AI4ALL-Project

Link to our repo: https://github.com/NnennaN123/AI4ALL-Project

Link to Streamlit App: https://cranberryai4allproject.streamlit.app/

CRANBerry Team’s Project!

Overview

This project builds machine learning models to predict whether a location (from NREL’s Wind Toolkit) has a wind turbine based on various features. The workflow includes data loading, spatial matching, feature engineering, model training, and evaluation.

Project Timeline & Model Evolution

This project demonstrates an iterative machine learning workflow, progressively improving model performance through algorithm selection, feature engineering, and hyperparameter optimization.

Phase 1: Baseline Models (Initial Exploration)

1. Logistic Regression Model

Location: logistic_regression/logistic_regression.ipynb

The initial baseline model uses Logistic Regression to establish a performance benchmark. This linear model was chosen for its interpretability and computational efficiency.

Model Configuration:

• Algorithm: Logistic Regression with balanced class weights
• Features: 4 base features
  • fraction_of_usable_area: Fraction of grid cell usable for wind development
  • capacity: Potential capacity of the site
  • wind_speed: Average wind speed at the site
  • capacity_factor: Expected capacity factor (efficiency)
• Preprocessing: StandardScaler for feature normalization
• Hyperparameters:
  • max_iter=1000
  • class_weight="balanced" (to handle class imbalance)
  • n_jobs=-1 (parallel processing)

Performance Metrics:

• ROC-AUC Score: 0.732
• Accuracy: 0.643

• Precision (No Turbine): 0.779

  Recall: 0.591

  F1-Score: 0.672

• Precision (Turbine): 0.522

  Recall: 0.728

  F1-Score: 0.608

Key Finding: Good recall for detecting turbines (72.8%) but lower precision (52.2%), indicating many false positives. This established the baseline for comparison.

---

2. Random Forest Model

Location: random_forest/random_forest.ipynb

The second iteration uses Random Forest, an ensemble method that combines multiple decision trees to capture non-linear relationships and improve prediction accuracy.

Model Configuration:

• Algorithm: Random Forest Classifier
• Features: Same 4 base features as Logistic Regression
• Preprocessing: No scaling required (tree-based models are scale-invariant)
• Hyperparameters:
  • n_estimators=500 (number of trees)
  • max_leaf_nodes=16 (limits tree depth)
  • n_jobs=-1 (parallel processing)
  • random_state=42 (reproducibility)

Performance Metrics:

• ROC-AUC Score: 0.770
• Accuracy: 0.703

• Precision (No Turbine): 0.717

  Recall: 0.860

  F1-Score: 0.782

• Precision (Turbine): 0.663

  Recall: 0.448

  F1-Score: 0.535

Improvement over Logistic Regression:

• +5.2% improvement in ROC-AUC (0.770 vs 0.732)
• +6.0% improvement in accuracy (0.703 vs 0.643)
• Better overall F1-scores for both classes

Key Finding: Better overall performance but more conservative in predicting turbines (44.8% recall vs 72.8%), suggesting the need for more sophisticated algorithms.

---

3. XGBoost Model (Initial)

Location: xgboost/xgboost.ipynb

The third iteration uses XGBoost (Extreme Gradient Boosting), a powerful gradient boosting framework known for superior performance in structured data problems.

Model Configuration:

• Algorithm: XGBoost Classifier
• Features: Same 4 base features as previous models
• Preprocessing: No scaling required (tree-based models are scale-invariant)
• Hyperparameters:
  • n_estimators=300 (number of boosting rounds)
  • learning_rate=0.05 (step size shrinkage)
  • max_depth=6 (maximum tree depth)
  • subsample=0.8 (row subsampling ratio)
  • colsample_bytree=0.8 (column subsampling ratio)
  • objective='binary:logistic' (binary classification)
  • eval_metric='logloss' (evaluation metric)

Performance Metrics:

• ROC-AUC Score: 0.847
• Accuracy: 0.766

• Precision (No Turbine): 0.796

  Recall: 0.839

  F1-Score: 0.817

• Precision (Turbine): 0.708

  Recall: 0.645

  F1-Score: 0.675

Improvement over Previous Models:

• +10.0% improvement in ROC-AUC over Random Forest (0.847 vs 0.770)
• +15.8% improvement in ROC-AUC over Logistic Regression (0.847 vs 0.732)
• +6.3% improvement in accuracy over Random Forest (0.766 vs 0.703)
• Best overall balance between precision and recall

Key Finding: XGBoost showed the best performance, confirming it as the optimal algorithm. However, further improvements were possible through feature engineering.

---

Phase 2: Feature Engineering & Optimization

4. Feature Engineering Exploration

Location: feature_engineering.ipynb

Conducted systematic feature engineering to identify the optimal feature combination. Tested 12 different feature configurations (X_train_1 through X_train_12) including:

• Base features: fraction_of_usable_area, capacity_factor
• Numeric features: wind_speed, capacity
• Categorical features: wind_speed_category, capacity_category (converted to numeric)
• Geographic features: State (one-hot encoded, ~50+ columns) or Region (one-hot encoded, ~7 columns)
• Engineered interaction features:
  • combined_wind_rescource = wind_speed × capacity_factor
  • potential_with_constraints = capacity × fraction_of_usable_area

Key Findings:

• Best feature combination: X_train_5 (Base + Numeric + State + New Features)
  • ROC-AUC: 0.9305 (before hyperparameter tuning)
  • Accuracy: 0.8548
  • F1-Score: 0.8134
• Geographic features (State) provided significant performance boost (+0.05-0.08 ROC-AUC)
• Numeric features outperformed categorical versions
• Interaction features added value, especially when combined with geographic features

Documentation: See X_TRAIN_COMBINATIONS_README.md for complete analysis of all 12 feature combinations.

