Dual-path one-dimensional convolutional BiLSTM with attention for silk cocoon price forecasting

Abstract

Accurately forecasting silk cocoon market prices is very crucial for enhancing decision-making silk cocoon supply chain. This work introduces a hybrid deep learning model, the dual-path one-dimensional convolutional neural network–bidirectional long short-term memory with lightweight attention mechanism (DPCBLA), which combines temporal sequence data with static features to predict daily maximum silk cocoon prices. The model captures short-term temporal patterns by using a one-dimensional convolutional neural network (1D CNN), then passes to a bidirectional long short-term memory (BiLSTM) network that captures long-term dependencies, and a lightweight attention mechanism that highlights key time steps for better interpretability. Static inputs such as market variety, seasonal indicators, and weather conditions are processed through a dense layer and integrated with temporal features in a dual-pathway design. The model was trained using a multi-source dataset obtained from the Ramanagara Government Silk Market in Karnataka, India. The dataset comprises historical transaction records, weather information, and seasonal disease data. The proposed model DPCBLA performed better than the traditional convolutional neural network (CNN), long short-term memory (LSTM), and BiLSTM models, achieving a mean absolute error (MAE)of24.73, a root mean square error (RMSE) of 33.93, a mean absolute percentage error (MAPE) of 4.05, a coefficient of determination (R2) of 0.94, and a prediction
accuracy of 95.95%.