A novel nowcasting (estimation) model based on an adaptive network neutrosophic hesitant fuzzy inference system (ANNHFIS): a case study of Istanbul – Scientific Reports
Biomass power plants are often viewed as cleaner alternatives to fossil-fuel facilities, yet they still emit nitrogen dioxide (NO₂)—a pollutant linked to respiratory issues and urban air quality degradation. For cities like Istanbul, where industrial activity intersects with dense populations, accurately estimating (nowcasting) NO₂ levels around such plants can guide smarter health advisories, regulatory oversight, and operational decisions. A new study introduces an adaptive network-based neutrosophic hesitant fuzzy inference system optimized with particle swarm optimization (ANNHFIS-PSO) to tackle this challenge, delivering fast, reliable predictions of near-plant NO₂ concentrations.
What sets this model apart
The standout innovation is the integration of neutrosophic hesitant fuzzy sets within an adaptive neuro-fuzzy inference system (ANFIS). While traditional ANFIS frameworks handle uncertainty with fuzzy logic, the neutrosophic hesitant layer adds a richer representation of ambiguity and conflict—ideal for environmental data streams that can be noisy, incomplete, or influenced by rapidly changing conditions. In practical terms, it allows the model to accommodate multiple plausible membership degrees simultaneously, capturing the uncertainty inherent in urban air quality sensing.
Under the hood: hybrid intelligence for complex dynamics
The ANNHFIS-PSO approach fuses three ingredients:
- A neural network backbone to learn nonlinear patterns from multi-source environmental signals.
- Neutrosophic hesitant fuzzy membership functions to encode uncertainty and hesitation in the data, improving resilience to variability and sensor disagreement.
- A two-stage training pipeline: particle swarm optimization (PSO) conducts a global search over parameters to escape poor local minima, followed by Adam-based fine-tuning that sharpens the solution with gradient-driven updates.
This hybrid learning scheme aims to balance exploration and precision, helping the model generalize better when conditions shift—think sudden weather changes, traffic fluctuations, or plant operation cycles that modulate NO₂ emissions.
How it was tested
To validate performance around biomass plants in Istanbul, the researchers benchmarked ANNHFIS-PSO against several strong baselines:
- Multilayer perceptron artificial neural network (MLP-ANN)
- ANFIS with particle swarm optimization (ANFIS-PSO)
- Grid-search-tuned ANFIS (ANFIS-GS)
- Long short-term memory (LSTM) network
- ANNHFIS with grid search (ANNHFIS-GS)
Accuracy was assessed using standard metrics: root mean square error (RMSE) to measure average prediction error and the coefficient of determination (R²) to gauge how well the model explains variance in observed NO₂. On the held-out test set, ANNHFIS-PSO posted an RMSE of 3.6488 µg/m³ and an R² of 0.8938—delivering the lowest RMSE and a high R² among all evaluated models. In short, it improved precision without sacrificing explanatory power.
Why it matters for cities and climate policy
Reliable nowcasts of NO₂ near biomass facilities can translate into concrete public health and operational benefits:
- Faster alerts for vulnerable populations and healthcare providers during pollution spikes.
- Dynamic compliance checks and smarter mitigation actions by plant operators.
- Evidence-based planning for traffic management, zoning, and siting of future facilities.
- Stronger foundations for emissions policy, calibration of dispersion models, and sensor network design.
Because the model explicitly handles uncertainty, it’s well-suited to real-world deployments where data can be messy—an important advantage when scaling to citywide networks or integrating low-cost sensors.
How it compares to the status quo
Classical machine learning and deep learning models (like MLPs and LSTMs) often perform well but can overfit or falter when data distributions drift. Conventional ANFIS variants, while interpretable, may struggle with the level of ambiguity typical in urban environmental signals. ANNHFIS-PSO bridges that gap. By weaving in neutrosophic hesitant fuzzy logic and using PSO for robust global search, it anchors the model in both interpretability and resilience, then leverages Adam to refine details. The result is a system that can adapt to complex, nonlinear interactions without losing stability.
Limitations and the road ahead
As with any modeling approach, broader validation will be key. Future work could explore:
- Real-time deployment across diverse seasons and operational regimes to stress-test robustness.
- Transferability to other pollutants and urban contexts with different meteorological patterns.
- Integration with edge devices and streaming pipelines for continuous, low-latency nowcasting.
- Explainability tools that trace predictions back to fuzzy rule activations for transparent decision support.
Still, the Istanbul case study indicates that ANNHFIS-PSO can provide a credible step forward in estimating near-source NO₂, bringing both accuracy and uncertainty-awareness to the front lines of air quality management.
Bottom line
ANNHFIS-PSO delivers a pragmatic and technically novel path to nowcasting NO₂ around biomass plants, pairing a flexible neuro-fuzzy core with a savvy optimization routine. With test-set results of 3.6488 µg/m³ RMSE and 0.8938 R², it sets a high bar in a tough prediction arena—offering policymakers and operators a sharper, faster lens on urban air health.
