Ensemble learning approach for advanced metering infrastructure in future smart grids.

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Traditional batch learning for load forecasting struggles with real-time data integration.
• Existing methods for Electricity Load and Price Forecasting (ELPF) face challenges with large datasets, non-linearity, high variance, and high dimensions.
• Online learning offers efficient adaptation to new data for improved forecasting.

Purpose of the Study:

• To develop a compact and improved algorithm for synchronized Electricity Load and Price Forecasting (ELPF).
• To enhance the accuracy and efficiency of load and price forecasting in smart grids.
• To address limitations of existing methods in handling complex, high-dimensional, and non-linear energy data.

Main Methods:

• A novel ELPF framework incorporating data separation (high/low consumers), missing/unstandardized data handling, and feature selection/reduction.
• Implementation of an improved Residual Network (ResNet-152) and a machine-improved Support Vector Machine (SVM) for forecasting.
• Application of distinct mechanisms including regularization, base learner selection, and hyperparameter tuning to optimize ResNet-152 and SVM performance.

Main Results:

• The proposed method demonstrated an 8% improvement in performance measures compared to existing schemes.
• The framework effectively handles complex consumer data by reducing time complexity and mitigating overfitting.
• Exploration of various ResNet-152 and SVM structures enhanced regularization, base learner selection, and parameter tuning for superior fitting capabilities.

Conclusions:

• The developed ELPF framework offers a significant advancement in forecasting accuracy and efficiency for smart grids.
• The proposed online learning approach effectively synchronizes with diverse data procedures in ELPF.
• The method is suitable for industry-based applications, improving energy grid stability and management.