ML-Powered Renewable Energy Trading Platform

An end-to-end automated energy trading system that predicts ERCOT electricity prices using ensemble ML models and makes buy/sell decisions informed by RAG-based market intelligence.

Live Demo: [Coming Soon]

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    ML-Powered Energy Trading Platform                │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│   Open-Meteo API ──→ ┌──────────────┐                               │
│   (Dallas weather)    │    Data      │                               │
│                       │   Service    │──→ Weather + ERCOT data       │
│   ERCOT API ────────→ │   :8001      │                               │
│   (market prices)     └──────┬───────┘                               │
│                              │                                       │
│                              ▼                                       │
│                       ┌──────────────┐    ┌──────────────┐          │
│                       │  Forecast    │    │     RAG      │          │
│                       │  Service     │    │   Service    │          │
│                       │  :8002       │    │   :8004      │          │
│                       │              │    │              │          │
│                       │ RF + XGBoost │    │ LangChain +  │          │
│                       │  Ensemble    │    │ FAISS +      │          │
│                       │              │    │ Gemini       │          │
│                       └──────┬───────┘    └──────┬───────┘          │
│                              │                   │                   │
│                              ▼                   ▼                   │
│                       ┌──────────────────────────────┐              │
│                       │      Trading Service         │              │
│                       │          :8003               │              │
│                       │                              │              │
│                       │  ML prediction + RAG risk    │              │
│                       │  → Adjusted thresholds       │              │
│                       │  → Confidence-based sizing   │              │
│                       │  → Position management       │              │
│                       │  → SQLite persistence        │              │
│                       └──────────────┬───────────────┘              │
│                                      │                               │
│                                      ▼                               │
│                       ┌──────────────────────────────┐              │
│                       │    Streamlit Dashboard        │              │
│                       │         :8501                │              │
│                       │                              │              │
│                       │  Price chart with CI bands   │              │
│                       │  Trade history + P&L         │              │
│                       │  RAG insights + Weather      │              │
│                       │  Model metrics + Importance  │              │
│                       └──────────────────────────────┘              │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

Key Features

Real-Time Data Ingestion

Live weather data from Open-Meteo (NOAA HRRR model) for Dallas, TX — real GHI, DNI, direct radiation
Real ERCOT settlement point prices from HB_NORTH hub every 15 minutes
Graceful fallback to synthetic data when APIs are unavailable

Ensemble ML Forecasting

RandomForest + XGBoost averaged via VotingRegressor
Bootstrap confidence intervals (30 iterations) for uncertainty quantification
Permutation-based feature importance analysis
Naive baseline comparison for model validation

RAG Market Intelligence

LangChain orchestration with FAISS vector store
Google Gemini embeddings + chat for document retrieval and synthesis
Ingests real EIA energy articles via RSS and ERCOT grid conditions
Returns structured risk assessments that adjust trading behavior

Intelligent Trading Logic

RAG-adjusted thresholds: high upward risk lowers sell threshold
Confidence-based position sizing: trade more when model is certain
Full position tracking with weighted average cost basis
Profit calculation, win rate tracking, trade audit trail
SQLite persistence for trade history and portfolio state

Production Architecture

5 Docker microservices communicating via REST APIs
Health check endpoints on every service
Environment-based configuration for local and cloud deployment
Graceful degradation: each service handles failures independently

Tech Stack

Layer	Technologies
ML Models	scikit-learn, XGBoost, RandomForest, VotingRegressor
RAG Pipeline	LangChain, FAISS, Google Gemini (embeddings + chat)
Backend APIs	FastAPI, Uvicorn
Database	SQLite
Data Sources	Open-Meteo API, ERCOT Public API, EIA RSS Feed
Frontend	Streamlit, Plotly
Infrastructure	Docker, Docker Compose
Language	Python 3.10

Performance Metrics

Metric	Value
Demand Forecast MAPE	~12%
Price Forecast MAPE	~14%
Energy Output MAPE	~10%
Improvement Over Naive Baseline	5-8 percentage points
Orchestration Interval	15 minutes (aligned with ERCOT RTM)
Time Intervals Tested	1,500+

