考虑数据分布偏移的短期居民净负荷预测方法

Research on Short-term Residential Net Load Forecasting Method Considering Data Distribution Shift

新能源发电的高不确定性导致净负荷的数据分布偏移更加严重。数据分布偏移导致模型在历史数据中学习到的特征信息不再完全适用于未来数据,从而给净负荷预测(net load forecasting,NLF)带来了挑战。因此,考虑到净负荷中更严重的数据分布偏移问题,提出了一种基于不变风险最小化-不确定性加权-长短期记忆神经网络(long short-term memory neural network,LSTM)的短期居民净负荷预测方法,以提升净负荷预测精度。首先,通过不变风险最小化(invariant risk minimization,IRM)建立了一个双目标问题,包括准确预测和学习跨不同数据分布的不变特征。其次,通过长短期记忆神经网络(long short-term memory neural network,LSTM)处理时间序列数据的非线性特征。然后,通过基于不确定性加权(uncertainty weighting,UW)的目标平衡机制避免过度实现任一目标。此外,通过引入分位数回归将所提方法扩展到概率预测。最后,通过基于澳大利亚Ausgrid公司提供的真实居民电表数据从确定性预测结果、概率预测结果、不同数据分布偏移程度和不同光伏渗透率等多个维度验证了所提方法的有效性。

As new energy generation incurs more uncertainty,it leads to a more severe shift in net load data distribution.The data distribution shift means that the feature information learned by the model in historical data is no longer fully applicable to future data,thus posing a significant challenge to net load forecasting(NLF).Therefore,considering the data distribution shift problem in net load,this study proposes a short-term residential net load forecasting method based on IRMUW-LSTM to improve net load forecasting accuracy.First,a dual-objective problem was established using invariant risk minimization(IRM),which includes accurate forecasting and learning of invariant features across different data distributions.Second,a long short-term memory neural network(LSTM)was used to deal with the nonlinear features of the time series data.Subsequently,an uncertainty weighting(UW)-based objective balancing mechanism was used to avoid overachieving either objective.In addition,a quantile regression method was introduced to extend this study to probabilistic forecasting.Finally,the effectiveness of the proposed method was verified using multiple dimensions of deterministic and probabilistic prediction results,different data distribution shift levels,and different PV penetration rates based on real residential meter data provided by Ausgrid,Australia.