🤖 AI Methodology Documentation

EcoPower AI utilizes a Hybrid AI Engine that combines the power of Generative AI (Google Gemini 2.5 Flash) with robust statistical forecasting.

This approach ensures you get the deep reasoning and pattern recognition capabilities of a Large Language Model (LLM), backed by a failsafe statistical model to ensure service reliability even without internet connectivity.

Purpose

Identify recurring consumption patterns in your historical energy data.

How It Works

1. Analyzes the last 30 days of hourly energy consumption data
2. Groups data by hour of day (0-23) and day of week (Monday-Sunday)
3. Calculates average consumption for each hour across all days
4. Identifies peak usage times and low-demand periods

Example

If you consistently use 3.0 kWh between 12 PM - 2 PM (peak solar hours), the AI detects this pattern and uses it for future predictions.

Code Implementation

// Build hourly pattern averages
foreach ($historical_data as $record) {
    $hour = $record['hour_of_day'];
    $consumption = floatval($record['consumption_kwh']);
    
    if (!isset($hourly_patterns[$hour])) {
        $hourly_patterns[$hour] = [];
    }
    $hourly_patterns[$hour][] = $consumption;
}

// Calculate average consumption per hour
foreach ($hourly_patterns as $hour => $consumptions) {
    $hourly_averages[$hour] = array_sum($consumptions) / count($consumptions);
}

Purpose

Detect whether your overall energy consumption is increasing, decreasing, or stable.

How It Works

1. Calculates total daily consumption for the last 7 days
2. Applies linear regression to detect trend slope
3. Classifies trend as: Increasing (slope > 0.5), Decreasing (slope < -0.5), or Stable
4. Adjusts future predictions based on trend direction

Formula

Linear regression slope: slope = Σ[(x - x̄)(y - ȳ)] / Σ[(x - x̄)²]

Impact on Predictions

If consumption is trending upward, predictions are adjusted slightly higher. If trending downward, predictions are adjusted lower. This ensures predictions adapt to changing usage patterns.

Purpose

Smooth out short-term fluctuations and predict future consumption.

How It Works

1. Takes the last 7 days (168 hours) of consumption data
2. Calculates weighted moving average
3. Uses this average as baseline for predictions
4. Combines with hourly patterns for more accurate forecasts

Formula

Moving Average = Σ(consumption values) / number of values

Why Moving Averages?

Moving averages reduce noise in data and provide a stable baseline for predictions. They're particularly effective for energy consumption, which has regular daily patterns.

Purpose

Identify unusual energy consumption spikes that deviate from normal patterns.

How It Works

1. Calculates mean (average) consumption across all historical data
2. Calculates standard deviation (σ) to measure variability
3. Sets anomaly threshold at: mean + 2σ
4. Flags any consumption above this threshold as an anomaly

Statistical Basis

In a normal distribution, 95% of values fall within 2 standard deviations of the mean. Values beyond this range are statistically unusual and warrant investigation.

Use Cases

• Detect equipment malfunctions (e.g., faulty appliances consuming excess power)
• Identify unusual usage patterns (e.g., forgot to turn off high-power devices)
• Alert users to potential energy waste

Purpose

Recommend the optimal energy source (solar, grid, or generator) for each time period to minimize fuel consumption and costs.

Decision Logic

Fuel Savings Calculation

Fuel Saved = (Predicted kWh) × 0.3 L/kWh
Based on average diesel generator consumption rate of 0.3 liters per kWh

Cost Savings Calculation

Cost Savings = (Generator Cost - Recommended Source Cost) × Predicted kWh
Where: Generator = ₦0.35/kWh, Grid = ₦0.15/kWh, Solar = ₦0.05/kWh

CO₂ Emission Calculation

CO₂ Reduced (kg) = Fuel Saved (liters) × 2.68 kg CO₂/liter
Based on EPA standard emission factor for diesel fuel

Baseline Comparison

All savings are calculated against a baseline scenario where 100% of energy comes from a generator. This represents the worst-case environmental impact.

Real-World Equivalents

• Trees Planted: CO₂ Reduced ÷ 21 kg (average CO₂ absorbed by one tree per year)
• Car Kilometers: CO₂ Reduced ÷ 0.12 kg/km (average car emissions)

How Confidence is Calculated

Each prediction includes a confidence score (0-100%) based on:

1. Data Availability: More historical data points = higher confidence
2. Pattern Stability: Consistent patterns = higher confidence
3. Trend Volatility: Stable trends = higher confidence

Formula

Base Confidence = min(95, 60 + (data_points × 2))
Final Confidence = Base Confidence - (trend_volatility_penalty)

Interpretation

• 80-95%: High confidence - strong historical patterns
• 70-79%: Good confidence - adequate data available
• 50-69%: Moderate confidence - limited data or volatile patterns

Minimum Data Requirements

• For Predictions: At least 24 hours of historical data
• For Accurate Patterns: 7+ days of data recommended
• For Trend Analysis: 7+ days required

Limitations

• Predictions assume future patterns will resemble historical patterns
• Cannot predict sudden changes in usage (e.g., new appliances, seasonal changes)
• Solar recommendations assume functional solar panels with adequate capacity
• Grid recommendations assume grid availability and stability

Best Practices

• Input data regularly (daily or hourly) for best results
• Include notes for unusual consumption periods
• Review and update predictions weekly
• Adjust recommendations based on actual grid/solar availability

Transparency Commitment

EcoPower AI is committed to full transparency in AI decision-making. Every prediction, recommendation, and calculation is based on documented statistical methods. No hidden algorithms, no proprietary black boxes.

All source code is available for review, and we provide detailed explanations for every AI output. If you have questions about how a specific recommendation was generated, please contact us.