Research on Hybrid Architecture Neural Networks for Time Series Prediction

In recent years, agricultural product price forecasting has played a crucial role in supply chain management and market decision-making. However, due to the influence of multiple factors such as supply-demand dynamics, seasonality, and weather conditions, agricultural product prices exhibit highly nonlinear and dynamic characteristics, making traditional methods often insufficient for accurate price change prediction. To address these challenges, this paper proposes a novel time series prediction model based on Temporal Convolutional Network (TCN), Multi-Layer Perceptron (MLP), and attention mechanism (referred to as “TCN-MLP-Attention”), using daily average price data from the U.S. avocado market from 2015 to 2023. The model aims to provide more accurate price trend forecasts for various stages of the supply chain. In the model architecture design, the TCN module utilizes causal convolution structure and dilated convolution to handle long-term temporal dependencies; the MLP module performs feature transformation and nonlinear mapping through a two-layer network; and the attention mechanism identifies important features at key time steps. This multi-module collaborative architecture effectively processes multi-scale features of time series data while providing model interpretability. Through comparative analysis of various optimization algorithms’ convergence performance and prediction accuracy, this study found that the AdamW optimizer, with its effective weight decay mechanism and adaptive learning rate, demonstrated superior performance in training stability and generalization capability, with MSE and R2 metrics outperforming traditional optimizers. Therefore, the AdamW optimizer with an adaptive learning rate adjustment strategy was adopted for model training. Experimental results show that the proposed TCN-MLP-Attention model exhibits better prediction performance compared to the baseline Transformer model: MSE decreased by approximately 1.24%, and R2 increased by about 0.17%, indicating stronger generalization ability in capturing time series features. Although MAE was slightly higher, the overall error control remained within an acceptable range, validating the model’s stability and effectiveness. The importance of the attention mechanism was further verified through ablation experiments, with the complete model’s R2 value improving by 87.04% and MSE decreasing by 87.18% compared to using TCN alone, confirming the effectiveness of the multi-module collaborative design. The model’s generalization capability is further validated through noise robustness tests and cross-product validation experiments. Even with medium-level noise ( $\sigma \approx 0.05$ ), the model maintains 77% of its R2 value, and when applied to strawberry price prediction using the dataset published by the United States Department of Agriculture Economic Research Service (USDA ERS)—a product with significantly different market characteristics—it achieves an R2 value of 0.7499 without any retraining, demonstrating strong adaptability to different data distributions. This research not only demonstrates the superiority of this hybrid architecture in handling complex time series prediction tasks, especially showing strong adaptability during market fluctuation periods, but also provides new research directions for deep learning applications in agricultural price prediction.