A novel stepwise regression method for predicting September Pan-Arctic sea-ice extent: Comparison with long short-term memory neural networks

Accurately forecasting the Pan-Arctic sea-ice extent (SIE) in September is crucial for understanding climate change impacts and ensuring safe Arctic navigation. This study introduces the SWR DY-method, a novel prediction approach that combines the stepwise regression (SWR) and interannual increment methods and applies it to predictive performance. The model identifies grid points significantly correlated to September SIE using the temporal correlation coefficients between various monthly climate variables and the Pan-Arctic SIE in September. A database of potential predictors was established, from which SWR and long short-term memory (LSTM) neural networks were used to develop single-predictor models with monthly initializations from January to September. These single-predictor models were then ensemble-averaged to create multi-predictor ensemble prediction models. Evaluation of the models from 2014 to 2022 was carried out, comparing the performances of the SWR DY-method and LSTM DY-method. Results indicated that the SWR DY-method had higher single-predictor prediction skill than the LSTM DY-method. Multi-predictor models using the SWR DY-method demonstrated lower mean absolute errors and higher predictive skill for January to September initializations, outperforming the LSTM and the SIO (Sea Ice Outlook) median predictions. This study presents a new and effective strategy for improving seasonal predictions of Pan-Arctic SIE. 精准预测9月北极海冰范围对理解气候变化的影响与保障北极航道安全至关重要.基于逐步回归和年际增量预测方法(SWR-DY), 本文研制了一种9月北极海冰范围预测新模型.该模型基于各月气候变量与9月北极海冰范围的时间相关系数, 筛选具有统计显著性的预测因子, 建立潜在预测因子库.在此基础上, 分别采用SWR和长短期记忆神经网络(LSTM)构建1−9月逐月起报的单因子预测模型, 并通过集合平均得到多因子集合预测模型.本文通过对比了不同模型对2014−2022年9月北极海冰范围的预测效能.结果表明:SWR DY方法的单因子预测能力普遍优于LSTM DY方法;其多因子模型在1−9月起报时均表现出更低的平均绝对误差和更高的预测能力.新方法对2014−2022年9月北极海冰范围的预测效能也高于国际海冰预测网络活动多模型预测结果中位数的预测效能.本研究为9月北极海冰范围预测提供了新方法.