基于特征构造和改进PSO算法的分布式光伏功率预测

Distributed photovoltaic power prediction based on feature construction and improved PSO algorithm

分布式光伏由于其历史数据缺乏,光伏出力的预测精度不高,提出改进粒子群优化算法(PSO)+长短时记忆网络(LSTM)与注意力机制结合的神经网络模型。构造基于聚类算法的特征工程扩充数据集;给出局部最优判据改进粒子群算法并应用于模型的超参数优化,提升模型泛化性;采用注意力机制与LSTM相结合的架构进行短期功率预测。在澳大利亚公开数据集上的实验表明,新的特征工程与光伏出力具有相关性,预测精度相比传统LSTM模型精度提高17.4%,且改进PSO算法相比标准算法收敛性更好。

Due to the lack of historical data of distributed PV,the prediction accuracy of PV output is not high.The improved neural network model combining PSO,LSTM and attention mechanism was proposed.The extended data set of feature engineering was construct based on clustering algorithm.A local optimal criterion was given to improve the particle swarm optimization algorithm and it was applied to the model's hyperparametric optimization,so as to improve the generalization of the model.The architecture combining attention mechanism and LSTM was used for short-term power prediction.The experiment on the Australian open data set shows that the new feature engineering is related to the PV output,the prediction accuracy is 17.4%higher than that of the traditional LSTM model,and the improved PSO algorithm has better convergence than the standard algorithm.