energy-trade/ML_IMPLEMENTATION.md

ML Implementation Strategy

Overview

This document outlines the machine learning components for the energy trading system, including gradient boosting price prediction models and Q-Learning reinforcement learning for battery optimization.

Package Location: backend/app/ml/ (integrated within the FastAPI backend)

Related Documents:

• BACKEND_IMPLEMENTATION.md - API endpoints, services, and ML integration layer
• FRONTEND_IMPLEMENTATION.md - React frontend that consumes backend API

Training Split Strategy: Time-based split (first 7 days train, next 1.5 days validation, last 1.5 days test) to prevent look-ahead bias and ensure realistic evaluation.

Data Source: ~/energy-test-data/data/processed/*.parquet

---

Architecture

┌──────────────────────────────────────────────────────────────┐
│                    ML Pipeline                                │
├──────────────────────────────────────────────────────────────┤
│                                                              │
│  ┌──────────────┬──────────────────────┬──────────────────┐ │
│  │  Feature     │                      │                  │ │
│  │  Engineering │  Model Training       │  Model Serving   │ │
│  │              │                      │                  │ │
│  │  - Lags      │  - XGBoost           │  - Load Model    │ │
│  │  - Rolling   │  - Q-Learning        │  - Predict       │ │
│  │  - Time      │  - Validation        │  - Return Action │ │
│  │  - Regions   │  - Evaluation        │                  │ │
│  └──────────────┴──────────────────────┴──────────────────┘ │
│                          │                                     │
│  ┌──────────────────────┴──────────────────────────────┐    │
│  │              Model Management                        │    │
│  │  - Versioning  - Persistence  - Registry  - Backup    │    │
│  └──────────────────────────────────────────────────────┘    │
└──────────────────────────────────────────────────────────────┘
                               │
                               ▼
┌──────────────────────────────────────────────────────────────┐
│              FastAPI Backend (app.services.ml_service)       │
│  /api/v1/models/train, /api/v1/models/predict               │
└──────────────────────────────────────────────────────────────┘

---

Project Structure

backend/app/ml/
├── __init__.py
│
├── features/
│   ├── __init__.py
│   ├── lag_features.py          # Price lag feature extraction
│   ├── rolling_features.py      # Rolling statistics
│   ├── time_features.py         # Time-of-day encoding
│   ├── regional_features.py     # Cross-region differentials
│   └── battery_features.py      # Battery state features
│
├── price_prediction/
│   ├── __init__.py
│   ├── model.py                 # XGBoost model wrapper
│   ├── trainer.py               # Training pipeline
│   ├── evaluator.py             # Evaluation metrics
│   └── predictor.py             # Prediction interface
│
├── rl_battery/
│   ├── __init__.py
│   ├── environment.py           # Battery MDP environment
│   ├── agent.py                 # Q-Learning agent
│   ├── trainer.py               # RL training loop
│   └── policy.py                # Policy inference
│
├── model_management/
│   ├── __init__.py
│   ├── registry.py              # Model registry
│   ├── persistence.py           # Save/load models
│   ├── versioning.py            # Version handling
│   └── comparison.py            # Model comparison
│
├── evaluation/
│   ├── __init__.py
│   ├── metrics.py               # Common evaluation metrics
│   ├── backtest_evaluator.py   # Backtest performance evaluation
│   └── reports.py               # Generate evaluation reports
│
├── training/
│   ├── __init__.py
│   ├── cli.py                   # CLI commands for retraining
│   ├── pipeline.py              # End-to-end training pipeline
│   └── scheduler.py             # Training job scheduler
│
└── utils/
    ├── __init__.py
    ├── data_split.py            # Time-based data splitting
    ├── preprocessing.py         # Data preprocessing
    ├── config.py                # ML configuration
    └── evaluation.py            # Evaluation metrics

---

Configuration

backend/app/ml/utils/config.py

from dataclasses import dataclass
from typing import List, Dict, Any
from pathlib import Path

@dataclass
class PricePredictionConfig:
    """Configuration for price prediction models."""

