Accurate prediction of soybean demand is of profound practical significance for safeguarding national food security, optimizing industrial decision-making, and responding to fluctuations in international trade. Traditional soybean demand forecasting methods are plagued by inadequacies such as limited capacity to excavate data dimensionality and multivariate interactive features, insufficient ability to capture nonlinear relationships under the coupling of multi-dimensional dynamic factors, and challenges in model interpretability and domain adaptability. These limitations render them incapable of effectively supporting accurate prediction and interpretable analysis of China's soybean demand. When the temporal fusion transformers (TFT) model is applied to forecast China's soybean demand, it exhibits certain constraints in aspects like feature interaction layers and attention weight allocation. Consequently, there is an urgent need to explore a forecasting method based on the improved TFT model to enhance the accuracy and interpretability of soybean demand prediction.

Drawing on relevant studies, this research applied the deep learning-based TFT model to China's soybean demand forecasting and proposed the MA-TFT (improved TFT model based on MDFI and AAWO) model, which was enhanced through multi-layer dynamic feature interaction (MDFI) and adaptive attention weight optimization (AAWO). Firstly, a dataset for analyzing China's soybean demand, covering eight dimensions: consumption, production, trade, inventory, market, economy, policy, and international factors, was collated. This dataset, encompassing 4652 relevant indicators spanning from 1980 to 2024, was subjected to data cleaning, transformation, augmentation, and feature engineering. The training, validation, and test sets for the model were constructed using the rolling window method. Secondly, based on the architecture of the TFT model for China's soybean demand forecasting, a multi-layer dynamic feature interaction module and an adaptive attention weight optimization strategy were designed. Additionally, the model's loss function, training strategy, and Bayesian hyperparameter tuning method were formulated, and the model performance evaluation metrics were determined. Subsequently, experiments were designed to compare the prediction performance of the MA-TFT model with that of the autoregressive integrated moving average model (ARIMA), the long short-term memory (LSTM) model, and the original TFT model. Ablation experiments on the MDFI and AAWO modules were conducted separately. The SHapley Additive exPlanations (SHAP) tool was employed for interpretability analysis to identify key feature variables influencing China's soybean demand and their interaction relationships. Error analysis was performed between the predicted and actual values of China's historical soybean demand, and a comparative analysis of the predicted soybean demand in China from 2025 to 2034 was carried out.

The mean squared error (MSE) and mean absolute percentage error (MAPE) of the MA-TFT model were 0.036 and 5.89%, respectively, with a coefficient of determination R² of 0.91, all of which outperformed those of the comparative models, namely ARIMA (1, 1, 1), LSTM, and TFT. Compared with the benchmark TFT model, the root mean square error (RMSE) and MAPE of the MA-TFT model decreased cumulatively by 21.84% and 3.44%, respectively. These results indicated that the MA-TFT model, as an improved version of TFT, could capture complex relationships between features and enhance prediction performance and accuracy. Interpretability analysis using the SHAP tool revealed that the MA-TFT model exhibited high stability in explaining key feature variables affecting China's soybean demand. It was projected that China's soybean demand would reach 117.99 million tons, 110.33 million tons, and 113.78 million tons in 2025, 2030, and 2034, respectively.

The MA-TFT model, developed by improving the TFT model, provides an innovative solution to address the practical issues of insufficient accuracy and poor interpretability in existing soybean demand forecasting methods. It also offers valuable references for method optimization and application in time series forecasting of other bulk agricultural products.

The fluctuations in the supply, consumption, and prices of agricultural products directly affect market monitoring and early warning systems. With the ongoing transformation of China's agricultural production methods and market system, advancements in data acquisition technologies have led to an explosive growth in agricultural data. However, the complexity of the data, the narrow applicability of existing models, and their limited adaptability still present significant challenges in monitoring and forecasting the interlinked dynamics of multiple agricultural products. The efficient and accurate forecasting of agricultural market trends is critical for timely policy interventions and disaster management, particularly in a country with a rapidly changing agricultural landscape like China. Consequently, there is a pressing need to develop deep learning models that are tailored to the unique characteristics of Chinese agricultural data. These models should enhance the monitoring and early warning capabilities of agricultural markets, thus enabling precise decision-making and effective emergency responses.

An integrated forecasting methodology was proposed based on deep learning techniques, leveraging multi-dimensional agricultural data resources from China. The research introduced several models tailored to different aspects of agricultural market forecasting. For production prediction, a generative adversarial network and residual network collaborative model (GAN-ResNet) was employed. For consumption forecasting, a variational autoencoder and ridge regression (VAE-Ridge) model was used, while price prediction was handled by an Adaptive-Transformer model. A key feature of the study was the adoption of an "offline computing and visualization separation" strategy within the Chinese agricultural monitoring and early warning system (CAMES). This strategy ensures that model training and inference are performed offline, with the results transmitted to the front-end system for visualization using lightweight tools such as ECharts. This approach balances computational complexity with the need for real-time early warnings, allowing for more efficient resource allocation and faster response times. The corn, tomato, and live pig market data used in this study covered production, consumption and price data from 1980 to 2023, providing comprehensive data support for model training.

The deep learning models proposed in this study significantly enhanced the forecasting accuracy for various agricultural products. For instance, the GAN-ResNet model, when used to predict maize yield at the county level, achieved a mean absolute percentage error (MAPE) of 6.58%. The VAE-Ridge model, applied to pig consumption forecasting, achieved a MAPE of 6.28%, while the Adaptive-Transformer model, used for tomato price prediction, results in a MAPE of 2.25%. These results highlighted the effectiveness of deep learning models in handling complex, nonlinear relationships inherent in agricultural data. Additionally, the models demonstrate notable robustness and adaptability when confronted with challenges such as sparse data, seasonal market fluctuations, and heterogeneous data sources. The GAN-ResNet model excels in capturing the nonlinear fluctuations in production data, particularly in response to external factors such as climate conditions. Its capacity to integrate data from diverse sources—including weather data and historical yield data—made it highly effective for production forecasting, especially in regions with varying climatic conditions. The VAE-Ridge model addressed the issue of data sparsity, particularly in the context of consumption data, and provided valuable insights into the underlying relationships between market demand, macroeconomic factors, and seasonal fluctuations. Finally, the Adaptive-Transformer model stand out in price prediction, with its ability to capture both short-term price fluctuations and long-term price trends, even under extreme market conditions.

This study presents a comprehensive deep learning-based forecasting approach for agricultural market monitoring and early warning. The integration of multiple models for production, consumption, and price prediction provides a systematic, effective, and scalable tool for supporting agricultural decision-making. The proposed models demonstrate excellent performance in handling the nonlinearities and seasonal fluctuations characteristic of agricultural markets. Furthermore, the models' ability to process and integrate heterogeneous data sources enhances their predictive power and makes them highly suitable for application in real-world agricultural monitoring systems. Future research will focus on optimizing model parameters, enhancing model adaptability, and expanding the system to incorporate additional agricultural products and more complex market conditions. These improvements will help increase the stability and practical applicability of the system, thus further enhancing its potential for real-time market monitoring and early warning capabilities.