Differentiable physics (demand, calendar, and weather)

This theme covers the demand model of a six-part series on forecasting with physical terms.

Background: Differentiable physics: background explains the models in plain terms.

The series builds models of Primary substation demand from a daily profile, holiday and heating terms, and a solar term, fitted by gradient-based optimisation. This theme holds the part that builds the model and asks whether more years of history help it. The series' finding so far is in experiment 104: the convex model trained on 12 months beat lagless XGBoost with its default start by 0.25 points (post hoc), and all years with drift are level with 12 months, so more years are not yet shown to help. Experiment 63 is an earlier, smaller linear test of the same decomposition into a weather baseline, a solar term, and transfer terms. The Experiments page lists every experiment, and the Research plan lists what is planned next.