    # Data
    data_path: str = "~/energy-test-data/data/processed"
    target_column: str = "real_time_price"

    # Training split (time-based)
    train_end_pct: float = 0.70  # First 70% for training
    val_end_pct: float = 0.85    # Next 15% for validation
    # Last 15% for testing

    # Features
    price_lags: List[int] = None
    rolling_windows: List[int] = None
    include_time_features: bool = True
    include_regional_features: bool = True

    # Model
    n_estimators: int = 200
    max_depth: int = 6
    learning_rate: float = 0.1
    subsample: float = 0.8
    colsample_bytree: float = 0.8
    random_state: int = 42

    # Early stopping
    early_stopping_rounds: int = 20
    early_stopping_threshold: float = 0.001

    def __post_init__(self):
        if self.price_lags is None:
            self.price_lags = [1, 5, 10, 15, 30, 60]
        if self.rolling_windows is None:
            self.rolling_windows = [5, 10, 15, 30, 60]


@dataclass
class RLBatteryConfig:
    """Configuration for RL battery optimization."""

    # State space
    charge_level_bins: int = 10
    price_bins: int = 10
    time_bins: int = 24  # Hours

    # Action space
    actions: List[str] = None

    # Q-Learning
    learning_rate: float = 0.1
    discount_factor: float = 0.95
    epsilon: float = 1.0
    epsilon_decay: float = 0.995
    epsilon_min: float = 0.05

    # Training
    episodes: int = 1000
    max_steps: int = 14400  # 10 days * 1440 minutes

    # Battery constraints
    min_reserve: float = 0.10  # 10%
    max_charge: float = 0.90  # 90%
    efficiency: float = 0.90

    # Reward scaling
    reward_scale: float = 1.0

    def __post_init__(self):
        if self.actions is None:
            self.actions = ["charge", "hold", "discharge"]


@dataclass
class MLConfig:
    """Overall ML configuration."""

    # Paths
    models_path: str = "models"
    results_path: str = "results"

    # Price prediction
    price_prediction: PricePredictionConfig = None
    rl_battery: RLBatteryConfig = None

    # Training
    enable_gpu: bool = False
    n_jobs: int = 4
    verbose: bool = True

    # Retraining
    keep_backup: bool = True
    max_backups: int = 5

    def __post_init__(self):
        if self.price_prediction is None:
            self.price_prediction = PricePredictionConfig()
        if self.rl_battery is None:
            self.rl_battery = RLBatteryConfig()


# Default configuration
default_config = MLConfig()

---

Feature Engineering Interface

Key Functions (backend/app/ml/features/init.py)

def build_price_features(
    df: pd.DataFrame,
    price_col: str = "real_time_price",
    lags: List[int] = None,
    windows: List[int] = None,
    regions: List[str] = None,
    include_time: bool = True,
    include_regional: bool = True,
) -> pd.DataFrame:
    """
    Build complete feature set for price prediction.

    Args:
        df: Input DataFrame
        price_col: Name of price column
        lags: List of lag periods
        windows: List of rolling window sizes
        regions: List of regions for differential features
        include_time: Whether to include time features
        include_regional: Whether to include regional features

    Returns:
        DataFrame with all features
    """


def build_battery_features(
    df: pd.DataFrame,
    price_df: pd.DataFrame,
    battery_col: str = "charge_level_mwh",
    capacity_col: str = "capacity_mwh",
    timestamp_col: str = "timestamp",
    battery_id_col: str = "battery_id",
) -> pd.DataFrame:
    """
    Build features for battery RL model.

    Args:
        df: Battery DataFrame
        price_df: Price DataFrame
        battery_col: Name of battery charge level column
        capacity_col: Name of battery capacity column
        timestamp_col: Name of timestamp column
        battery_id_col: Name of battery ID column

    Returns:
        DataFrame with battery features
    """

---

Price Prediction Interface

PricePredictor (backend/app/ml/price_prediction/predictor.py)

class PricePredictor:
    """Interface for making price predictions."""

    def __init__(self, models_dir: str = "models/price_prediction"):
        """
        Initialize predictor.

