研究目的
Investigating probabilistic forecasting schemes of day-ahead photovoltaic (PV) generations with the auto-regressive recurrent neural network model named DeepAR, and evaluating their performance based on the normalized residues.
研究成果
The study demonstrates that probabilistic forecasting of PV generations using the DeepAR model with local weather forecast data provides tighter prediction intervals and higher reliability, as indicated by normalized residues close to the standard normal distribution. Future work could focus on quantitative statistical methods for comparing distributions and analyzing the contribution of individual weather components to forecast accuracy.
研究不足
The study does not quantitatively compare the empirical distribution of standard scores with the normal distribution using statistical methods. Additionally, not all weather components may contribute equally to PV forecast accuracy, suggesting the need for feature importance analysis.
1：Experimental Design and Method Selection:
The study employs the DeepAR model, an auto-regressive RNN model, for probabilistic forecasting of PV generations. The model is trained using historical PV data and optional weather forecast data.
2：Sample Selection and Data Sources:
Hourly historical PV generation data from Hadong, Korea, and 3-hourly weather forecast data including temperature, precipitation type, probability of precipitation, humidity, wind speed, and cloud amount are used.
3：List of Experimental Equipment and Materials:
The study utilizes the DeepAR model developed by Amazon Web Service (AWS).
4：Experimental Procedures and Operational Workflow:
The model generates multiple sample traces of PV generations for the next day based on historical data and optional weather forecasts. The tightness and reliability of the forecast are evaluated using standard deviation and standard score.
5：Data Analysis Methods:
The performance is evaluated using root mean square error (RMSE) and mean absolute error (MAE), and the distribution of standard scores is compared to the standard normal distribution.
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
