast needs to consider both ensemble NWP and ensemble model chain. Furthermore, owing to the higher-quality probabilistic forecasts, the point forecast accuracy is also improved substantially through pairing. Overall, the recommended paring strategy achieves a mean-normalized continuous ranked probability score and a root mean square error of 18.4% and 42.1%, respectively. • New physical ensemble methods are presented for photovoltaic power forecasting. • The added value of ensemble numerical weather prediction is quantified. • Underdispersed raw ensembles can be effectively calibrated by linear quantile regression. • Improved probabilistic forecasting methods also lead to better deterministic forecasts. [ABSTRACT FROM AUTHOR]

Irradiance-to-power conversion is an essential step of state-of-the-art photovoltaic (PV) power forecasting, regardless of the source and post-processing of irradiance forecasts. The two distinct approaches for mapping the irradiance forecasts to PV power are physical and data-driven, which can also be hybridized. The contribution of this paper is twofold; first, it proposes a concept and identifies the best implementation of a hybrid physical and machine learning irradiance-to-power conversion method. Second, a head-to-head comparison of the physical, data-driven, and hybrid methods is performed for the operational day-ahead power forecasting of 14 PV plants in Hungary based on numerical weather prediction (NWP). To respect the rule of consistency but still obtain as complete picture as possible, two directives are set, namely minimizing the mean absolute error (MAE) and minimizing the root mean square error (RMSE), and separate sets of forecasts are optimized for both directives. The results reveal that for two years of training data, the hybrid method that involves the most physically-calculated predictors can reduce the MAE by 5.2% and 10.4% compared, respectively, to the optimized physical model chains and the machine learning without any physical considerations. The two most important physical modeling steps are separation and transposition modeling, and the rest of the physical PV simulation can be left to machine learning in hybrid models without a significant increase in the errors. The optimization of the physical model chains is found to be important even in the case of hybrid modeling; therefore, it should become a standard procedure in practical applications. Finally, the hybrid method is only beneficial for at least one year of training data, while in the initial period of the operation of a PV plant, it is advised to stay with optimized physical modeling. The guidelines and recommendations of this paper can help both researchers and practitioners design and optimize their power conversion model to increase the accuracy of the PV power forecasts. • New concept for hybrid physical and machine learning irradiance-to-power conversion. • Comparison of physical, data-driven, and hybrid approaches for two directives. • Physical global tilted irradiance calculation is sufficient in hybrid modeling. • Optimization of physical model chains is important even in hybrid models. • The recommended method depends on the length of the available training data. [ABSTRACT FROM AUTHOR]

The increase of the worldwide installed photovoltaic (PV) capacity and the intermittent nature of the solar resource highlights the importance of power forecasting for the grid integration of the technology. This study compares 24 machine learning models for deterministic day-ahead power forecasting based on numerical weather predictions (NWP), tested for two-year-long 15-min resolution datasets of 16 PV plants in Hungary. The effects of the predictor selection and the benefits of the hyperparameter tuning are also evaluated. The results show that the two most accurate models are kernel ridge regression and multilayer perceptron with an up to 44.6% forecast skill score over persistence. Supplementing the basic NWP data with Sun position angles and statistically processed irradiance values as the inputs of the learning models results in a 13.1% decrease of the root mean square error (RMSE), which underlines the importance of the predictor selection. The hyperparameter tuning is essential to exploit the full potential of the models, especially for the less robust models, which are prone to under or overfitting without proper tuning. The overall best forecasts have a 13.9% lower RMSE compared to the baseline scenario of using linear regression. Moreover, the power forecasts based on only daily average irradiance forecasts and the Sun position angles have only a 1.5% higher RMSE than the best scenario, which demonstrates the effectiveness of machine learning even for limited data availability. The results of this paper can support both researchers and practitioners in constructing the best data-driven techniques for NWP-based PV power forecasting. • 24 machine learning models tested for day-ahead photovoltaic power forecasting. • Kernel ridge and multilayer perceptron are the overall most accurate models. • Predictor selection is even more important than model selection. • Hyperparameter optimization is essential for the highest accuracy. • Up to 13.9% RMSE improvement over a baseline linear regression model. [ABSTRACT FROM AUTHOR]

• Irradiance separation and transposition modeling are the two most critical steps. • 13% MAE, 12% RMSE and 23–33% skill score improvement is possible by model selection. • Detailed physical model chains, inlc. reflection and shading, have the lowest MAE. • Simple models have the lowest RMSE, but high bias and underdispersed forecast. • Wind speed has only marginal effect on physical photovoltaic power forecasting. Forecasting the power production of grid-connected photovoltaic (PV) power plants is essential for both the profitability and the prospects of the technology. Physically inspired modelling represents a common approach in calculating the expected power output from numerical weather prediction data. The model selection has a high effect on physical PV power forecasting accuracy, as the difference between the most and least accurate model chains is 13% in mean absolute error (MAE), 12% in root mean square error (RMSE), and 23–33% in skill scores for a PV plant on average. The power forecast performance analysis performed and verified for one-year 15-min resolution production data of 16 PV plants in Hungary for day-ahead and intraday time horizons on all possible combinations of nine direct and diffuse irradiance separation, ten tilted irradiance transposition, three reflection loss, five cell temperature, four PV module performance, two shading loss, and three inverter models. The two most critical calculation steps are identified as irradiance separation and transposition modelling, while the inverter models are the least important. Absolute and squared errors are two conflicting metrics, as the more detailed models result in the lowest MAE, while the simplest ones have the lowest RMSE. Wind speed forecasts have only a marginal effect on the PV power prediction. The results of this study contribute to a deeper understanding of the physical forecasting approach in the research community, while the main conclusions are also beneficial for PV plant owners in preparing their generation forecasts. [ABSTRACT FROM AUTHOR]

Since directly measuring beam and diffuse irradiance is not feasible on many occasions, one often has to resort to estimating the beam and diffuse irradiance components from the global irradiance, which is known as separation modeling. Separation modeling is essentially a nonlinear regression problem, with the clearness index being the main input and the diffuse fraction being the output. Hundreds of separation models with various complexities have been proposed, among which the Yang4 model was recently validated using worldwide data as the quasi-universal choice for 1-min data. In this work, Yang4 is further improved by regime-dependent fitting, i.e., fitting a separate set of model coefficients for each climatological regime. Different regimes are determined through clustering of cloud cover frequency, aerosol optical depth, and surface albedo climatology maps. The new Yang5 model is able to outperform its predecessor at the 126 stations tested, covering a wide range of climate types. Overall, the normalized root mean square errors for beam normal irradiance (BNI) and diffuse horizontal irradiance (DHI) of Yang5 are 17.55% and 32.92% on average, as compared to 19.13% and 34.94% for the next best model, namely, Yang4. Furthermore, through conducting pairwise Diebold–Mariano tests, Yang5 is shown superior to Yang4 at 110/126 sites for BNI prediction and 93/126 for DHI. [Display omitted] • A concise summary of recent advances in separation modeling is given. • A regime-dependent version of the Yang4 model is proposed. • Regimes are defined by clustering cloud, aerosol, and albedo climatology. • The new Yang5 model outperforms the quasi-universal Yang4. [ABSTRACT FROM AUTHOR]