A Novel Ensemble Algorithm for Solar Power Forecasting Based on Kernel Density Estimation

Abstract

A novel ensemble algorithm based on kernel density estimation (KDE) is proposed to forecast distributed generation (DG) from renewable energy sources (RES). The proposed method relies solely on publicly available historical input variables (e.g., meteorological forecasts) and the corresponding local output (e.g., recorded power generation). Given a new case (with forecasted meteorological variables), the resulting power generation is forecasted. This is performed by calculating a KDE-based similarity index to determine a set of most similar cases from the historical dataset. Then, the outputs of the most similar cases are used to calculate an ensemble prediction. The method is tested using historical weather forecasts and recorded generation of a PV installation in Portugal. Despite only being given averaged data as input, the algorithm is shown to be capable of predicting uncertainties associated with high frequency weather variations, outperforming deterministic predictions based on solar irradiance forecasts. Moreover, the algorithm is shown to outperform a neural network (NN) in most test cases while being exceptionally faster (32 times). Given that the proposed model only relies on public locally-metered data, it is a convenient tool for DG owners/operators to effectively forecast their expected generation without depending on private/proprietary data or divulging their own.

Inverter Efficiency Analysis Model Based on Solar Power Estimation Using Solar Radiation

“…This study utilized a simple primary linear model, whereas studies for comparison [34,35] used neural networks such as Recurrent Neural Network(RNN) and LSTM. It can be safely said that our study yielded qualitatively good results when comparing the computing power required for the calculation of each model as well as the complexity of the calculations.…”

“…Some studies are ongoing with the goal of estimating solar radiation to predict future power output [30][31][32]. There are also studies on power output estimation based on ambient temperature, wind velocity, and incident light [33][34][35]. The method is based on historical weather data; there is a high correlation between the weather conditions in the present or past, and the solar power generation in the future.…”

Regression and Generalized Additive Model to Enhance the Performance of Photovoltaic Power Ensemble Predictors

“…Unlike parametric regression that exploits a known, predetermined relationship, nonparametric regression discovers the relationships through curve-fits that minimize a prescribed error measure, RMSE in this case. Different models exist, such as local averaging [49], local regression [50], kernel smoothing [51], [52], and wavelet transforms [53]- [56]. The proposed approach summarized in Algorithm 2 applies polynomial fitting with up to 9 • to capture dependencies, and the degree with the lowest error score is used as the best fit.…”

Generation Data of Synthetic High Frequency Solar Irradiance for Data-Driven Decision-Making in Electrical Distribution Grids

“…Finally, the prerequisite on long historical data for downscaling and data discovery of solar irradiance variation has been relaxed in [24], which has used the datasets existing in the World Radiation Monitoring Center (WRMC), the central archive of the Baseline Surface Radiation Network (BSRN) [25]. In the direction of the third strategy (i.e., using grey-or black-box modeling strategy), the most recent models are copula [26], neural-network based [27], ensemble methods [28], deep learning [29], and MC [30,31]. Moreover, a synthetic long-term dataset of CSI timeseries has been generated in [32] that is statistically indistinguishable from the observed data.…”