研究目的
To improve the accuracy of energy yield forecasts for photovoltaic systems, especially in grid areas of high photovoltaic shares, by presenting a uni?ed methodology for hourly-averaged day-ahead photovoltaic power forecasts based on data-driven machine learning techniques and statistical post-processing.
研究成果
The proposed location and setup independent day-ahead PV power forecasting model can form an important part of grid management processes of grid managers in shaping more cost-e?ective, stable and reliable power systems. in parallel, it further enables the optimal participation of PV power plant operators and aggregators to the energy market.
研究不足
The technical and application constraints of the experiments, as well as potential areas for optimization, are not explicitly mentioned in the paper.
1：Experimental Design and Method Selection
The methodology comprised of a data quality stage, data-driven power output machine learning model development (arti?cial neural networks), weather clustering assessment (K-means clustering), post-processing output optimisation (linear regressive correction method) and the ?nal performance accuracy evaluation.
2：Sample Selection and Data Sources
The development of the day-ahead PV power production forecasting model was performed by utilising a 1-year dataset of PV system electrical and meteorological measurements that were acquired from a system installed at the OTF of the PV Technology Laboratory at the University of Cyprus (UCY), in Cyprus. Additionally, the location and setup performance independent behaviour of the developed day-ahead PV power production forecasting model was validated by employing datasets from a PV system located in Albuquerque, NM USA, administered by Sandia National Laboratories.
3：List of Experimental Equipment and Materials
The test-bench PV system comprises of 5 poly-crystalline Silicon (poly-c-Si) PV modules (rated at 235 Wp each as depicted from the manufacturer’s datasheet) that were connected in series to form a string of nominal power capacity 1.175 kWp at the input of a string inverter. The PV system was installed in an open-?eld mounting arrangement due South and at the optimum annual energy yield inclination angle of 27.5°.
4：Experimental Procedures and Operational Workflow
The methodology followed for the development of the optimal hourly-averaged day-ahead PV power production forecasting model is presented in Fig. 1. The methodology comprised of data quality assessment (to ensure data integrity and validity), data-driven power output model development (for training/validation and testing), weather clustering assessment (based on clustering of daily irradiance patterns), post-processing performance optimisation, and ?nal forecasting performance evaluation.
5：Data Analysis Methods
The prediction performance accuracy was assessed based on several prede?ned metrics. The mean absolute percentage error (MAPE), the nRMSE which is the relative RMSE normalised to the nominal capacity of the PV system, and the skill score (SS) which describes the accuracy and degree of association of the network’s output to a baseline model.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
