A Sequence‐to‐Sequence Approach for Short‐Term Temperature Prediction With Data Decomposition and Reconstruction

Temperature is an important factor affecting daily life, and accurate multi‐step temperature prediction can provide essential support for weather prediction, energy management, agricultural planning, and disaster mitigation. However, as the prediction horizon extends, the nonlinear nature of temperature data becomes more prominent, making it increasingly difficult to accurately model temporal variation patterns. Regarding this issue, this paper proposes a multi‐step temperature prediction system which comprises a multi‐frequency information extraction subsystem and a high‐precision prediction subsystem. The multi‐frequency information extraction subsystem employs a decomposition–clustering–reconstruction strategy to obtain high, medium, and low‐frequency data from the original temperature data. In the high‐precision prediction subsystem, we employ an encoders‐decoder architecture with an adaptive weighting mechanism to extract features from multi‐frequency data and produce prediction results. In addition, to further enhance prediction accuracy, time–frequency domain fusion is applied during training to capture differences between predictions and actual values from multiple angles. To evaluate the performance of the prediction system, three sets of comparative experiments are conducted on four datasets, demonstrating that the proposed system outperforms other benchmark models.