Federated Machine Learning Enables Risk Management and Privacy Protection in Water Quality.

Area of Science:

• Environmental Engineering
• Data Science
• Machine Learning

Background:

• Real-time water quality management in wastewater treatment plants (WWTPs) is crucial but hindered by data privacy concerns, limiting effective data sharing.
• Current data sharing practices are insufficient, often remaining theoretical due to the inability to access sensitive raw data.

Purpose of the Study:

• To develop a novel framework for federated learning (FL) that facilitates secure data sharing for water quality risk management in WWTPs.
• To enhance the performance of machine learning models for effluent quality prediction through privacy-preserving parameter sharing.

Main Methods:

• An adaptive water system federated averaging (AWSFA) framework was developed, utilizing federated learning (FL) principles.
• Ten machine learning models were trained for effluent indicators using data from six WWTPs (2018-2024), with a bidirectional long-term memory network (BM) as the baseline.
• AWSFA was compared against direct training and classical federated averaging (FedAvg) in terms of mean absolute percentage error (MAPE).

Main Results:

• AWSFA significantly reduced the mean absolute percentage error (MAPE) of the bidirectional long-term memory network (BM) compared to direct training and FedAvg.
• Performance gains were attributed to parameter sharing for data sharing, not algorithmic complexity.
• The model demonstrated robustness, maintaining performance even with 50% of key features missing during simulated water quality disturbances.

Conclusions:

• AWSFA offers a viable solution for data sharing and privacy preservation in water systems.
• The framework provides theoretical support for the digital transformation of WWTPs in the big data and big model era.
• This approach enables effective real-time water quality risk management through secure, collaborative model training.