Wind Power Forecasting with Tensor Decomposition (CP & Tucker)

This project evaluates how tensor decomposition (CP and Tucker) affects the performance of short-term wind turbine power forecasting. We use real SCADA data and compare multiple forecasting models trained on raw vs. decomposed reconstructions.

Overview

1. Data Preprocessing

• Load SCADA CSV
• Parse timestamps
• Normalize features
• Build tensor (Days × Window × Features)

2. Tensor Decomposition

• CP ranks
• Tucker ranks
• Reconstruction error analysis

3. Forecasting Models

• Baseline
• Ridge
• MLP
• CNN
• LSTM

4. Evaluation

• CSV summaries
• Plots for CP/Tucker errors
• Model performance

Repository Structure

MLGproject/
│
├── preprocess.py
├── windows.py
├── decompose.py
├── models.py
├── run.py
├── analyze_results.py
├── results/
└── data/

Installation

1. Install dependencies:

pip install -r requirements.txt

2. Set up your data:
   • Option A: Set environment variable: set DATA_CSV=C:\path\to\your\scada.xlsx
   • Option B: Create data/ folder and place scada.xlsx there
   • Option C: Edit DATA_CSV path in config.py

Running

1. Test setup (recommended first):

python smoke_test.py

2. Run full pipeline:

python run.py          # Main training pipeline
python viz.py           # Generate visualizations
python analyze.py       # Additional analysis plots

Data Requirements

• Format: Excel (.xlsx) or CSV file
• Columns: Must include Timestamps, Power, and 15 SCADA features
• Frequency: 10-minute intervals recommended
• Duration: At least 30 days for meaningful results

Troubleshooting

• Import errors: Run pip install -r requirements.txt
• Data not found: Check DATA_CSV configuration in config.py
• CUDA errors: Decomposition will automatically fall back to CPU
• Memory issues: Reduce MAX_DAYS in config.py or use smaller Tucker ranks