Adaptive Neuro-Fuzzy System Offers Explainable Uncertainty for Wastewater Energy Forecasts

Global: Adaptive Neuro-Fuzzy System Offers Explainable Uncertainty for Wastewater Energy Forecasts

Researchers at Melbourne Water have introduced an Interval Type-2 Adaptive Neuro-Fuzzy Inference System (IT2-ANFIS) to improve short-term electricity demand forecasting at the Eastern Treatment Plant, a study first posted to arXiv in January 2026. The system aims to provide both accurate point predictions and interpretable uncertainty intervals, addressing the need for risk-aware operational decisions in a sector that accounts for roughly 1-3 % of global electricity consumption.

Energy Consumption Context

Wastewater treatment facilities worldwide consume a significant share of electrical power, making precise load forecasting essential for cost reduction and sustainability goals. Traditional forecasting approaches often focus solely on minimizing error metrics without quantifying the confidence of each prediction.

Limitations of Existing Machine-Learning Models

Conventional machine-learning models, including deep neural networks and standard fuzzy inference systems, typically generate single-value forecasts and rely on black-box probabilistic techniques to estimate uncertainty. Such methods can obscure the sources of ambiguity, limiting their usefulness for safety-critical infrastructure management.

IT2-ANFIS Architecture and Uncertainty Decomposition

The proposed IT2-ANFIS extends the first-order ANFIS by incorporating interval type-2 fuzzy sets, enabling three distinct levels of uncertainty analysis: feature-level identification of ambiguous inputs, rule-level assessment of local model confidence, and instance-level prediction intervals that capture overall forecast uncertainty.

Validation Results

Using a dataset collected from Melbourne Water’s Eastern Treatment Plant, the IT2-ANFIS achieved predictive accuracy comparable to the baseline first-order ANFIS while exhibiting a markedly lower variance across multiple training runs. The system consistently produced narrower and more stable prediction intervals.

Explainability and Operational Insight

By linking interval widths to specific input variables, the framework allows operators to pinpoint which operational conditions—such as influent flow rate or temperature—contribute most to forecast uncertainty. This transparency supports more informed scheduling and load-balancing decisions.

Implications and Future Directions

The study demonstrates that explainable uncertainty quantification can be integrated into existing plant control workflows without sacrificing accuracy. Researchers suggest that the approach could be adapted to other utility sectors, including water distribution and power grid management, to enhance resilience against demand fluctuations.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.