        Args:
            models_dir: Directory containing trained models
        """

    def predict(
        self,
        current_data: pd.DataFrame,
        horizon: int = 15,
        region: Optional[str] = None,
    ) -> float:
        """
        Predict price for a specific horizon.

        Args:
            current_data: Current/historical price data
            horizon: Prediction horizon in minutes
            region: Specific region to predict for (optional)

        Returns:
            Predicted price
        """

    def predict_all_horizons(
        self,
        current_data: pd.DataFrame,
        region: Optional[str] = None,
    ) -> Dict[int, float]:
        """
        Predict prices for all available horizons.

        Returns:
            Dictionary mapping horizons to predictions
        """

    def predict_with_confidence(
        self,
        current_data: pd.DataFrame,
        horizon: int = 15,
        region: Optional[str] = None,
    ) -> Dict:
        """
        Predict price with confidence interval.

        Returns:
            Dictionary with prediction and confidence interval
        """

    def get_feature_importance(self, horizon: int) -> pd.DataFrame:
        """
        Get feature importance for a specific horizon.

        Returns:
            DataFrame with feature importance
        """

PricePredictionTrainer (backend/app/ml/price_prediction/trainer.py)

class PricePredictionTrainer:
    """Training pipeline for price prediction models."""

    def __init__(self, config: PricePredictionConfig = None):
        """Initialize trainer."""

    def load_data(self) -> pd.DataFrame:
        """Load price data."""

    def prepare_data(self, df: pd.DataFrame) -> Tuple[pd.DataFrame, List[str]]:
        """
        Prepare data with features.

        Returns:
            Tuple of (features DataFrame, feature names)
        """

    def train_for_horizon(
        self,
        df_features: pd.DataFrame,
        feature_cols: List[str],
        horizon: int,
    ) -> Dict:
        """
        Train model for a specific horizon.

        Returns:
            Training results dictionary with metrics
        """

    def train_all(self, horizons: List[int] = None) -> Dict:
        """
        Train models for all horizons.

        Returns:
            Dictionary with all training results
        """

    def save_models(self, output_dir: str = "models/price_prediction") -> None:
        """Save all trained models."""

    @classmethod
    def load_models(
        cls,
        models_dir: str = "models/price_prediction",
        horizons: List[int] = None,
    ) -> Dict[int, PricePredictionModel]:
        """
        Load trained models.

        Returns:
            Dictionary mapping horizons to models
        """

---

RL Battery Optimization Interface

BatteryPolicy (backend/app/ml/rl_battery/policy.py)

class BatteryPolicy:
    """Interface for RL battery policy inference."""

    def __init__(self, policy_path: str = "models/rl_battery"):
        """
        Initialize policy.

        Args:
            policy_path: Path to trained policy
        """

    def get_action(
        self,
        charge_level: float,
        current_price: float,
        price_forecast_1m: float = 0,
        price_forecast_5m: float = 0,
        price_forecast_15m: float = 0,
        hour: int = 0,
    ) -> Dict:
        """
        Get action for current state.

        Returns:
            Dictionary with action, q_values
        """

BatteryRLTrainer (backend/app/ml/rl_battery/trainer.py)

class BatteryRLTrainer:
    """Training pipeline for RL battery policy."""

    def __init__(self, config: RLBatteryConfig = None):
        """Initialize trainer."""

    def load_data(self) -> None:
        """Load price data for environment."""

    def train(self, n_episodes: int = 1000, region: str = "FR") -> Dict:
        """
        Train RL agent.

        Returns:
            Training results with metrics
        """

    def save(self, output_dir: str = "models/rl_battery") -> None:
        """Save trained policy."""

