研究目的
To improve the accuracy of short-term photovoltaic power forecasting by exploiting the capacity of feature learning and nonlinear fitting implied in deep structure, using a model based on a deep convolutional network (DCN).
研究成果
The proposed deep-based forecasting model combined with multi-dimensional power data achieves the desired goals overall, with a better performance compared to several advanced models. However, there are several open problems regarding the proposal that need to be addressed in future studies.
研究不足
The model has room for improvement against the instability caused by weather and other factors. The impact from pre-processing power tensor, power matrices with multiple sizes, and more meteorological elements are not considered.
1：Experimental Design and Method Selection:
The study proposes a DCN-based forecasting model that combines feature learning with multi-dimensional power arrays. The model is designed to exploit historical daily data to build a power tensor for improved forecast performance.
2：Sample Selection and Data Sources:
The power data used for evaluation comes from four PV farms of the ELIA, covering the period from January 2013 to December 2017 with a resolution of 15 minutes. Meteorological elements from Brussels, Arlon, Bruges, and Maastricht are selected as data sources.
3：List of Experimental Equipment and Materials:
Not explicitly mentioned in the paper.
4：Experimental Procedures and Operational Workflow:
The model consists of three modules: input layer for initial splicing of the power tensor and the forecasted day’s MEV, feature learning module with convolutional layers, and feature transformation module with hidden layers.
5：Data Analysis Methods:
The performance of the proposed model is evaluated using MAE and RMSE criteria.
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
