Abstract: Garlic is one of China’s major agricultural products and export commodities. Accurate early warning of garlic price fluctuations is of great significance for safeguarding farmers’ income and maintaining industrial stability. Given that the garlic price time series exhibits nonlinear, non-stationary, and “sharp-peaked and heavy-tailed” distribution characteristics, traditional early warning methods often fail to effectively capture complex temporal features, resulting in limited warning accuracy. To address these issues, this study constructs a price alert indicator system and a scientific warning threshold division method, and proposes a VMD-TCAN-based garlic price forecasting and graded early warning model to achieve precise prediction and automatic warning.first, to mitigate the high noise and non-stationarity of the original price series, Variational Mode Decomposition (VMD) is introduced. Unlike Empirical Mode Decomposition (EMD), which suffers from mode mixing and end effects, VMD iteratively searches for the optimal solution of variational modes and adaptively decomposes the signal into intrinsic mode functions (IMFs) with specific sparsity and finite bandwidth. To overcome the subjectivity in VMD parameter selection, a Particle Swarm Optimization (PSO) algorithm is employed, using minimum fuzzy entropy as the fitness function to dynamically determine the optimal number of modes and penalty factor. The complex garlic price series is thereby decomposed into eight IMFs with distinct center frequencies, effectively separating multi-scale components such as long-term trends, seasonal fluctuations, and short-term random shocks.second, a multi-source information fusion TCAN prediction model is developed. Six key external influencing factors covering the production side (planting area, planting cost), circulation side (inventory, export volume), and market side (exchange rate, CPI) are selected. Together with the VMD-derived IMF components and daily garlic prices, these variables form a high-dimensional input tensor. Based on the TCAN architecture, dilated causal convolution is utilized to efficiently extract long-range historical dependency features through exponentially expanding receptive fields while preserving temporal causality. A self-attention mechanism is further introduced to globally perceive multi-source inputs, dynamically assigning correlation weights to focus on critical time steps that significantly influence price movements. This design substantially enhances the model’s capability to capture abnormal price volatility intensity and trend turning points under complex market dynamics.finally, an integrated “forecasting–early warning” framework is established. The Shapiro–Wilk test reveals that garlic price volatility follows a non-normal, sharp-peaked and heavy-tailed distribution, rendering the traditional 3\sigma rule inapplicable. Therefore, a quantile-based method grounded in the historical volatility distribution is adopted to scientifically define five warning levels: severe negative warning, mild negative warning, no warning, mild positive warning, and severe positive warning. Continuous numerical forecasts are thus mapped into discrete risk-level signals.using daily price data from the Jinxiang production area spanning 2005 to 2024, experimental results demonstrate that the proposed VMD-TCAN model achieves a root mean square error (RMSE) of 0.042 and a coefficient of determination (R²) of 0.998, significantly outperforming benchmark models and their corresponding VMD-based hybrid versions. In terms of early warning performance, the accuracy and F1-score improve by approximately 4.86% and 4.97%, respectively, compared with the TCN model. The proposed model enables more precise identification of abnormal fluctuations in the garlic market, providing timely and reliable price forecasts and warning information for farmers and related enterprises, and offering decision support for governmental macro-regulation.