Perceiving Unpredictability for New Energy Power and Electricity Consumption Forecasting.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Accurate sensor network data prediction is crucial for electric power systems and traffic planning.
• Current deep learning models assume uniform predictability, neglecting temporal distance and signal randomness.
• This uniform approach can hinder the learning of generalizable patterns in sensor data.

Purpose of the Study:

• To develop a novel loss function that accounts for the intrinsic unpredictability of forecasting tasks.
• To improve the accuracy and reliability of sensor network data prediction models.
• To enhance the stability of critical infrastructures and urban operational efficiency.

Main Methods:

• Introduced an Unpredictability Perception loss function with dynamically computed supervision weights.
• Unified two dimensions of unpredictability: signal randomness (local spectral entropy) and temporal distance (exponential decay).
• Applied the proposed loss function to the TimeMixer model for experimental validation.

Main Results:

• The Unpredictability Perception loss demonstrated performance improvements on multiple public benchmark datasets.
• Reduced supervision on random signal segments and distant future points.
• Enhanced forecasting accuracy for sensor network data by matching supervision to signal predictability.

Conclusions:

• The Unpredictability Perception loss function offers a more reliable technical foundation for critical infrastructure stability.
• Improved forecasting accuracy supports enhanced grid stability and optimized urban traffic systems.
• This method provides a more robust approach to deep learning for time-series forecasting in complex domains.