研究目的
To develop and apply artificial neural network (ANN) models for estimating and forecasting solar radiation (GHI and DNI) from meteorological data to support energy system integration and management.
研究成果
ANN models are effective for estimating 5-min GHI (nRMSE 19.35% with 6 inputs) and converting GHI to GTI (nRMSE around 8%). For forecasting, hourly GHI and DNI can be predicted with nRMSE ranging from 22.57% to 34.85% for GHI and 38.23% to 61.88% for DNI from h+1 to h+6. ANN outperforms naive persistence models, especially for longer horizons. The methods support solar energy integration by providing reliable radiation data where measurements are sparse.
研究不足
The ANN models are black-box and may lack physical interpretability. Performance depends on data quality and station characteristics; results may not generalize to other locations. Forecasting accuracy decreases with longer time horizons, especially for DNI due to its higher variability. The study uses specific datasets from two stations, limiting broader applicability.
1：Experimental Design and Method Selection:
The study uses multilayer perceptron (MLP) artificial neural networks with feed-forward back-propagation for estimation and forecasting. The Levenberg-Marquardt learning algorithm is employed, and k-fold cross-validation (k=10) is used for robustness.
2：Sample Selection and Data Sources:
Data from two meteorological stations are used: Bouzareah, Algeria (for GHI and GTI estimation, 5-min data from April 2011 to April 2013) and Odeillo, France (for GHI and DNI forecasting, hourly data over two years). Data include measured parameters like temperature, humidity, wind speed, and calculated astronomical data.
3：List of Experimental Equipment and Materials:
Pyranometers for measuring solar irradiance, meteorological sensors for temperature, humidity, wind speed, etc. Specific models and brands are not detailed in the paper.
4：Experimental Procedures and Operational Workflow:
Data preprocessing includes cleaning, filtering outliers, removing data during sunrise/sunset, and transforming to stationary series using clear sky indices. For estimation, various input combinations (up to 1023 for GHI) are tested. For forecasting, past data are used to predict future values with horizons from h+1 to h+
5：Data Analysis Methods:
Statistical error metrics are used: mean absolute error (MAE), root mean square error (RMSE), mean bias error (MBE), and their normalized versions (nMAE, nRMSE). Pearson correlation coefficients are calculated for input selection.
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
