Edge Layer

The Edge Layer hosts Predictive AI models on low-power compute devices attached directly to energy assets, performing short-term forecasts and maintaining local data resilience.

Overview

The Edge Layer is the foundation of the Ona Platform’s distributed intelligence architecture. It runs on low-power compute devices attached directly to energy assets, enabling real-time forecasting and fault prediction at the source. This layer ensures operational continuity even when connectivity to the central control layer is interrupted.

Key Capabilities:

• Hosts Predictive AI models for short-term forecasting (0-48 hours)
• Performs local model execution on each node
• Stores 48 hours of data locally for resilience
• Operates independently during connectivity loss
• Provides real-time predictions for each asset node

Forecasting Capabilities

Short-Term Forecasts (0-48 Hours)

The Edge Layer generates energy production forecasts with a horizon of up to 48 hours. These forecasts are critical for:

• Operational Planning: Enabling operators to anticipate energy production
• Grid Integration: Supporting dispatch decisions and grid stability
• Maintenance Scheduling: Optimizing maintenance windows based on predicted production
• Financial Planning: Supporting energy trading and revenue forecasting

Forecast API Endpoint

GET /forecast

Generate energy production forecast for a customer’s assets.

Query Parameters:

• customer_id (required): Customer identifier
• site_id (optional): Specific site identifier
• horizon_hours (optional): Forecast horizon in hours (default: 48)

Example Request:

curl -X GET "https://api.asoba.co/forecast?customer_id=demo-customer&horizon_hours=48" \
  -H "X-API-Key: your-api-key"

Example Response:

{
  "forecast_id": "fc-demo-12345678",
  "customer_id": "demo-customer",
  "site_id": "demo-site-cape-town-01",
  "generated_at": "2025-01-15T12:00:00Z",
  "horizon_hours": 48,
  "forecasts": [
    {
      "timestamp": "2025-01-15T13:00:00Z",
      "predicted_power_kw": 1250.5,
      "confidence_interval": {
        "lower": 1180.2,
        "upper": 1320.8
      },
      "weather_conditions": {
        "temperature_c": 25.3,
        "irradiance_w_m2": 850.2,
        "wind_speed_m_s": 3.2
      }
    }
  ],
  "metadata": {
    "model_version": "v2.1.0",
    "training_data_points": 8760,
    "model_accuracy": 0.94
  }
}

Lambda Services

The Edge Layer is powered by several AWS Lambda functions that handle forecasting and data processing:

Forecasting API Service

ona-forecastingApi-prod

• Memory: 3008MB
• Timeout: 60 seconds
• Purpose: Generates energy production forecasts using trained ML models
• Capabilities:
  • Loads trained LSTM models from S3
  • Processes nowcast data for forecast generation
  • Integrates weather forecast data
  • Returns structured forecast responses with confidence intervals

Interpolation Service

ona-interpolationService-prod

• Memory: 3008MB
• Timeout: 900 seconds (15 minutes)
• Purpose: Fills data gaps and enriches time series data
• Capabilities:
  • ML-powered gap filling using adaptive multi-output methods
  • Data enrichment with weather features
  • Performance metrics calculation (RMSE, MAE, R²)
  • Handles missing telemetry data gracefully

Data Standardization Service

ona-dataStandardizationService-prod

• Memory: 1024MB
• Timeout: 300 seconds
• Purpose: Normalizes and standardizes data from various OEM sources
• Capabilities:
  • Converts OEM-specific formats to standardized schema
  • Handles multiple inverter manufacturers (Huawei, Enphase, etc.)
  • Validates data quality and completeness

Data Storage and Resilience

Local Data Buffer

Each edge node maintains a 48-hour local data buffer to ensure:

• Resilience: Continued operation during connectivity loss
• Performance: Reduced latency for local predictions
• Reliability: No data loss during network interruptions

Cloud Storage (S3)

Input Bucket: sa-api-client-input

• observations/: Raw sensor data from assets
• nowcast/: Real-time data for forecasting
• historical/: Historical data for model training

Output Bucket: sa-api-client-output

• forecasts/: Generated forecast results
• models/: Trained ML model artifacts
• training_data/: Processed training datasets

DynamoDB Tables

ona-platform-locations

• Stores location and customer data
• Supports geographic data queries
• Enables location-based weather data retrieval

ona-platform-weather-cache

• Caches weather data to reduce API calls
• Improves forecast generation performance
• Reduces external API dependencies

Local Model Execution

Model Deployment

Predictive AI models are deployed to edge devices and execute locally:

• LSTM Models: Long Short-Term Memory networks for time series forecasting
• Model Versioning: Supports multiple model versions for A/B testing
• Model Updates: Can be updated remotely without device downtime
• Resource Efficiency: Optimized for low-power compute devices

Model Performance

Typical Model Metrics:

• Training RMSE: 0.072-0.082
• Validation RMSE: 0.089-0.096
• MAPE: 5.4-6.8%
• SMAPE: 6.8-8.2%
• Model Accuracy: 94%+

Model Artifacts

Models are stored in S3 with the following structure:

s3://sa-api-client-output/customer_tailored/{customer_id}/models/{version}/
├── model.h5              # Trained model weights
├── encoders.pkl          # Feature encoders
└── config.json           # Model configuration

Resilience and Independence

Graceful Degradation

Each edge node operates independently if the central connection fails:

• Forecasting Continues: Local models continue generating predictions
• Fault Detection Active: Anomaly detection runs locally
• Data Queuing: Decisions and data are queued for transmission
• Automatic Recovery: Queued data transmits automatically when connectivity is restored

Connectivity Requirements

• Minimum: Intermittent connectivity sufficient for data synchronization
• Optimal: Continuous connectivity for real-time updates
• Fallback: 48-hour local buffer ensures operation during outages

Integration with Control Layer

The Edge Layer communicates with the Control Layer through:

• Encrypted Channels: TLS 1.3 encryption for all communications
• Certificate-Based Authentication: Secure device authentication
• API Gateway: Centralized API endpoint management
• Data Synchronization: Automatic sync when connectivity is restored

Performance Characteristics

Latency

• Forecast Generation: < 5 seconds per node
• Data Processing: < 1 second per data point
• Model Loading: < 2 seconds (cached after first load)

Throughput

• Forecasts per Hour: 1000+ per node
• Data Points Processed: 10,000+ per minute
• Concurrent Requests: 20+ simultaneous forecasts

Resource Usage

• CPU: Optimized for low-power ARM processors
• Memory: 512MB-3GB depending on model complexity
• Storage: 48-hour buffer requires ~100MB per node

Next Steps

• Control Layer - Learn about the central coordination layer
• Interface Layer - Explore the user interface and dashboards
• User Guide - Get started with Ona Platform deployment

Get Help & Stay Updated