基于TPE-LSTM的区域超短期风电功率预测

REGIONAL ULTRA-SHORT-TERM WIND POWER FORECASTING METHOD BASED ON TPE-LSTM

针对目前区域超短期风电功率预测精度较低且辅助信息不足的问题,提出一种结合贝叶斯优化和长短期记忆(LSTM)神经网络的预测方法。对历史风电功率数据进行数据修复和预处理并搭建LSTM网络模型;依据贝叶斯优化中的TPE算法对模型的超参数寻优,以获得更好的预测性能;为了验证所提出的TPE-LSTM模型的泛化能力,加入同样经过TPE算法优化的其他模型与其比较,同时加入误差校正环节降低LSTM算法在预测过程中存在的预测误差。实验结果表明,以区域历史风电功率数据为训练数据,该模型能够得到较高的预测精度。

Aiming at the problem of low accuracy and insufficient auxiliary information of regional ultra-short-term wind power prediction,this paper proposes a prediction method combining Bayesian optimization and long short-term memory(LSTM)neural network.The LSTM network model was built after repairing and preprocessing the historical wind power data.The tree-structured Parzen estimator(TPE)algorithm in Bayesian optimization was used to optimize the hyper-parameters of the model to achieve better prediction performance.Other existing models optimized by TPE algorithm were added for comparison in order to verify the generalization ability of TPE-LSTM model proposed in this paper.Besides,error correction was introduced to reduce the prediction error of LSTM algorithm in the prediction process.The experimental results show that this model can obtain higher prediction accuracy,with the regional historical wind power data as training data.