Multihorizon KPI Forecasting in Complex Industrial Processes: An Adaptive Encoder-Decoder Framework With Partial...

Area of Science:

• Industrial Engineering
• Data Science
• Process Control

Background:

• Key Performance Indicators (KPIs) are crucial for monitoring manufacturing quality and efficiency.
• Existing KPI forecasting methods struggle with multi-cycle predictions, hindering precise industrial process control.
• Effective advance prediction of KPIs is essential for optimizing complex manufacturing operations.

Purpose of the Study:

• To develop a novel framework for flexible multihorizon KPI forecasting.
• To address the limitations of current methods in predicting KPIs across multiple cycles.
• To improve the precision and timeliness of control in industrial processes through enhanced forecasting.

Main Methods:

• Proposed an adaptive encoder-decoder framework with a partial teacher forcing strategy (PTF-ED).
• Utilized an encoder with an attention layer for context vector generation from input time series.
• Designed a dual-decoder structure (delayed and current) with a partial teacher forcing strategy to mitigate exposure bias.
• Incorporated a weighted multihorizon forecasting constraint into the model training loss.

Main Results:

• The PTF-ED framework demonstrated effective flexible multihorizon KPI forecasting.
• The partial teacher forcing strategy efficiently utilized measured KPIs and addressed exposure bias.
• The weighted multihorizon constraint improved input-output correspondence across different sample intervals.
• Validation confirmed the model's effectiveness in numerical simulations and a real-world zinc flotation process.

Conclusions:

• The proposed PTF-ED framework offers a significant advancement in multihorizon KPI forecasting.
• This approach enables more precise and timely control in complex industrial environments.
• The method provides a robust solution for challenges in advance KPI prediction across multiple cycles.