<p>Long-term time series forecasting is significant in many scenarios. Although various deep neural network models have been proposed in recent years, many of them lack the ability to detect the underlying mechanisms of large-scale time series generation. This limitation has resulted in bottlenecks in improving long-term prediction accuracy. Research indicates that one-to-many relationship and concept drift are essential features of large-scale sequences, and they are principal factors affecting long-term prediction. The one-to-many refers to a scenario in a time series where the observed values at different time points are identical, but the corresponding states differ. Concept drift refers to the change in the distribution of data over time. In this paper, it specifically refers to the variation in the probability distribution of the latent states corresponding to the observed sequence over time. Based on these two characteristics, a Concept Recursion and Generation-based model (CRG) is proposed. The CRG model utilizes Kullback–Leibler divergence and probability graphs to tackle the challenges of one-to-many relationships and concept drift, respectively. The goal is to accurately estimate the encoding of concepts within a sequence. To enhance the model’s extensibility, an encoder-decoder architecture is designed to implement the CRG model. The encoder estimates the distribution conditioned on the challenges posed by one-to-many relationships and concept drift, while the decoder estimates the distributions of future concepts and generates the target sequences. Extensive experimental results conducted on seven datasets demonstrate that the CRG model significantly outperforms baseline models in both efficiency and accuracy. Source code is publicly available at <a href="https://github.com/wc005/CRG">https://github.com/wc005/CRG</a>.</p>