f97.be · Three views of tomorrow

Solar forecasts

Compare the physical GTI forecast, the AI model and their adaptive Fusion blend.

Loading the latest data…

Forecast Summary

Today · measured so far and full-day forecasts

7-day forecast · AI, GTI and Fusion

Earlier model development

Historical notes from the Python implementation. The current native model is described above and in the methodology link.

Version Date Key Improvements
v6 July 2026
  • Alignment with Deployed Implementation: The paper now documents the production system exactly. The estimation layer is a box-constrained grid search with time-series cross-validation and a robust median-residual intercept (replacing the earlier Huber/ridge regression).
  • Anchored Online Calibration: The scale factor is now gated, anchored to a long-term baseline, hard-clamped, and exponentially smoothed, limiting day-to-day movement.
  • Forecast Floor & UQ Update: A forecast floor (15% of mean AI/GTI) was introduced, uncertainty now blends short/long horizons equally with a 40% interval cap, and statistically inert winsorization steps were removed.
v5 April 2026
  • Documentation Update: Paper revisions to enhance clarity, define mathematical formulas explicitly, and add real-world evaluation metrics. No changes to the underlying model logic.
v4 April 2026
  • UQ Logic Standardization: Addressed a drift between the formalized mathematical model and the deployed Python implementation, aligning Uncertainty Quantification (UQ) logic to match production realities.
  • Confidence Interval Stabilization: Adjustments ensure confidence intervals do not artificially explode on low-yield days (via relative mapping and hard capping) and remain smoothly reactive rather than locking onto seasonal maximums (via the weighted blend).
v3 March 2026
  • Dual-Window Uncertainty: Computes a robust maximum between short-term reactive MAD and long-term seasonal MAD to prevent overly reactive prediction bands.
  • Zero-Value Regimes: Safely retains previous stable scale factors during snow cover and persistent zero-generation to prevent downward drift.
  • Model Selection Logic: Clarified separation between the internal Huber fitting objective and the MAE-based out-of-sample selection metric.
  • Fallback Scaling: Ensured cold-start fallback intercept strictly mirrors the scaling algebra of the primary model.
v2 March 2026
  • Bias Components: Formalized distinction between additive bias (intercept) and multiplicative drift (online scale).
  • Safeguards: Physical output bounds defined explicitly as conservative engineering constraints.
  • Math & Notation: Rewritten calibration step using strict set notation; uncertainty formulation corrected.
home | top | rss
About AI, GTI and Fusion

The native fff service on the Mac publishes this forecast independently of the retired Python server. It reads Fronius production and Open-Meteo weather forecasts, trains a boosted-tree AI model and combines AI with its calibrated physical GTI forecast.

The seven-day forecast is issued daily. Production readings and the published feed update every 15 minutes. Today's measured production is still incomplete; AI, GTI and Fusion describe the full day. Missing forecasts are shown as unavailable.

The GTI values here use the native fff calibration. The separate 14-day GTI forecast preserves the original calculation now supplied by eDaemon; the two model versions can produce different values.

Current model details: native forecast methodology. Earlier paper: Fusion Forecasting for Residential PV. Original AI, GTI and Fusion histories remain available below.

Original forecast archive

Original AI and Fusion records are preserved unchanged. These files describe the retired Python models, separately from the current native fff forecast.

Original GTI history

Preserved daily GTI forecasts and measured production. This archive uses the original GTI model.

—

Loading the preserved archive…

PeriodActual · kWhSaved GTI · kWhActual / forecastRecorded days