---

5. Final Model: Feature-Engineered XGBoost with Hyperparameter Tuning

Location: xgboost/xgboost_hyperparameter_tuning.ipynb

Model File: xgboost/xgboost_tuned_feat_eng_wind_model.pkl

The final production model combines the best feature engineering (X_train_5) with systematic hyperparameter optimization using GridSearchCV.

Model Configuration:

• Algorithm: XGBoost Classifier (tuned)
• Features: 56 features (X_train_5 configuration)
  • Base: fraction_of_usable_area, capacity_factor
  • Numeric: wind_speed, capacity
  • Geographic: State (one-hot encoded, ~50+ columns)
  • Engineered: combined_wind_rescource, potential_with_constraints
• Preprocessing: No scaling required (tree-based models are scale-invariant)
• Hyperparameter Tuning: GridSearchCV with 5-fold cross-validation
  • Tested 54 parameter combinations
  • Scoring metric: ROC-AUC
• Best Hyperparameters:
  • max_depth: 8
  • learning_rate: 0.1
  • n_estimators: 300
  • subsample: 0.7
  • colsample_bytree: 1.0
  • scale_pos_weight: 1.6201 (handles class imbalance)

Final Performance Metrics:

• ROC-AUC Score: 0.9545 ⭐
• Accuracy: 0.8783 ⭐
• F1-Score: 0.8537 ⭐

• Precision (No Turbine): 0.960

  Recall: 0.840

  F1-Score: 0.896

• Precision (Turbine): 0.780

  Recall: 0.943

  F1-Score: 0.854

Improvement Journey:

• +12.7% improvement in ROC-AUC over initial XGBoost (0.9545 vs 0.847)
• +11.2% improvement in accuracy over initial XGBoost (0.8783 vs 0.766)
• +30.4% improvement in ROC-AUC over Logistic Regression baseline (0.9545 vs 0.732)
• +23.5% improvement in accuracy over Logistic Regression baseline (0.8783 vs 0.643)

Key Achievements:

• Excellent turbine recall (94.3%) - captures nearly all turbine locations
• High precision for both classes (96.0% for No Turbine, 78.0% for Turbine)
• Best-in-class performance across all metrics
• Production-ready model deployed in Streamlit app

---

Complete Model Comparison

Metric	Logistic Regression	Random Forest	XGBoost (Initial)	XGBoost (Final)	Winner
ROC-AUC	0.732	0.770	0.847	0.9545	Final XGBoost
Accuracy	0.643	0.703	0.766	0.8783	Final XGBoost
F1-Score	0.640	0.658	0.746	0.8537	Final XGBoost
Turbine Recall	0.728	0.448	0.645	0.943	Final XGBoost
Turbine Precision	0.522	0.663	0.708	0.780	Final XGBoost
No Turbine F1	0.672	0.782	0.817	0.896	Final XGBoost

Performance Evolution:

• Phase 1 (Baseline): Logistic Regression → Random Forest → XGBoost
  • ROC-AUC: 0.732 → 0.770 → 0.847 (+15.7% total improvement)
• Phase 2 (Optimization): Feature Engineering → Hyperparameter Tuning
  • ROC-AUC: 0.847 → 0.9305 → 0.9545 (+12.7% additional improvement)
• Total Improvement: +30.4% ROC-AUC from baseline to final model

Key Insights:

• The final feature-engineered XGBoost model achieves the best performance across all metrics
• Feature engineering provided the largest performance boost (+0.0835 ROC-AUC)
• Hyperparameter tuning further refined the model (+0.0240 ROC-AUC)
• The iterative approach demonstrates systematic ML engineering practices
• Final model achieves excellent balance: 94.3% recall for turbines with 78.0% precision

---

Technical Highlights

Feature Engineering Process

• Systematic testing: 12 feature combinations evaluated
• Geographic features: State-level encoding proved most valuable
• Interaction features: Captured non-linear relationships between variables
• Feature selection: Numeric features outperformed categorical discretizations

Hyperparameter Optimization

• Method: GridSearchCV with 5-fold cross-validation
• Search space: 54 parameter combinations tested
• Optimization metric: ROC-AUC (appropriate for imbalanced classification)
• Result: Significant improvement over default parameters

Model Selection Strategy

1. Algorithm comparison: Logistic Regression → Random Forest → XGBoost
2. Feature engineering: Systematic evaluation of 12 configurations
3. Hyperparameter tuning: Grid search on best feature set
4. Final validation: Test set performance confirms production readiness

---

Data

The project uses:

• USWTDB (US Wind Turbine Database): Contains information about existing wind turbines
• NREL WTK (Wind Toolkit): Grid cell locations and wind resource data

Spatial matching is performed using geospatial joins to match turbines to NREL grid cells within a 25 km radius.

---

Project Structure

• logistic_regression/: Baseline Logistic Regression model
• random_forest/: Random Forest model for comparison
• xgboost/:
  • xgboost.ipynb: Initial XGBoost model (4 features)
  • xgboost_hyperparameter_tuning.ipynb: Final optimized model with feature engineering
  • xgboost_tuned_feat_eng_wind_model.pkl: Production model
  • xgboost_tuned_feat_eng_model_metrics.json: Final model metrics
• feature_engineering.ipynb: Systematic feature engineering exploration
• X_TRAIN_COMBINATIONS_README.md: Complete documentation of 12 feature combinations tested
• datasets/: Training and source data files
• streamlit/: Production Streamlit application

---

Deployment

The final model is deployed in a Streamlit web application:

• Live App: https://cranberryai4allproject.streamlit.app/
• Model: Feature-engineered XGBoost with hyperparameter tuning
• Performance: ROC-AUC 0.9545, Accuracy 0.8783