How we know what we know.
Meridian's pipeline is documented end-to-end. There are no black boxes between an observation in the world and a number on a Meridian dashboard.
- 01
Ingest
Telemetry, market feeds, satellite imagery, and operator filings stream into a partitioned lakehouse with full provenance metadata.
- 02
Validate
Each signal is cross-checked against ≥2 independent sources. Conflicts are surfaced, never silently averaged.
- 03
Reconcile
Bayesian state estimators reconcile partial telemetry with physical constraints and historical priors.
- 04
Attribute
Every published metric carries an attribution graph: source, transform, model version, and confidence interval.
- 05
Publish
Outputs are versioned and immutable. Corrections issue as new versions; nothing is rewritten in place.
- 06
Audit
External auditors are granted read-only access to the full pipeline. Audit reports are public.
Every Meridian figure is published with a 95% confidence interval and a method version string. If you can read the number, you can audit how we got there.
energy pulse
Real-time composite of grid stress and resilience. Computed per region or globally.
Inputs
| Input | Unit | Range | Direction | Weight | Definition |
|---|---|---|---|---|---|
| Demand | load / capacity | 0 – 1.1 | ↓ better | 20.0% | Instantaneous load divided by available capacity. |
| Reliability | score 0..1 | 0 – 1 | ↑ better | 20.0% | 1 − (frequency deviation + outage minutes) normalized. |
| Reserve Margin | fraction | 0 – 0.4 | ↑ better | 15.0% | (Capacity − Peak Demand) / Peak Demand. |
| Carbon Intensity | gCO₂/kWh | 20 – 900 | ↓ better | 20.0% | Generation-weighted grid carbon intensity. |
| Congestion | utilization | 0 – 1 | ↓ better | 10.0% | Mean transmission line utilization. |
| Pricing Volatility | σ/μ (24h) | 0 – 1 | ↓ better | 15.0% | Coefficient of variation of wholesale price over 24 hours. |
Confidence
Gaussian propagation of per-input (1 − confidence) variance; 95% CI = ±1.96·σ.
global energy index
Country-level composite of structural energy-system performance across six pillars.
Inputs
| Input | Unit | Range | Direction | Weight | Definition |
|---|---|---|---|---|---|
| Reliability | score 0..1 | 0 – 1 | ↑ better | 16.7% | SAIDI/SAIFI proxy + frequency stability. |
| Affordability | price / GDPpc | 0 – 0.15 | ↓ better | 16.7% | Retail price per kWh as fraction of GDP per-capita-hour. |
| Sustainability | renewable share | 0 – 1 | ↑ better | 16.7% | Low-carbon generation share weighted against carbon intensity. |
| Capacity | MW/capita | 0 – 0.004 | ↑ better | 16.7% | Installed generating capacity per capita. |
| Infrastructure Health | score 0..1 | 0 – 1 | ↑ better | 16.7% | Asset age, investment trend, transmission coverage. |
| Energy Security | score 0..1 | 0 – 1 | ↑ better | 16.7% | Import dependency + fuel-mix diversity (Shannon entropy). |
Confidence
Gaussian propagation of per-pillar (1 − confidence) variance; 95% CI = ±1.96·σ.
Programmatic access
Every published methodology and every historical calculation is queryable without authentication.
/api/public/methodologyList every registered methodology version.
/api/public/methodology?index=energy_pulseFilter to a single index.
/api/public/energy-pulseLatest global Energy Pulse with 95% CI.
/api/public/energy-pulse?history=1Latest + last 100 historical snapshots.
/api/public/global-energy-index?country=DELatest GEI for a country (ISO code).
/api/public/global-energy-index?country=DE&history=1Historical GEI series.
/api/public/forecast?metric=demand_mw&scope=USLatest demand forecasts (Holt-Winters, Prophet, XGBoost, LightGBM) with 80/95% intervals + scorecards.
/api/public/forecast?metric=demand_mw&scope=USRetrain & backtest all four model families (requires INGEST_API_KEY).
Four-model ensemble, scored
The MVP rule-based stress engine has been replaced by a four-family ML stack. Every retrain runs all four in parallel, ranks them by backtest sMAPE, and writes per-fold predictions, residuals and confidence intervals to the historical archive.
Triple exponential smoothing with daily seasonality. Best for stable, strongly seasonal demand.
Linear trend + Fourier seasonality (daily/weekly/yearly). Robust to gaps and changepoints.
Gradient boosted trees on 24 lag features + cyclical encodings. Captures non-linear effects.
Leaf-wise tree growth, deeper interactions. Often wins on long horizons.
Pipeline
- Train — fit each model on the full observation series; record sMAPE, RMSE, MAE, R² in
forecast_runs. - Predict — emit point forecasts with 80% and 95% gaussian prediction intervals derived from training residual σ.
- Backtest — rolling-origin cross-validation with k folds and fixed horizon; per-fold scorecards stored in
forecast_backtests. - Grade — aggregate sMAPE → A (<5%), B (<10%), C (<20%), D (<35%), F otherwise. Best family per metric/scope flagged automatically.
- Backfill — once truth lands,
actual,abs_error,apeare written on each historical prediction for live MAPE tracking.