Research on time series prediction of multi-process based on deep learning.

Area of Science:

• Industrial Engineering
• Artificial Intelligence
• Data Science

Background:

• Data fluctuation in multi-process production poses significant forecasting challenges.
• Existing time series models often struggle to adapt to dynamic production environments.

Purpose of the Study:

• To develop an advanced time series forecasting model to address data fluctuation in multi-process production.
• To improve prediction accuracy by correcting GRU model outputs using a novel deep reinforcement learning approach.

Main Methods:

• A Soft Update Dueling Double Deep Q-learning (SU-D3QN) algorithm was integrated with a Gate Recurrent Unit (GRU) network, forming the SU-D3QN-G model.
• The SU-D3QN algorithm learns to add bias to GRU predictions, minimizing absolute error.
• Experiments were conducted on real-world production data (outlet temperature, moisture, etc.) using training, inspection, and test sets.

Main Results:

• The SU-D3QN-G model demonstrated substantial improvements over GRU, LSTM, and ARIMA models.
• Significant reductions in Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) were observed.
• MSE reductions ranged from 0.846-23.930% compared to GRU, 5.132-36.920% vs. LSTM, and 10.606-70.714% vs. ARIMA.

Conclusions:

• The proposed SU-D3QN-G model effectively mitigates data fluctuation issues in multi-process production forecasting.
• The integration of deep reinforcement learning with recurrent neural networks offers a powerful approach for enhancing time series prediction accuracy.