AI-POWERED ICING FORECASTING
Operational Icing Forecast for Power Grids Based on WRF-ICE and Machine Learning Correction
Authors: Yujie Li*, Yang Yang, Meng Li, Mingguan Zhao, Xiaojing Yang
Published: 14 November 2025
This study pioneers a multi-data source approach for enhancing icing forecasts on power grids, integrating advanced machine learning with numerical weather prediction.
The research introduces a novel correction model utilizing actual observations, reanalysis data (CLDAS), and WRF-ICE forecasts. By incorporating Light Gradient Boosting Machine (LightGBM) and Bayesian Optimization, the model achieves real-time correction of icing forecasts across 1-24 hour lead times. Comparative results demonstrate a significant reduction in Mean Absolute Error (MAE) from 1.84 mm to 0.3 mm and an improvement in the correlation coefficient (R) from 0.31 to 0.9 at a 24-hour lead time. For 8-hour nowcasting, MAE remains below 0.2 mm, effectively mitigating time lags and underestimation in original forecasts.
Executive Impact & Key Outcomes
Implementing this advanced AI-driven icing forecast system offers substantial benefits, enhancing grid reliability and operational efficiency for power utilities.
0.0mm
MAE at 24-Hour Lead
0.0R
Correlation Coefficient
0.0mm
MAE for 8-Hour Nowcasting
0
Max Forecast Lead Time
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Advanced Model Correction Techniques
This section details the core methodology behind the improved icing forecast, focusing on the synergistic use of LightGBM and Bayesian Optimization to refine WRF-ICE model outputs.
0.3mm
MAE at 24-hour lead time, down from 1.84 mm (WRF-ICE).
0.9R
Correlation Coefficient (R) at 24-hour lead time, up from 0.31 (WRF-ICE).
Enhanced Icing Forecast Workflow
Actual Observations + CLDAS Data
→
WRF-ICE Forecasts
→
LightGBM + Bayesian Optimization
→
Real-time Icing Correction
Multi-Source Data Integration & Processing
Understanding how diverse data sources are harmonized and processed is crucial for the model's accuracy. This includes real-time observations, reanalysis data, and numerical forecasts, alongside robust data cleaning protocols.
| Feature | LightGBM | Legacy Models |
|---|---|---|
| Training Speed | Fast | Moderate |
| Memory Usage | Low | High |
| Data Handling | Robust for large datasets | Can struggle with large datasets |
| Accuracy | Significantly improved | Limited by non-linearity |
| Scalability | Excellent for enterprise | Good, but less efficient |
0.2mm
MAE for 1-8 hour nowcasting, demonstrating high precision for short lead times.
Performance Validation & Real-World Application
Quantitative evaluation metrics and real-world case studies confirm the model's superior performance, highlighting its stability and generalization capabilities across various forecast lead times.
24hours
Maximum forecast lead time with corrected data, supporting both short-term nowcasting and long-term planning.
Real-World Impact: Tower 42 Icing Forecast
The LightGBM correction model demonstrated remarkable accuracy at Tower 42, significantly improving upon WRF-ICE's forecasts. At a 6-hour lead time, the corrected predictions closely matched actual observations, effectively mitigating the time lags and underestimation issues. For 24-hour forecasts, the model maintained high consistency, validating its robustness for both short and long lead times. This real-time, precise forecasting enables proactive measures for grid stability.
Calculate Your Potential ROI
Estimate the significant operational savings and efficiency gains your organization could achieve with AI-powered forecasting.
Annual Savings
$0
Annual Hours Reclaimed
0
Your AI Implementation Roadmap
A clear, phased approach to integrating advanced AI into your power grid operations for optimal icing forecast and management.
Phase 1: Discovery & Customization
Comprehensive analysis of your existing grid infrastructure, data sources, and operational requirements. Customization of the WRF-ICE and LightGBM models to local meteorological conditions and power line specifics.
Phase 2: Data Integration & Model Training
Secure integration of real-time monitoring data, CLDAS reanalysis, and WRF-ICE forecasts. Initial training and validation of the machine learning correction model using historical icing events.
Phase 3: Pilot Deployment & Refinement
Deployment of the corrected forecast system in a pilot region or for critical transmission lines. Continuous monitoring of forecast accuracy and iterative refinement based on operational feedback and performance metrics.
Phase 4: Full-Scale Rollout & Ongoing Support
Expansion of the AI forecasting system across your entire power grid. Provision of ongoing technical support, model updates, and performance optimization to ensure sustained benefits.
Ready to Enhance Your Grid's Resilience?
Book a personalized consultation with our AI specialists to explore how our operational icing forecast solution can protect your power infrastructure.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat