Strategic Explainable Electricity Load Forecasting ⚡

Overview

Energy load forecasting models are typically evaluated using predictive accuracy metrics such as RMSE and MAE. However, high forecasting performance does not necessarily imply alignment with long-term sustainability objectives.

This project introduces the Strategic Alignment Score (SAS) — a novel metric that evaluates how well machine learning models align with sustainability-driven energy planning goals.

Using explainable AI techniques (SHAP and LIME), we analyze how different models make predictions and assess their strategic relevance.

---

🎯 Objectives

• Build accurate electricity load forecasting models
• Apply explainable AI (SHAP & LIME)
• Map model features to policy-relevant strategic categories
• Develop a new evaluation metric (SAS)
• Compare models on both accuracy and strategic alignment

---

📊 Models Used

• XGBoost
• Reduced XGBoost
• Random Forest
• LSTM (Deep Learning)

---

🧠 Key Concept: Strategic Alignment Score (SAS)

Features are grouped into three strategic categories:

• Reactive → Short-term demand (Lag features)
• Climate → Temperature, humidity (cooling demand)
• Structural → Time-based patterns (hour, day)

SAS measures how much a model relies on these groups, enabling evaluation beyond accuracy.

---

📈 Results

Model	RMSE	SAS
XGBoost	~29	0.21
Reduced XGB	~33	0.22
Random Forest	~30	0.20
LSTM	~51	0.33

Key Findings

• Tree-based models → High accuracy, low strategic alignment
• LSTM → Lower accuracy, higher strategic alignment
• Clear accuracy vs interpretability trade-off

---

🔍 Explainability

• SHAP → Global feature importance
• LIME → Local prediction explanations
• Feature contributions mapped to Strategic KPIs

---

📊 Additional Analysis

• Statistical testing (bootstrap confidence intervals)
• Entropy analysis of explanations
• Scenario-based SAS weighting
• Multi-model comparison

---

💡 Key Insights

• High accuracy ≠ strategic usefulness
• Model architecture influences reasoning behavior
• SAS provides a bridge between ML models and real-world policy decisions
• LSTM models better capture long-term planning signals

---

🧪 Research Questions

• Can SHAP feature importance be mapped into strategic categories?
• Do different models produce different strategic alignment levels?
• Does SAS improve interpretability over raw SHAP?

---

🧾 Contributions

1. Proposed Strategic Alignment Score (SAS)
2. Demonstrated accuracy–alignment trade-off
3. Integrated explainable AI with energy policy evaluation
4. Compared tree-based vs deep learning models for strategic reasoning

---

⚙️ Installation

pip install -r requirements.txt