@neural-trader/example-energy-forecasting

@neural-trader/example-energy-forecasting

Self-learning energy forecasting with conformal prediction and swarm-based ensemble models.

Features

• Multi-Model Ensemble: ARIMA, LSTM, Transformer, and Prophet models
• Conformal Prediction: Guaranteed coverage prediction intervals with @neural-trader/predictor
• Swarm-Based Optimization: Parallel hyperparameter exploration for optimal model selection
• Adaptive Learning: Online model updates and horizon-specific model weighting
• Seasonal Pattern Detection: Automatic detection and modeling of seasonal components
• Dual Implementation: Pure TypeScript with automatic WASM/Native acceleration via @neural-trader/predictor
• Uncertainty Quantification: Statistically rigorous prediction intervals with configurable confidence levels

Supported Domains

• ☀️ Solar Generation: Photovoltaic power forecasting with daily cycles
• 💨 Wind Power: Wind turbine generation with high variability
• ⚡ Electricity Demand: Load forecasting with multi-seasonal patterns
• 🌡️ Temperature: Weather prediction with seasonal trends

Installation

npm install @neural-trader/example-energy-forecasting

Quick Start

import { EnergyForecaster, EnergyDomain, ModelType } from '@neural-trader/example-energy-forecasting';

// Create forecaster
const forecaster = new EnergyForecaster(EnergyDomain.SOLAR, {
  alpha: 0.1,              // 90% confidence intervals
  horizon: 48,             // 48-hour forecast
  seasonalPeriod: 24,      // Daily seasonality
  enableAdaptive: true,
  ensembleConfig: {
    models: [ModelType.ARIMA, ModelType.LSTM, ModelType.PROPHET]
  }
});

// Train with historical data
const historicalData = [
  { timestamp: Date.now(), value: 150 },
  // ... more data points
];
await forecaster.train(historicalData);

// Generate forecast
const forecast = await forecaster.forecast(48);

// Access predictions with confidence intervals
forecast.forecasts.forEach(f => {
  console.log(`Time: ${new Date(f.timestamp).toISOString()}`);
  console.log(`Forecast: ${f.pointForecast}`);
  console.log(`Interval: [${f.interval.lower}, ${f.interval.upper}]`);
  console.log(`Confidence: ${f.confidence * 100}%`);
});

Architecture

Core Components

1. EnergyForecaster

Main forecasting system that orchestrates ensemble models and conformal prediction.

const forecaster = new EnergyForecaster(EnergyDomain.SOLAR, config);
await forecaster.train(data);
const forecast = await forecaster.forecast(48);

2. EnsembleSwarm

Swarm-based hyperparameter optimization and adaptive model weighting.

const ensemble = new EnsembleSwarm({
  models: [ModelType.ARIMA, ModelType.LSTM],
  adaptiveLearningRate: 0.05,
  retrainFrequency: 100
});
await ensemble.trainEnsemble(trainingData, validationData);

3. EnergyConformalPredictor

Wrapper around @neural-trader/predictor for uncertainty quantification.

const predictor = new EnergyConformalPredictor(ModelType.ARIMA, {
  alpha: 0.1,
  calibrationSize: 2000
}, true); // adaptive mode

await predictor.initialize();
await predictor.calibrate(historicalData, predictions);

Model Implementations

All models implement the ForecastingModel interface:

• ARIMAModel: Exponential smoothing with trend
• LSTMModel: Recurrent patterns with attention-like weights
• TransformerModel: Self-attention mechanism
• ProphetModel: Trend + seasonality decomposition

Examples

Run the included examples:

# Solar generation forecasting
npm run example:solar

# Wind power forecasting
npm run example:wind

# Electricity demand forecasting
npm run example:demand

# Temperature prediction
npm run example:temperature

Configuration

ForecasterConfig

interface ForecasterConfig {
  alpha?: number;                    // Conformal prediction miscoverage rate (default: 0.1)
  calibrationSize?: number;          // Max calibration samples (default: 2000)
  horizon?: number;                  // Default forecast horizon (default: 24)
  seasonalPeriod?: number;           // Seasonal period in hours (default: 24)
  enableAdaptive?: boolean;          // Enable adaptive conformal prediction (default: true)
  ensembleConfig?: EnsembleConfig;   // Ensemble configuration
  weatherIntegration?: WeatherConfig; // Weather API integration
}

