Predicting Energy Consumption Using LSTM, Multi-Layer GRU and Drop-GRU Neural Networks

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• Smart grid development and advanced measuring infrastructure necessitate accurate short-term power consumption forecasting.
• Predicting future power loads is essential for energy waste reduction and effective power management strategies.
• Energy consumption data, as historical time series, requires analysis for future consumption forecasting.

Purpose of the Study:

• To model and compare three machine learning algorithms for time series power forecasting.
• To evaluate the effectiveness of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Drop-GRU models.
• To identify the most accurate model for predicting future power loads and preventing consumption peaks.

Main Methods:

• Utilized time series power consumption data for model training and prediction.
• Implemented and compared Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Drop-GRU algorithms.
• Conducted experiments using real-world power consumption data from French cities across various time horizons.

Main Results:

• The Long Short-Term Memory (LSTM) model demonstrated superior performance compared to GRU and Drop-GRU.
• LSTM achieved fewer prediction errors and finer precision in forecasting power consumption.
• Experimental results validated LSTM's effectiveness across different time horizons.

Conclusions:

• LSTM-based power consumption predictions enable proactive decision-making and load shedding.
• Accurate forecasting significantly impacts power quality planning and power equipment maintenance.
• The study highlights LSTM as a key technology for optimizing smart grid operations.