Project Structure

Renewable_AI_Trading/
├── data_service/                  # Weather + ERCOT data ingestion
│   ├── main.py                    # Open-Meteo + ERCOT API endpoints
│   ├── Dockerfile
│   └── requirements.txt
│
├── forecast_service/              # Ensemble ML predictions
│   ├── main.py                    # FastAPI service with prediction endpoints
│   ├── ensemble_forecaster.py     # RF + XGBoost + confidence intervals
│   ├── Dockerfile
│   └── requirements.txt
│
├── rag_service/                   # RAG market intelligence
│   ├── main.py                    # FastAPI endpoints for analysis
│   ├── rag_engine.py              # LangChain + FAISS + Gemini pipeline
│   ├── energy_documents.py        # ERCOT domain knowledge base
│   ├── news_fetcher.py            # Real EIA + ERCOT news ingestion
│   ├── Dockerfile
│   └── requirements.txt
│
├── trading_service/               # Automated trading decisions
│   ├── main.py                    # Trading logic with RAG integration
│   ├── database.py                # SQLite for trades + positions
│   ├── Dockerfile
│   └── requirements.txt
│
├── dashboard/                     # Streamlit visualization
│   ├── app.py                     # Full dashboard with charts + metrics
│   ├── Dockerfile
│   └── requirements.txt
│
├── src/
│   └── train_ensemble.py          # Train RF + XGBoost ensemble models
│
├── models/                        # Trained model files (.joblib)
├── data/                          # Training data + SQLite database
├── docker-compose.yml             # Orchestrates all services
└── .env.example                   # Environment variables template

Quick Start

Prerequisites: Docker, Docker Compose, Python 3.10+

# 1. Clone the repo
git clone https://github.com/aloksharma3/Renewable-Energy-Trading-Platform.git
cd Renewable-Energy-Trading-Platform/Renewable_AI_Trading

# 2. Set up environment variables
cp .env.example .env
# Edit .env and add your GEMINI_API_KEY (free from https://aistudio.google.com/apikey)

# 3. Train the ML models
pip install scikit-learn xgboost joblib pandas numpy
python src/train_ensemble.py

# 4. Start all services
docker-compose up --build

# 5. Open the dashboard
# http://localhost:8501

How Trading Decisions Are Made

Every 15 minutes:

1. DATA SERVICE fetches:
   → Dallas weather from Open-Meteo (temp, wind, GHI, DNI)
   → ERCOT HB_NORTH settlement point price

2. FORECAST SERVICE predicts:
   → Energy output, price, and demand using RF + XGBoost ensemble
   → Returns confidence intervals via bootstrap sampling

3. RAG SERVICE analyzes:
   → Searches FAISS for relevant market documents
   → Generates risk assessment via Gemini
   → Returns risk_score (0-1) and price_direction (up/down/stable)

4. TRADING SERVICE decides:
   → Adjusts thresholds: high upward risk → lower sell threshold
   → Sizes trade: tight confidence → trade more
   → Checks position limits: can't sell more than held
   → Executes trade and logs to SQLite with full reasoning

Environment Variables

# Required
GEMINI_API_KEY=your_gemini_api_key    # Free from Google AI Studio

# Optional (defaults shown)
LAT=32.7767                           # Dallas, TX
LON=-96.7970
SELL_THRESHOLD=55                     # Base sell threshold ($/MWh)
BUY_THRESHOLD=30                      # Base buy threshold ($/MWh)
MAX_POSITION_MWH=500                  # Maximum position size
LOG_LEVEL=INFO

Why These Design Decisions

Decision	Reasoning
ERCOT HB_NORTH	Highest demand zone in Texas, matches Dallas weather location
15-minute intervals	Aligned with ERCOT Real-Time Market settlement periods
Open-Meteo over OpenWeatherMap	Real GHI/DNI solar irradiance, free, no API key
Ensemble over single model	Errors cancel out, more robust predictions
RAG over static rules	Markets driven by events ML can't see in numbers
SQLite over PostgreSQL	Zero-setup for MVP, same interface for production swap
Microservices over monolith	Independent scaling, fault isolation, production-realistic

Author

Alok Sharma

M.S. Applied Machine Intelligence, Northeastern University
5+ years in SCADA/EMS systems and industrial automation at Schneider Electric, L&T, NCDEX
LinkedIn | GitHub

aloksharma3/Renewable-Energy-Trading-Platform