Research on Agricultural Product Price Fluctuation Prediction Based on Time Series Analysis

Abstract

To address the limitation of traditional LSTM models in capturing the spatiotemporal correlation characteristics of agricultural product price fluctuations, this paper proposes an improved spatiotemporal fusion long short-term memory network (ST-LSTM). By adding a spatiotemporal attention mechanism and a cross-regional feature fusion module, this model achieves the joint extraction of "long-term temporal dependencies and spatial correlation features." The spatiotemporal attention module dynamically weights key spatiotemporal node information, while the cross-regional module uses graph convolution to model regional price transmission relationships. The improved LSTM core module introduces residual connections and leaky ReLU to mitigate gradient vanishing. Experiments based on daily data for six agricultural products (wheat, rice, etc.) from 2018 to 2023 show that the ST-LSTM algorithm achieves an average MAE of 0.12 yuan/kg, RMSE of 0.18 yuan/kg, and a MAPE of 3.2%, which are 33.3% lower than the MAPE of the traditional LSTM algorithm and 61.4% lower than that of the ARIMA algorithm. Ablation experiments show that without spatiotemporal attention or cross-regional fusion, the MAPE increases to 4.1% and 4.3%, respectively, an increase of over 28%. In predicting cucumber prices during the 2022 drought, the ST-LSTM algorithm achieved a MAPE of 4.5%, outperforming both the traditional LSTM algorithm (6.8%) and the ARIMA algorithm (10.2%). This demonstrates the algorithm's accuracy and robustness in both normal and abnormal scenarios, providing an effective solution for agricultural product price forecasting.