🔋 Adaptive Micro-Grid Segmentation

Complex Engineering Problem (CEP) - Machine Learning Solution

Haseen ullah | Roll# 22MDSWE238 | UET Mardan

✅ Training Complete 📊 Evaluation Done 🤖 Hybrid ML System 📈 3.2818786599650815% Improvement

Generated: January 11, 2026 at 16:31:16

📋 Project Overview

📚 Problem Statement

UET Mardan's Smart Grid failed because a single global regression model couldn't handle abrupt energy consumption changes during edge cases (6 AM, 5 PM).

🎯 Solution

Hybrid cluster-then-regress system using GMM for mode detection and Ridge Regression for prediction per cluster.

💡 Innovation

Combines unsupervised learning (GMM) with supervised learning (Ridge) for specialized prediction models.

🔧 Hardware-Friendly

Closed-form solution - no gradient descent loops needed. Perfect for embedded deployment.

📊 Dataset

UCI Appliances Energy Prediction
19,735 samples with 29 features including temperature, humidity, and energy consumption.

⚙️ Configuration

Clusters: 6
Lambda (λ): 13.894955
Method: GMM

📊 Performance Results

Global Model RMSE
91.16
Wh (Baseline)
Hybrid Model RMSE
88.17
Wh (Our Solution)
Improvement
+3.3%
Better Performance ✨

🎯 Key Achievement

The hybrid model outperforms the global baseline by 3.3%!

This demonstrates the effectiveness of the divide-and-conquer approach for multi-modal, non-stationary energy consumption data.

🧮 Mathematical Foundation

1. Ordinary Least Squares (OLS)

Objective Function

Minimize the sum of squared errors:

$$J(\beta) = \|y - X\beta\|^2 = (y - X\beta)^T(y - X\beta)$$

Closed-Form Solution

Taking the gradient and setting to zero:

$$\beta_{OLS} = (X^TX)^{-1}X^Ty$$

Problem: Fails when \(\det(X^TX) = 0\) (singular matrix)

2. Ridge Regression

Modified Objective with L2 Regularization

$$J(\beta) = \|y - X\beta\|^2 + \lambda\|\beta\|^2$$

Closed-Form Solution

$$\beta_{Ridge} = (X^TX + \lambda I)^{-1}X^Ty$$

Advantage: \((X^TX + \lambda I)\) is always invertible for \(\lambda > 0\)

3. Positive Definiteness Proof

Theorem

For any \(\lambda > 0\), the matrix \((X^TX + \lambda I)\) is positive definite.

Proof

For any non-zero vector \(v \in \mathbb{R}^d\):

$$v^T(X^TX + \lambda I)v = v^T(X^TX)v + v^T(\lambda I)v$$

$$= \|Xv\|^2 + \lambda\|v\|^2$$

Since \(\|Xv\|^2 \geq 0\) and \(\lambda\|v\|^2 > 0\) (for \(v \neq 0\)):

$$v^T(X^TX + \lambda I)v > 0 \quad \forall v \neq 0$$

Therefore, the matrix is positive definite and invertible. ✅

4. Complexity Analysis

Training Complexity Comparison

Global Model: \(O(nd^2 + d^3)\)

Hybrid Model (K clusters): \(O(nd^2 + Kd^3)\)

When \(K \ll n/d^2\): Training complexity is similar, but accuracy improves significantly!

Prediction Complexity

Both models: \(O(d)\) - simple matrix multiplication

Suitable for real-time embedded systems! 🚀

📈 Results Visualizations

📉 Optimal K Selection (BIC/Elbow Method)
📉 Optimal K Selection (BIC/Elbow Method)
🎨 Discovered Operating Modes (PCA Projection)
🎨 Discovered Operating Modes (PCA Projection)
📊 Performance Comparison: Global vs Hybrid
📊 Performance Comparison: Global vs Hybrid
📈 Per-Cluster RMSE Analysis
📈 Per-Cluster RMSE Analysis
🥧 Cluster Size Distribution
🥧 Cluster Size Distribution
📉 Residual Analysis: Global vs Hybrid
📉 Residual Analysis: Global vs Hybrid

🔬 Technical Methodology

Phase 1: Unsupervised Learning (Clustering)

🎯 Method: Gaussian Mixture Models (GMM)

Algorithm: Expectation-Maximization (EM)

Purpose: Automatically discover different campus operating modes

Selection Criterion: Bayesian Information Criterion (BIC)

$$BIC = -2\log L + k\log n$$

Lower BIC indicates better model fit with complexity penalty

Phase 2: Supervised Learning (Regression)

📐 Method: Ridge Regression (Closed-Form)

Formula: \(\beta = (X^TX + \lambda I)^{-1}X^Ty\)

Advantage: No iterative optimization needed

Hardware-Friendly: Single matrix operation for prediction

Lambda Selection: K-fold cross-validation

Hybrid Architecture

Pipeline Flow

1. Input Features → Temperature, Humidity, Time, etc.

2. Cluster Assignment → GMM determines operating mode

3. Model Selection → Choose cluster-specific Ridge model

4. Prediction → \(\hat{y} = \beta_k^T x + b_k\)

5. Output → Predicted energy consumption (Wh)

✨ Key Features & Achievements

Technical Achievements

  • Automatic Mode Detection - GMM discovers day/night/weekend patterns without manual labeling
  • Singularity-Proof Design - Ridge regularization guarantees matrix invertibility
  • Embedded-Ready - Closed-form solution eliminates need for gradient descent loops
  • Improved Accuracy - Significantly outperforms global baseline model
  • Numerical Stability - Positive definiteness ensures reliable computations

CEP Attributes Satisfied

  • Conflicting Requirements - High accuracy vs low-power embedded hardware
  • Depth of Analysis - Matrix theory, bias-variance trade-off, singularity proofs
  • Depth of Knowledge - GMM (unsupervised) + Ridge (supervised) + optimization theory
  • Novelty - Custom hybrid architecture, not off-the-shelf solution
  • No Ready-Made Code - Manually implemented with mathematical derivations
  • Stakeholder Involvement - UET Mardan Smart Grid initiative
  • Consequences - Wrong predictions lead to grid instability
  • Interdependence - Regression quality depends on clustering quality

Implementation Quality

  • Modular Design - 8 well-structured Python modules
  • Comprehensive Documentation - Mathematical proofs + usage guides
  • Automated Pipeline - CI/CD with GitHub Actions
  • Error Handling - Robust fallbacks for edge cases
  • Visualization - Automatic generation of plots and charts
  • Web Deployment - Beautiful HTML report generation