Decision tree and reinforcement learning for contextual electricity consumption forecasting in buildings

Area of Science:

• Energy Systems
• Artificial Intelligence
• Machine Learning

Background:

• Accurate energy consumption prediction is crucial for optimizing energy control and storage.
• Forecasting algorithms like Neural Networks, K-Nearest Neighbors (KNN), and XGBoost are vital for energy pattern estimation.
• Context-specific application of forecasting models is recommended for enhanced performance.

Purpose of the Study:

• To develop a method for predicting building energy consumption at five-minute intervals over a week.
• To leverage unique building contexts for more accurate energy demand forecasting.
• To dynamically select the most effective forecasting algorithm based on identified contexts.

Main Methods:

• Utilized K-Nearest Neighbors (KNN) and Artificial Neural Networks (ANN) for energy consumption prediction.
• Employed a decision tree to identify unique building contexts based on energy and sensor data patterns.
• Implemented a Multiarmed Bandit algorithm with reinforcement learning to select the optimal forecasting model for each context.

Main Results:

• Context identification using decision rules improved confidence bounds in evaluating forecasting algorithms.
• Achieved low SMAPE (Symmetric Mean Absolute Percentage Error) forecasting errors: 3.54% with KNN and 4.79% with ANN in the third context.
• Obtained SMAPE errors of 4.91% with KNN and 4.54% with ANN in the fourth context.

Conclusions:

• Context-specific forecasting significantly enhances the reliability of energy prediction models.
• The proposed method effectively identifies unique building energy consumption patterns.
• Dynamic algorithm selection via Multiarmed Bandit optimizes prediction accuracy for different contexts.