Explainable artificial intelligence for reliable water demand forecasting to increase trust in predictions.

Area of Science:

• Water resource management
• Artificial Intelligence (AI) regulation
• Machine Learning (ML)

Background:

• The EU Artificial Intelligence Act categorizes AI in water supply systems as high-risk due to potential impacts on infrastructure and reliability.
• AI applications like water demand forecasting for automatic tank operation must meet stringent requirements for transparency and robustness.

Purpose of the Study:

• To systematically evaluate the accuracy, transparency, and robustness of various machine learning models for AI-driven water demand forecasting.
• To assess compliance with the EU Artificial Intelligence Act's requirements for high-risk AI in water supply systems.

Main Methods:

• Applied six established machine learning models (transparent and opaque) to diverse datasets for daily water demand forecasting.
• Systematically evaluated model performance against accuracy, transparency (traceability, explainability), and technical robustness criteria.

Main Results:

• Opaque models generally offer higher prediction accuracy by capturing complex relationships (e.g., weather data) but are more susceptible to input forecast errors.
• Transparent models, primarily using historical demand data, demonstrate greater robustness against data irregularities.
• Both model types can achieve explainability, but differ in transparency levels and robustness to input errors.

Conclusions:

• The choice between transparent and opaque AI models for water demand forecasting depends on balancing prediction accuracy with robustness to data variations.
• Model selection should consider operator preferences and specific application contexts within the regulatory framework of the EU Artificial Intelligence Act.