Introduction to the LSTM-based Smart Analysis and Prediction System for Household Energy Consumption

This is an end-to-end machine learning system that uses LSTM neural networks as its core to monitor, analyze, and predict household energy consumption. The system features real-time multi-granularity energy consumption tracking, device-level electricity usage insights, next-hour energy consumption prediction with over 90% accuracy, and personalized energy-saving recommendations based on usage patterns. The project builds an interactive dashboard via Flask and is open-sourced on GitHub, with both technical demonstration and practical application value.

Project Background and Data Sources

• Project Source: Developed and maintained by Poojitha47672, released on GitHub on May 24, 2026 (repository link: https://github.com/Poojitha47672/Smart-Energy-Consumption-Analysis-and-Prediction).
• Data Foundation: The raw data comes from the Individual Household Electric Power Consumption dataset in the UCI Machine Learning Repository, containing over 2 million records with a time span from December 2006 to November 2010.
• Technology Selection: Compared to traditional linear regression models, LSTM can better capture long-term dependencies in time series, so it was chosen as the core prediction algorithm.

Technical Methods and Architecture

Model Design

Two-layer LSTM architecture: 64 units in the first layer, 32 units in the second layer, input is 24 time steps ×29 features, output is the energy consumption value at the next time point.

Feature Engineering

Constructed over 80 engineered features, including:

• Lag features (1/2/3/6/12/24/48/168-hour historical values)
• Rolling statistics (6/24/168-hour mean, 24-hour standard deviation)
• Time features (hour, day of week, month, season)
• Cyclical encoding (sine/cosine encoding for hour/month)
• Device features (sub-meter lag values)
• Derived metrics (differences, exponential moving averages, momentum indicators)

Data Preprocessing

Missing value handling → aggregating minute-level data to hour-level → splitting into training/validation/test sets in a 70/15/15 ratio.

Model Performance

Training Set Metrics

• R² score: 90-95%
• Mean Absolute Error (MAE): 0.10-0.20 kW
• Root Mean Squared Error (RMSE): 0.15-0.30 kW

Test Set Metrics

• R² score: 85-90%
• MAE: 0.15-0.25 kW

The model has good generalization ability and maintains stable prediction performance on unseen data.

Practical Application Value

• Household Users: Helps identify electricity usage patterns, predict electricity bills, and discover energy-saving opportunities.
• Researchers: Provides a reference implementation of a complete energy consumption prediction pipeline.
• Developers: Demonstrates the method of building an end-to-end system integrating deep learning models with web applications.

Deployment and Usage Guide

Tech Stack

• Backend: Python3.11, Flask3.0, TensorFlow2.13, Scikit-learn1.6, Pandas & NumPy
• Frontend: HTML5/CSS3, JavaScript, Chart.js3.9, Font Awesome6.4

Deployment Steps

1. Clone the repository
2. Create a virtual environment and install dependencies
3. Run python app.py to start the application (default port 5000)

Functional Modules

• Home page, dashboard, prediction page, energy-saving recommendation page
• RESTful API supports obtaining statistical data, hourly/daily historical data (callable via curl).

Summary and Insights

This project demonstrates the application potential of machine learning in the energy management field. By using LSTM to capture complex patterns in time series and combining it with well-designed feature engineering, high-precision energy consumption prediction is achieved. The complete architecture (data processing → model deployment → API design → frontend visualization) provides an excellent reference template for similar prediction applications.

Recommendation: For developers who are new to time series prediction or energy data analysis, this is a high-quality open-source project worth in-depth study.