This repository contains the implementation and analysis of artificial neural network models used to predict emissions from gas turbines. The models were evaluated on their ability to forecast carbon monoxide (CO) and nitrogen oxides (NOx) based on turbine operating conditions. The project explores two ANN approaches: a Scikit-learn-based Multi-Layer Perceptron (MLP) and a Keras Sequential model, including hyperparameter optimization.
| File/Folder | Description |
|---|---|
ANN.ipynb |
Full notebook with data preprocessing, modeling, evaluation, and visualizations |
data/ |
Gas turbine emissions dataset from UCI (2011–2015) |
To build and evaluate neural network models that forecast CO and NOx emissions using turbine sensor data.
- Source: UCI Machine Learning Repository (link)
- Records: 36,733 hourly records (2011–2015)
- Features: 9 input features (e.g., temperature, pressure) + 2 targets (CO, NOx)
- Preprocessing: Yeo-Johnson transformation, PCA, chronological train/test split
- Hidden Layers: (128, 64)
- Activation: ReLU
- Solver: Adam with early stopping
- Regularization: L2
- Optimizer: Adam (learning rate: 0.001 or 0.0005)
- Layers: Tuned from (64, 32) to (128, 64, 32)
- Dropout: 0.1–0.2
- Early stopping enabled
| Model | MSE | R² |
|---|---|---|
| MLP | 1.9752 | 0.588 |
| Keras | 1.9809 | 0.587 |
Both models achieved moderate accuracy with R² ≈ 0.58
| Model | MSE | R² |
|---|---|---|
| MLP | 187.18 | -0.67 |
| Keras (baseline) | 192.64 | -0.72 |
| Keras (tuned) | 184.26 | -0.65 |
NOx proved significantly harder to predict; optimized model showed improvement but remains limited.
- Incorporate external features (e.g., weather, fuel mix)
- Explore LSTM or ensemble models for NOx
- Apply interpretability tools (e.g., SHAP)
- Test noise reduction and advanced feature engineering
Itorobong Akpan
MSc Data Science & Computational Intelligence
Coventry University, UK
📧 akpani4@uni.coventry.ac.uk
🔗 GitHub Profile