Differentiable physics: embedded generation: research plan

Id	Question	Priority	Waits for	State
Wind node	A typed wind node beside the flat generator node	Proposed	A flat node that stays off where no generator exists (experiment 110 found it reads 12 to 35 points at all five decision substations)	Not started
Dispatchable node	A typed dispatchable-generator node beside the flat generator node	Proposed	A flat node that stays off where no generator exists (experiment 110 found it reads 12 to 35 points at all five decision substations)	Not started
Solar node	A typed solar-farm node beside the flat generator node	Proposed	A flat node that stays off where no generator exists (experiment 110 found it reads 12 to 35 points at all five decision substations)	Not started
265, gas engine bands	Can an outside prior and a causal visibility switch make D8's widened bands honest in every quarter, and do they harm Primaries with no engine?	Running	None	Planned
352, capacity response surface	How much does the sunny-half-hour score of Primary demand research default 3 move with the solar capacity, with and without the Kalman filter?	Running	None	Planned
N13, PV orientation mix	Does replacing the single south-facing solar shape with three non-negative capacities (south, east, west) keep the model convex and lower the summer morning and evening error?	Dropped	None. Experiment 73 and experiment 77 found the mix unidentified or immaterial, and experiment 114 dropped it	Dropped

N13: PV orientation mix

Experiment 114 dropped N13, so the design below is not queued. It stays as the design for any later experiment on solar orientation.

The proposal gives the convex model three non-negative solar capacities, one each for south, east, and west, in place of one south-facing unit shape. The model stays convex, and an east-west array should get a larger east and west share. The shares count as identified only if they pass the coverage rule of experiment 113, which asks that 80% intervals cover the injected truth 70 to 90% of the time. The experiment passes if the shares are identified and the score does not get worse. The run takes minutes on the CPU. It was first considered for the same bundle as the interaction-term and heating-threshold experiment, which ran without it as experiment 114.

The later parts of the differentiable-physics series are planned one at a time, from the result of the part before.

Planned experiments

Experiment	Hypothesis	Relevance