研究目的
To construct an accurate forecasting model for newly built PV power plants lacking sufficient historical data by transferring knowledge from historical solar irradiance data to PV output prediction using a long short-term memory neural network (LSTM).
研究成果
The proposed shared-optimized-layer LSTM architecture significantly improves prediction accuracy for PV power forecasting when historical data is insufficient. Transfer learning is particularly advantageous in scenarios with limited target domain data, though its benefits decrease as data volume increases.
研究不足
The experimental data is insufficient, and the evaluation and analysis of the predicted results are not carried out in various time periods of the year. Additionally, the data of the target domain have few stages of increase, limiting the ability to make a more complete description of the changes.
1：Experimental Design and Method Selection:
The study employs a transfer learning approach with LSTM neural networks, optimizing hyperparameters using Sequential Model-Based Global Optimization (SMBO) and pre-training weights on solar irradiance data before fine-tuning on PV output data.
2：Sample Selection and Data Sources:
Data from a PV power station in southwest China, including 6 months of historical solar irradiance data and 45 days of historical output data, with a sampling interval of 10 minutes.
3：List of Experimental Equipment and Materials:
Not explicitly mentioned in the provided text.
4：Experimental Procedures and Operational Workflow:
The LSTM network structure is optimized and pre-trained on solar irradiance data, then fine-tuned on PV output data. The model's performance is evaluated using absolute percentage error (APE), mean absolute percentage error (MAPE), and root mean square error (RMSE).
5：Data Analysis Methods:
The autocorrelation function (ACF) is used to characterize the autocorrelation of solar radiation and PV output. Performance metrics include APE, MAPE, and RMSE.
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
