Multiscale temporal modeling of agricultural economic cycle fluctuations and early warning

Predicting fluctuations in the agricultural economic cycle provides a basis for macroeconomic regulation, with accurate predictions and timely early warnings supporting food security and sustainable agricultural development. However, agricultural economic time series data exhibit non-stationarity, multi-cycle superposition, and noise interference, making it difficult for traditional methods to capture complex dynamic patterns. This article proposes a multiscale temporal modeling approach to predict and issue early warnings for the agricultural economic cycle. The method includes three core components: (1) a multiscale feature extraction method coupling Variational Mode Decomposition (VMD) with wavelet transform to decompose agricultural economic sequences into different frequency components and extract time-frequency features; (2) a multiscale feature fusion attention module with intra-scale and cross-scale attention mechanisms to learn dependencies among multiscale features; (3) a joint optimization network architecture for prediction and early warning through a shared encoder and dual-task decoder. Experiments on the Chinese agricultural product price dataset and the FAO agricultural production dataset show that the proposed method reduces MAE by 19.1% on average and improves R 2 by 5.2% compared to baseline methods.