---

Model Management Interface

ModelRegistry (backend/app/ml/model_management/registry.py)

class ModelRegistry:
    """Registry for tracking model versions."""

    def __init__(self, registry_path: str = "models/registry.json"):
        """Initialize registry."""

    def register_model(
        self,
        model_type: str,
        model_id: str,
        version: str,
        filepath: str,
        metadata: Dict = None,
    ) -> None:
        """Register a model version."""

    def get_latest_version(self, model_id: str) -> Optional[Dict]:
        """Get latest version of a model."""

    def list_models(self) -> List[Dict]:
        """List all registered models."""

---

Training Task Interface

Training Jobs (app/tasks/training_tasks.py)

async def train_model_task(training_id: str, request: TrainingRequest):
    """
    Execute ML model training via Celery task.

    Dispatches to Celery for async processing of:
    - Price prediction training
    - RL battery policy training

    Emits WebSocket events for progress updates.
    """

@shared_task(name="tasks.train_price_prediction")
def train_price_prediction(training_id: str, request_dict: dict):
    """
    Celery task for price prediction model training.

    Process:
    1. Load and prepare data
    2. Train XGBoost models for specified horizon
    3. Save models
    4. Register in model registry
    5. Update training job status
    """

@shared_task(name="tasks.train_rl_battery")
def train_rl_battery(training_id: str, request_dict: dict):
    """
    Celery task for RL battery policy training.

    Process:
    1. Load environment and data
    2. Train Q-Learning agent
    3. Save policy
    4. Register in model registry
    5. Update training job status
    """

---

Key Integration Points

ML Service Integration (app/services/ml_service.py)

The ML service provides the bridge between API routes and ML models:

class MLService:
    """Service for ML model management and inference."""

    def list_models(self) -> List[ModelInfo]:
        """List all available trained models."""

    def get_model_metrics(self, model_id: str) -> Dict[str, float]:
        """Get performance metrics for a model."""

    def predict(
        self,
        model_id: str,
        timestamp: datetime,
        features: Optional[Dict[str, Any]] = None
    ) -> Dict[str, Any]:
        """
        Run prediction with on-demand model loading.

        Supports:
        - Price prediction models
        - RL battery policy
        """

    def get_feature_importance(self, model_id: str) -> Dict[str, float]:
        """Get feature importance for a model."""

WebSocket Events

Real-time training progress updates:

# Event types
- "model_training_progress"  # Training progress updates

# Payload
{
    "model_id": str,
    "progress": float,      # 0.0 to 1.0
    "epoch": int,           # Current epoch (optional)
    "metrics": dict         # Current metrics
}

---

Data Flow

Training Pipeline

API Request (POST /api/v1/models/train)
    ↓
training_tasks.train_model_task()
    ↓
Celery Task (train_price_prediction / train_rl_battery)
    ↓
PricePredictionTrainer / BatteryRLTrainer
    ↓
Feature Engineering → Model Training → Evaluation
    ↓
Save Models → Register in Registry
    ↓
WebSocket Events (progress updates)
    ↓
Update Training Status

Prediction Pipeline

API Request (POST /api/v1/models/predict)
    ↓
ml_service.predict()
    ↓
Load Model (on-demand)
    ↓
Feature Engineering
    ↓
Model.predict() / Policy.get_action()
    ↓
Return Prediction with Confidence

---

Model Artifacts

Price Prediction Models

Location: models/price_prediction/

model_1min.pkl       # 1-minute horizon model
model_5min.pkl       # 5-minute horizon model
model_15min.pkl      # 15-minute horizon model
model_60min.pkl      # 60-minute horizon model
training_results.json  # Training metrics and metadata

RL Battery Policy

Location: models/rl_battery/

battery_policy.pkl       # Trained Q-Learning policy
training_results.json    # Training metrics

Model Registry

Location: models/registry.json

{
    "models": {
        "price_prediction_15m": {
            "type": "price_prediction",
            "versions": ["v20260211_134500", "v20260210_100000"],
            "latest": "v20260211_134500"
        },
        "battery_policy": {
            "type": "rl_battery",
            "versions": ["v20260211_140000"],
            "latest": "v20260211_140000"
        }
    }
}