基于AM-LSTM的风电场内多点位风电功率预测

Multi-Point Wind Power Prediction in Wind Farms Based on Am-Lstm

为了提高大型风电场的风电功率预测的准确性,构建基于多点位的注意力机制(AM)和长短期记忆神经网络(LSTM)混合预测模型。对于风电功率影响因素特征提取模块,利用属性简约方法,在保证不影响原有的分类效果上,去掉多余属性,大幅度提高预测效率。对于风电功率预测模块,首先,利用15分钟数据间隔,不同高度的风速等作为模型输入,然后通过LSTM处理输入的时间序列与风电功率之间的非线性关系,接下来通过AM进一步优化LSTM的权重,最终得到风电功率的预测结果。通过国外某电场数据的验证:将数据集按照季度分开,各个模型预测结果最为优的是第四季度,对于基于多点位的注意力机制和长短期记忆混合预测模型的均方根误差和平均绝对误差分别比前馈神经网络预测模型降低48.8%和17.4%,比单一长短期记忆神经网络预测模型降低37.2%和7.8%。

In order to improve the accuracy of wind power prediction for large-scale wind farms, a hybrid prediction model based on multi-point attention mechanism(AM) and long short-term memory neural network(LSTM) was established. For the feature extraction module of wind power influencing factors, the method of attribute simplification was used to remove the redundant attributes and greatly improve the prediction efficiency without affecting the original classification effect. For the wind power prediction module, first of all, 15-minute data interval and wind speed at different heights were used as model inputs. Then, the nonlinear relationship between the input time series and wind power was processed through LSTM. Then, the weight of LSTM was further optimized through AM, and finally the prediction results of wind power were obtained. The validation of a foreign electric field data shows that the data sets are separated by quarter, and the most optimal prediction result is the fourth quarter;The root mean square error and mean absolute error of the multi-point based attention mechanism and the mixed prediction model of long-term and short-term memory are 48.8% and 17.4% lower than those of the feedforward neural network prediction model, and 37.2% and 7.8% lower than those of the single long-term and short-term memory neural network prediction model, respectively.