Energy Prediction Toolkit

A comprehensive machine learning toolkit for building energy consumption prediction, featuring advanced LSTM models with Firefly Algorithm optimization and support for multiple datasets.

🎯 Overview

This project implements and evaluates various energy prediction models including:

• ModLSTM: Modified LSTM with Firefly Algorithm optimization
• Standard LSTM: Baseline LSTM implementation
• SAMFOR: Sequential Attention-based Model for Forecasting
• SVR/RFR: Support Vector Regression and Random Forest baselines

The toolkit supports two datasets:

• Portuguese Dataset: Single-household energy consumption data
• ASHRAE Dataset: Multi-building commercial energy consumption data

📁 Repository Structure

Energy-Prediction/
├── models/                     # Model implementations
│   ├── LSTM_comb.py           # Main LSTM training pipeline
│   ├── LSTM_hyperpara_search.py # Firefly Algorithm hyperparameter optimization
│   ├── SAMFOR_trial1.py        # SAMFOR model implementation
│   └── SVR_energy_data_paper2.py # SVR/RFR baseline models
├── tools/                      # Data preprocessing utilities
│   ├── preprocess_data2.py    # Shared utilities for Portuguese dataset
│   ├── preprocessing_ashrae.py # ASHRAE-specific preprocessing
│   └── resample_dataset.py    # Data resampling utilities
├── dataset/                    # Datasets
│   ├── 1Hz/                   # Original Portuguese dataset (1Hz)
│   ├── resampled data/        # Resampled Portuguese data
│   └── ASHRAE/                # ASHRAE Great Energy Predictor III dataset
├── results/                    # Model outputs and results
│   ├── 1s/                    # Portuguese dataset results
│   └── ashrae/                # ASHRAE dataset results
├── niapy/                      # Firefly Algorithm implementation
└── config.py                  # Configuration settings

🚀 Quick Start

Prerequisites

• Python 3.8+
• TensorFlow/Keras
• scikit-learn
• pandas, numpy
• matplotlib, seaborn

Installation

1. Clone the repository:

git clone <repository-url>
cd Energy-Prediction

2. Create a virtual environment:

conda create -n FF python=3.8
conda activate FF

3. Install dependencies:

pip install tensorflow keras scikit-learn pandas numpy matplotlib seaborn

Usage

Portuguese Dataset Training

1. Train LSTM model:

python models/LSTM_comb.py

2. Run hyperparameter optimization:

python models/LSTM_hyperpara_search.py

3. Train baseline models:

python models/SVR_energy_data_paper2.py
python models/SAMFOR_trial1.py

4. Generate comparative results:

python plot_results.py

ASHRAE Dataset

• SVR baseline (disjoint buildings)

conda run -n FF python -m ashrae.call_svr_ashrae

• SAMFOR baseline

conda run -n FF python -m ashrae.call_samfor_ashrae

• LSTM Hyperparameter Search (overnight)

conda run -n FF python -m ashrae.call_lstm_search_ashrae

📁 Centralized ASHRAE Results

All ASHRAE models now use a centralized results saving system:

results/ashrae/
├── svr/
│   ├── metrics.csv              # Model performance metrics
│   ├── model_info.json          # Model parameters and metadata
│   └── artifacts/
│       └── SVR_ASHRAE.obj       # Unscaled predictions and test data
├── samfor/
│   ├── metrics.csv
│   ├── model_info.json
│   └── artifacts/
│       └── SAMFOR.obj
└── lstm/
    ├── metrics.csv              # LSTM results (from log_results_LSTM)
    ├── model_info.json          # Best parameters and metadata
    ├── training_log.json        # Training history and convergence
    └── artifacts/
        └── LSTM_ModFF.obj       # Unscaled predictions and test data

Standardized Artifacts Include:

• Unscaled ground truth (y_test) and predictions (y_test_pred)
• Model parameters, timing, and metadata
• Consistent file naming and structure across all models

📊 Datasets

Portuguese Dataset

• Source: Single-household energy consumption data
• Frequency: 1Hz (resampled to 1s)
• Features: Power consumption measurements
• Preprocessing: MinMax scaling, sliding window sequencing
• Sequence Length: 23 timesteps

ASHRAE Dataset

• Source: ASHRAE Great Energy Predictor III (Kaggle)
• Scope: Multi-building commercial energy consumption
• Features: Building metadata, weather data, temporal features
• Preprocessing: MinMax scaling (fit on train only), one-hot encoding; building_id preserved and not scaled for windowing
• Sequence Length: 23 timesteps
• Sample Size: ~130,000 total windows using disjoint building splits (Train≈52k, Val≈26k, Test≈52k)
• Resampling: None for ASHRAE 1s/1Hz; dataset already at target granularity

🔧 Key Features

Advanced LSTM Models

• ModLSTM: Modified LSTM architecture with optimized hyperparameters
• Firefly Algorithm: Bio-inspired optimization for hyperparameter tuning
• Early Stopping: Prevents overfitting during training
• Sequence-to-One: Predicts next timestep energy consumption

Comprehensive Preprocessing

• Feature Engineering: Temporal, building, and weather features
• Missing Value Imputation: Mean imputation for numerical features
• Normalization: Dataset-specific scaling (MinMax vs Z-score)
• Sequential Cropping: Time-series consistent data sampling

