Scholarlink

<p>Accurate forecasting of agricultural commodity prices is vital for maintaining market stability and supporting informed decision-making in production and policy planning. This study introduces the Frequency-domain Generative Price Diffusion (F-GPD) model, a novel approach that redefines agricultural price forecasting as a generative diffusion process operating in the frequency domain. Unlike traditional neural forecasting models that rely on pointwise regression losses, F-GPD captures both amplitude and phase dynamics of market signals through spectral noise injection and frequency-conditioned reconstruction. The model learns to generate coherent price trajectories by progressively denoising spectral representations, thereby aligning temporal patterns with their underlying frequency components. Experiments conducted on weekly potato price data from 25 Chinese provinces between 2012 and 2018 demonstrate that F-GPD consistently outperforms conventional baselines including CNN, LSTM, N-BEATS, Autoformer, and Informer across multiple evaluation metrics. The results confirm that frequency-domain diffusion effectively captures both short-term volatility and long-term periodicity, producing forecasts that are more accurate, stable, and physically interpretable. This work provides a new generative perspective for agricultural time series forecasting and offers a unified framework for modeling stochastic price dynamics under nonstationary 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

Frequency-Domain Generative Price Diffusion for Potato Time Series Forecasting in Agricultural Management

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Zenghua He,
  • Te Qi,
  • Xiangyu Liu,
  • Shuaichen Zhu,
  • Yuhan Yang

摘要

Accurate forecasting of agricultural commodity prices is vital for maintaining market stability and supporting informed decision-making in production and policy planning. This study introduces the Frequency-domain Generative Price Diffusion (F-GPD) model, a novel approach that redefines agricultural price forecasting as a generative diffusion process operating in the frequency domain. Unlike traditional neural forecasting models that rely on pointwise regression losses, F-GPD captures both amplitude and phase dynamics of market signals through spectral noise injection and frequency-conditioned reconstruction. The model learns to generate coherent price trajectories by progressively denoising spectral representations, thereby aligning temporal patterns with their underlying frequency components. Experiments conducted on weekly potato price data from 25 Chinese provinces between 2012 and 2018 demonstrate that F-GPD consistently outperforms conventional baselines including CNN, LSTM, N-BEATS, Autoformer, and Informer across multiple evaluation metrics. The results confirm that frequency-domain diffusion effectively captures both short-term volatility and long-term periodicity, producing forecasts that are more accurate, stable, and physically interpretable. This work provides a new generative perspective for agricultural time series forecasting and offers a unified framework for modeling stochastic price dynamics under nonstationary market conditions.