Abstract: To address the challenges of high complexity, information loss, and error accumulation in existing models based on Recurrent Neural Network (RNN) for time series forecasting tasks, along with the difficulty current forecasting models face in adapting to dynamically changing data lengths, a novel variable-length time series forecasting model based on segmented iteration, named PatchRNN was proposed. A subsequence segmentation approach is employed to divide time series data into multiple relatively short subsequence segments and Gated Recurrent units are utilized for iterative processing, effectively reducing the number of iterations required by RNN models and significantly improving forecasting accuracy. Meanwhile, through segmented iteration of these subsequence segments, the proposed model can achieve variable-length time series forecasting without requiring retraining, enhancing model’s adaptability for various applications. The experimental results demonstrate that, compared with several advanced baseline models including iTransformer and TimesNet, the proposed model maintains fewer parameters and achieves reductions in Mean Squared Error (MSE) by 7.3%-45.2% and Mean Absolute Error (MAE) by 3.3%-34.2% on long-term time series forecasting tasks across four datasets, including power transformer temperature and exchange rates, which confirms its superior forecasting accuracy and computational efficiency.

Key words: segmented iteration, variable-length time series forecasting, recurrent neural network, time series, deep learning

摘要： 针对现有基于循环神经网络（RNN）的模型在处理时间序列预测任务时存在复杂度高、信息遗忘、误差累积的问题以及现有预测模型难以匹配数据长度动态变化的问题，提出一种新的基于分段迭代的可变长时间序列预测模型PatchRNN。该模型使用子序列分割方法将时间序列分割为多个相对短的子序列片段，使用门控循环单元进行循环迭代，有效地降低了RNN模型的迭代次数，显著提高了预测准确性；同时通过子序列片段的分段迭代，在不重新训练模型的前提下所提出模型能实现可变长度的时间序列预测，提高了模型应用的普适性。实验结果表明，与iTransformer、TimesNet等多个先进基准模型相比，所提模型具有较少的参数量，在电力变压器温度、汇率等4个数据集的长时间序列预测任务的均方误差(MSE)减少了7.3%~45.2%，平均绝对误差(MAE)减少了3.3%~34.2%，表明所提出模型具有更高的预测精度和计算效率。

关键词: 分段迭代, 变长时间序列预测, 循环神经网络, 时间序列, 深度学习