EnsembleConfig

interface EnsembleConfig {
  models: ModelType[];                              // Models to include in ensemble
  horizonWeights?: Map<number, Map<ModelType, number>>; // Custom horizon weights
  adaptiveLearningRate?: number;                    // Learning rate for weight updates (default: 0.05)
  retrainFrequency?: number;                        // Retraining frequency (default: 100)
  minCalibrationSamples?: number;                   // Min samples for calibration (default: 50)
}

Conformal Prediction

This package uses @neural-trader/predictor for statistically rigorous uncertainty quantification:

• Split Conformal Prediction: Distribution-free prediction intervals
• Adaptive Conformal Inference: Dynamic alpha adjustment for target coverage
• Multiple Implementations: Pure TypeScript with automatic WASM/Native acceleration

Coverage Guarantees

With conformal prediction, the forecaster provides mathematically guaranteed coverage:

P(y_true ∈ [lower, upper]) ≥ 1 - α

For α = 0.1, at least 90% of true values will fall within the prediction intervals.

Performance

Model Selection by Domain

Recommended models for each domain based on empirical performance:

• Solar: Prophet, ARIMA (strong daily seasonality)
• Wind: LSTM, Transformer (high variability)
• Demand: Prophet, LSTM (multi-seasonal patterns)
• Temperature: Prophet, Transformer (seasonal trends)

Computational Complexity

• Training: O(n·m·k) where n=samples, m=models, k=hyperparameter candidates
• Prediction: O(h·m) where h=horizon, m=models in ensemble
• Calibration: O(n log n) for conformal predictor
• Update: O(log n) for online learning

Parallel Execution

The swarm exploration runs hyperparameter search in parallel:

// All model types explored concurrently
const explorationPromises = models.map(modelType =>
  ensemble.exploreHyperparameters(modelType, data, validation)
);
await Promise.all(explorationPromises);

Testing

Comprehensive test suite with >90% coverage:

npm test                 # Run all tests
npm run test:watch      # Watch mode
npm run test:coverage   # Coverage report

API Reference

Main Classes

• EnergyForecaster: Main forecasting system
• EnsembleSwarm: Swarm-based ensemble learning
• EnergyConformalPredictor: Conformal prediction wrapper
• ARIMAModel, LSTMModel, TransformerModel, ProphetModel: Model implementations

Types

• TimeSeriesPoint: Time series data point
• ForecastResult: Single forecast with prediction interval
• MultiStepForecast: Multi-step forecast results
• ModelPerformance: Model performance metrics
• EnergyDomain: Energy domain enum
• ModelType: Model type enum

Advanced Usage

Custom Model Weights

const horizonWeights = new Map([
  [1, new Map([[ModelType.LSTM, 0.6], [ModelType.ARIMA, 0.4]])],
  [24, new Map([[ModelType.PROPHET, 0.5], [ModelType.LSTM, 0.5]])],
]);

const forecaster = new EnergyForecaster(domain, {
  ensembleConfig: {
    models: [ModelType.LSTM, ModelType.ARIMA, ModelType.PROPHET],
    horizonWeights
  }
});

Online Learning

// Initial training
await forecaster.train(historicalData);

// Generate forecast
const forecast = await forecaster.forecast(24);

// Update with new observations as they arrive
for (const observation of newData) {
  await forecaster.update(observation);
}

Accessing Statistics

const stats = forecaster.getStats();

console.log('Domain:', stats.domain);
console.log('Training points:', stats.trainingPoints);
console.log('Models:', stats.ensembleStats.modelCount);
console.log('Seasonal strength:', stats.seasonalPattern?.strength);

// Per-model performance
stats.ensembleStats.performances.forEach(({ model, performance }) => {
  console.log(`${model}: MAPE=${performance.mape}%, RMSE=${performance.rmse}`);
});

Integration with Neural Trader

This package integrates seamlessly with the Neural Trader ecosystem:

• @neural-trader/predictor: Conformal prediction core
• agentdb: Memory-persistent model performance tracking (optional)
• claude-flow: Swarm orchestration (optional)
• openrouter: Weather pattern interpretation (optional)

Contributing

Contributions welcome! Please ensure:

• Tests pass: npm test
• Linting passes: npm run lint
• Types check: npm run typecheck
• Coverage >90%: npm run test:coverage

License

MIT

See Also

• @neural-trader/predictor - Conformal prediction core
• Neural Trader - Main repository
• Examples - More example packages