Skip to content

Techniques

This section holds durable explainers of the solution methods we bring to the forecasting problem — what each technique is, how it works, and why it fits. It complements the other sections: background describes the problem, the roadmap says what we plan to build and in what order, and architecture documents what is already built. Unlike roadmap pages, techniques pages are permanent: they stay put (and stay linkable — including from code docstrings) as the work that applies them ships.

  • Convex optimisation — the one-bowl property, why a certified global optimum beats gradient descent wherever it is available, exact zeros from ℓ1 penalties, hard limits as censoring, quantile envelopes, priors as penalties, and the project tooling rule (CVXPY for convex estimation subproblems; PyTorch for physics + learning). First applications: the edge-flow estimator on the switching-events roadmap page, and the convex candidate in the v0.7 capacity head-to-head.
  • Probabilistic forecasting from NWP ensembles — why pushing an NWP ensemble through a deterministic power model is systematically overconfident (under-dispersed), the three-term uncertainty decomposition, the fix (per-member conditional quantile forecasts recombined by the linear-pool mixture), the pooling recipe, why quantile-averaging (Vincentization) re-creates the under-dispersion, and the double-counting caveat. Applied by the probabilistic evaluation & calibration plan and the delivery tables.
  • Evaluation metrics — every metric compute_metrics produces, each with a plain-language summary, the equation, and the design decisions behind it: the deterministic four (MAE, NMAE, RMSE, MBE), fair CRPS, the Fortin-corrected spread-skill ratio, pinball loss at the NGED delivery quantiles, PICP with its finite-ensemble calibrated references, and interval width. Applied by the metrics & leaderboard plan and computed by the metrics Dagster asset.
  • Differentiable physics — inversion through differentiable forward models: the single-site solar plant and the aggregate fleet node. Applied by the capacity-estimation and net-demand disaggregation roadmap pages.
  • Learned encoders — shared WeatherEncoder / TimeEncoder modules and why they pair naturally with differentiable physics.
  • Disaggregation evaluation — the multi-pronged protocol for evaluating disaggregation when no single clean ground truth exists.