Scholarlink

<p>Forecasting agricultural commodity prices is critical for ensuring food security, stabilizing supply chains, and informing data-driven policy decisions in agricultural management. However, the volatile and nonstationary nature of crop prices, especially in settings lacking auxiliary signals, poses a significant challenge to conventional forecasting methods. In this study, we introduce PatchTST-DME, a frequency-aware deep forecasting framework tailored for univariate price sequences. The model incorporates a Spectral Residual Enhancement module to amplify anomaly-related frequency signals, a Frequency-Aware Attention mechanism to dynamically weight spectral bands, and a Dynamic Mixture-of-Experts architecture to disentangle long-term trends from short-term shocks via learnable routing. Additionally, we integrate a Fourier-domain contrastive learning objective to improve generalization under distributional shifts. Experiments on a large-scale, multi-year dataset of provincial potato prices across China demonstrate that PatchTST-DME achieves substantial improvements over state-of-the-art baselines in MAE, RMSE, and MAPE. Beyond accuracy, the model yields interpretable frequency-level insights that support practical agricultural decision-making in volatile market conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

PatchTST-DME: Frequency-Aware Transformer for Potato Price Forecasting in Agricultural Management

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • An Zhang,
  • Sheng Chen,
  • Chao Wu,
  • Guiyuan Li,
  • Li Liu

摘要

Forecasting agricultural commodity prices is critical for ensuring food security, stabilizing supply chains, and informing data-driven policy decisions in agricultural management. However, the volatile and nonstationary nature of crop prices, especially in settings lacking auxiliary signals, poses a significant challenge to conventional forecasting methods. In this study, we introduce PatchTST-DME, a frequency-aware deep forecasting framework tailored for univariate price sequences. The model incorporates a Spectral Residual Enhancement module to amplify anomaly-related frequency signals, a Frequency-Aware Attention mechanism to dynamically weight spectral bands, and a Dynamic Mixture-of-Experts architecture to disentangle long-term trends from short-term shocks via learnable routing. Additionally, we integrate a Fourier-domain contrastive learning objective to improve generalization under distributional shifts. Experiments on a large-scale, multi-year dataset of provincial potato prices across China demonstrate that PatchTST-DME achieves substantial improvements over state-of-the-art baselines in MAE, RMSE, and MAPE. Beyond accuracy, the model yields interpretable frequency-level insights that support practical agricultural decision-making in volatile market conditions.