Evaluation Metrics

• RMSE: Root Mean Square Error
• MAE: Mean Absolute Error
• MAPE: Mean Absolute Percentage Error
• RMSLE: Root Mean Squared Logarithmic Error
• R²: Coefficient of Determination

🔬 Research Context

Problem Statement

Building energy consumption prediction is crucial for:

• Smart Grid Management: Optimizing energy distribution and demand response
• Building Optimization: Improving HVAC system efficiency and occupant comfort
• Sustainability Goals: Supporting Net-Zero Energy Building (NZEB) initiatives
• Cost Reduction: Minimizing energy costs through predictive maintenance

Challenges in Energy Prediction

1. Temporal Dependencies: Energy consumption exhibits complex time-series patterns
2. Multi-scale Variability: Patterns vary across hours, days, seasons, and years
3. External Factors: Weather, occupancy, and building characteristics significantly impact consumption
4. Data Quality: Missing values, outliers, and measurement errors in sensor data
5. Generalizability: Models trained on single buildings may not generalize to diverse building types

Model Architecture Rationale

• LSTM Networks: Capture long-term dependencies in time-series data
• ModLSTM: Enhanced LSTM with optimized architecture for energy prediction
• Firefly Algorithm: Bio-inspired optimization for hyperparameter tuning
• Attention Mechanisms: Focus on relevant temporal patterns (SAMFOR)
• Ensemble Methods: Combine multiple models for robust predictions

🛠️ Configuration

Key hyperparameters (from Table II):

• LSTM Units: 72
• Learning Rate: 0.010
• Sequence Length: 23
• Batch Size: 64
• Epochs: 50 (with early stopping)

📝 Implementation Details

Data Preprocessing Pipeline

1. Data Loading: Load raw datasets
2. Merging: Join building metadata and weather data
3. Feature Engineering: Create temporal and categorical features
4. Normalization: Apply dataset-specific scaling
5. Sequencing: Generate sliding window sequences for LSTM
6. Splitting: Train/validation/test splits

Model Training Pipeline

1. Model Building: Construct LSTM architecture
2. Compilation: Configure optimizer and loss function
3. Training: Fit model with early stopping
4. Evaluation: Compute metrics on test set
5. Persistence: Save model and results

Dataset Diversity Challenge

This toolkit addresses a critical limitation in energy prediction research: Limited Dataset Diversity. Most studies focus on single buildings or homogeneous datasets, limiting model generalizability.

Our Approach:

• Multi-Dataset Validation: Portuguese residential + ASHRAE commercial data
• Diverse Building Types: Single-family homes vs. multi-building commercial complexes
• Different Climates: Portuguese climate vs. multiple US climate zones
• Varied Meter Types: Electricity vs. multiple utility types (electricity, chilled water, steam, hot water)
• Scale Differences: Single building vs. thousands of buildings

Methodological Contributions

1. Unified Preprocessing Pipeline: Consistent feature engineering across datasets
2. Adaptive Normalization: Dataset-specific scaling strategies
3. Memory-Efficient Processing: Sequential cropping for large datasets
4. Comprehensive Evaluation: Multiple metrics on original-scale data
5. Reproducible Research: Complete preprocessing and training pipelines

📚 Background Literature

Energy Prediction in Buildings

Building energy consumption prediction has evolved from simple regression models to sophisticated deep learning approaches. Key developments include:

• Traditional Methods: Linear regression, ARIMA, and support vector machines
• Machine Learning: Random forests, gradient boosting, and neural networks
• Deep Learning: LSTM, GRU, and transformer-based models
• Hybrid Approaches: Combining physical models with data-driven methods

Time Series Forecasting Challenges

Energy consumption exhibits unique characteristics:

• Seasonality: Daily, weekly, and annual patterns
• Non-stationarity: Changing patterns over time
• External Dependencies: Weather, occupancy, and building operations
• Multi-scale Dynamics: Short-term fluctuations and long-term trends

Optimization in Neural Networks

Hyperparameter optimization is crucial for model performance:

• Grid Search: Exhaustive but computationally expensive
• Random Search: More efficient than grid search
• Bayesian Optimization: Model-based optimization
• Bio-inspired Algorithms: Firefly, particle swarm, genetic algorithms

📚 References

• ASHRAE Great Energy Predictor III: Kaggle Competition
• Firefly Algorithm: Yang, X.S. (2008). Nature-Inspired Metaheuristic Algorithms
• LSTM Networks: Hochreiter, S. & Schmidhuber, J. (1997). Long Short-Term Memory
• Energy Prediction: Ahmad, T. et al. (2018). A review on renewable energy and electricity requirement forecasting

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🐛 Troubleshooting

Common Issues

1. Memory Errors: Reduce max_samples in preprocessing
2. Import Errors: Ensure all dependencies are installed
3. Path Issues: Check dataset paths in configuration files
4. CUDA Issues: Verify TensorFlow GPU installation

Support

For issues and questions:

• Check the troubleshooting section
• Review the preprocessing reports
• Examine the test scripts for examples

---

Note: This toolkit is designed for research purposes and addresses the challenge of limited dataset diversity in energy prediction research.