Shift-Aware Hierarchical Attention Network for Ammonia Price Prediction

Forecasting ammonia prices is essential for strategic planning in agriculture, chemical production, and energy-related industries. However, ammonia prices exhibit substantial fluctuations and repeated shifts in statistical properties, challenging conventional forecasting models. To address this, we propose SHARP (Shift-aware Hierarchical Attention for Robust Prediction), which decomposes multivariate time series into variable-specific streams through a channel-independent mechanism and captures inter-variable and temporal dependencies through hierarchical attention. The model incorporates the DILATE loss function to mitigate temporal distortion and RevIN normalization to enhance robustness against distributional shifts. Empirical evaluations on real-world ammonia price data demonstrate SHARP reduces mean absolute error by up to 31% over leading baselines for 8- and 12-week forecasts. Additionally, analysis of attention weights identifies major contributing variables, highlighting the model’s explanatory capability. SHARP thus achieves accurate and interpretable forecasting for dynamically changing time series.