【摘要】 针对不同应用场景下光伏数据波动模式差异较大时，现有的光伏发电功率预测模型存在精度及适应性不足的问题，提出了一种具有权参数自适应性的并行深度学习光伏发电功率预测框架。该框架包含2种可以并行预测的深度学习算法单元(Attention-Seq2Seq单元、Transformer单元)以及一个权参数自适应优化单元。基于所提出的并行深度学习框架，对光伏发电功率进行预测，并分别与Attention-Seq2Seq、Transformer模型的预测结果进行了对比验证。结果表明：基于权参数优化的并行深度学习光伏功率预测框架弥补了不同数据波动模式下单一算法预测精度和适应性不足的问题，也可以有效解决时间序列预测中的长距离依赖问题，较单一算法预测精度更高，其平均绝对误差和均方根误差在夏季典型日最大降幅分别是41.18%和45.59%,在冬季典型日最大降幅分别是81.13%和82.86%。更多还原

【Abstract】 A parallel deep learning framework with adaptive optimization of weight parameters was proposed to address the accuracy and adaptability short comings of existing photovoltaic（PV） power prediction models when the data fluctuation patterns varied greatly in different application scenarios. The framework contained two parallel deep learning algorithm units such as Attention-Seq2Seq unit and Transformer unit and a weight parameter adaptive optimization unit. Based on the proposed parallel deep learning framework, the prediction of PV power generation was carried out and compared with the prediction results by Attention-Seq2Seq and Transformer models. Results show that the proposed framework not only can offset the lack of prediction accuracy and adaptability of single algorithm under different data fluctuation patterns, but also can effectively solve the long-distance dependence problem in time series prediction, which results in higher forecasting accuracy. The maximum reduction of average absolute error(E_MAE) and root mean square error(E_RMSE) is 41.18% and 45.59% in a typical day of summer, respectively, while the maximum reduction of E_MAE and E_RMSE arrives at 81.13% and 82.86%, respectively in a typical day of winter.更多还原