A Machine Learning Model Ensemble for Mixed Power Load Forecasting across Multiple Time Horizons.

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• The shift towards renewable energy necessitates advanced smart grid technologies.
• Accurate load forecasting is crucial for efficient power system planning, operation, and management.
• Traditional grid models are evolving towards smart grid frameworks.

Purpose of the Study:

• To develop a novel mixed ensemble forecasting scheme for electrical load.
• To provide accurate predictions across multiple time horizons, from 15 minutes to 24 hours.
• To enhance the reliability and efficiency of smart grid operations.

Main Methods:

• Utilized a pool of diverse machine learning models: neural networks, linear regression, support vector regression, random forests, and sparse regression.
• Implemented an online decision mechanism to dynamically weight individual model performance for final predictions.
• Evaluated the scheme using real-world electrical load data from a high voltage/medium voltage substation.

Main Results:

• Achieved high accuracy with R-squared values ranging from 0.99 (15 min ahead) to 0.79 (24 h ahead).
• Demonstrated superior or competitive performance compared to state-of-the-art machine learning methods and other ensemble techniques.
• Validated the effectiveness of the proposed mixed ensemble forecasting scheme on practical grid data.

Conclusions:

• The proposed mixed ensemble load forecasting scheme is highly effective for smart grids.
• The method offers robust and accurate predictions for various short-term and medium-term horizons.
• This approach contributes to improved planning and operational efficiency in evolving power systems.