Leveraging scenario differences for cross-task generalization in water plant transfer machine learning models.

Area of Science:

• Environmental Engineering
• Computer Science
• Artificial Intelligence

Background:

• Machine learning (ML) models are vital for optimizing urban water systems and sustainability.
• Model transferability across different operational scenarios is limited by data variations, requiring extensive retraining.

Purpose of the Study:

• To develop a novel framework, the Environmental Information Adaptive Transfer Network (EIATN), to leverage scenario differences for improved ML model generalization in water systems.
• To demonstrate EIATN's effectiveness in enabling reuse of existing ML models across distinct prediction tasks within the same facility.

Main Methods:

• Evaluated the EIATN framework across four scenario categories and 16 ML architectures, resulting in 64 models.
• Utilized bidirectional long short-term memory (BiLSTM) as a top-performing architecture within the EIATN framework.
• Assessed performance using mean absolute percentage error (MAPE) and data volume requirements.

Main Results:

• The EIATN framework demonstrated feasibility, with BiLSTM achieving 3.8% MAPE using only 32.8% of typical data volume.
• In a Shenzhen case study, EIATN reduced carbon emissions by 40.8% compared to fine-tuning and 66.8% versus training from scratch.
• EIATN significantly enhances ML model generalization and reduces energy-intensive retraining needs.

Conclusions:

• EIATN effectively leverages scenario differences as prior knowledge for improved ML model generalization in urban water systems.
• This approach unlocks the reuse of existing ML models, leading to substantial energy savings and promoting low-carbon intelligent water management.
• EIATN fosters equitable and sustainable operations in urban water infrastructure.