Does a clear-sky-index forecast rescue the optimal home-battery controller on real weather?

What it found

Replaces the synthetic weather of energy-tou-lp-forecast with hourly irradiance from PVGIS typical-meteorological-year files (ERA5-based) for Golden CO, Phoenix AZ and Seattle WA, and adds the standard cheap PV forecast: yesterday's clear-sky index applied to the known clear-sky curve (“smart persistence”). The same rolling 36 h linear program, tariff and execution rule are used. The clear-sky forecast cuts hourly day-ahead nRMSE by 4–13% against yesterday's weather (0.63→0.55 at Golden, 0.68→0.60 at Seattle) but changes the annual saving by at most $5 in dispatch: for 10 kWh behind 6 kW PV the perfect-forecast optimiser saves $929, $880 and $852 a year at Golden, Phoenix and Seattle against the best fixed rule's $734, $668 and $618, while the realistic forecasts save $722–733, $779–785 and $677–691. They keep −6% to −1% of the gain at Golden (the fixed rule wins), 52–55% at sunny Phoenix and 25–31% at cloudy Seattle, and on life-limited break-even price they sit below the rule at Golden ($672–677/kWh against $713). So the forecast, not the optimiser, sets the value, and a better PV forecast alone does not buy it: a perfect load forecast adds only $15–18 a year. Real daily clear-sky index has lag-1 autocorrelation 0.29–0.36 against 0.50 for the synthetic chain. Not from a source: the tariff, load shape, battery cost and the 4,000-cycle life (inherited, assumed); the typical year stitches months from different years, and its irradiance is a reanalysis estimate, not a pyranometer record.

A figure from its own output

Paper, code and results

• README and method
• Experiments and results
• Code
• Tests

More in Energy supply and dense energy storage

• How wrong is the DC approximation of optimal power flow on the IEEE 30-, 39- and 118-bus cases?DC against AC optimal power flow in pandapower on the IEEE 30-, 39- and 118-bus cases: the DC model's cost is 1 to 3 % below the AC optimum (it ignores losses); on the 118-bus case its dispatch is feasible but costs 0.7 to 1.7 % more than the AC optimum once losses are paid, and on the 39-bus case it breaks voltage limits (15 buses at 0.6 of load) at almost no cost difference (modelled benchmark, not a measured system; tightened-limit runs often had no solution). Feeding the AC losses back into the DC model (a loss-adjusted DC) brings the 118-bus cost estimate to within 0.6 to 1.2 % of the AC optimum, but repairs only 13 to 32 % of the dispatch's excess cost and none of case39's voltage violations.
• How much grinding energy does a classifier loop save at the same product size, and does a ball-mill population balance agree with Bond's law?Ball-mill population balance (Austin form): at the same 100 um P80 a closed circuit at 278% circulating load uses 0.47 of the open-circuit energy, 1.8 to 3.7 times better across the swept mill and classifier inputs, far above Bond's 1.2 open-circuit factor and a vendor's ~10% for cement; the model's size exponent matches Bond's only at a selection exponent near 0.5.
• How much rooftop PV can the IEEE 13-node feeder take before voltage binds, and which cheap measure adds the most?Stochastic PV hosting capacity of the IEEE 13-node feeder in OpenDSS on 300 random placements: median 1,015 kW at midday minimum load, voltage binds first in 75% of placements; a 2 V lower regulator set-point adds +310 kW, while inverter fixed power factor and volt-var add about nothing once the regulators respond (modelled, not a measured feeder).
• Does regulator set-point still beat inverter volt-var for PV hosting on the IEEE 34- and 123-node feeders?PV hosting capacity of the IEEE 34- and 123-node feeders in OpenDSS (200 random placements): on the 123-node feeder a 2 V lower regulator set-point adds +790 kW to a 1,160 kW median (every placement better), inverter volt-var +100 kW and fixed power factor +180 kW; the published 34-node case has no headroom at all because its source is set to 1.05 pu (modelled, not measured feeders); over a one-minute day the set-point trades overvoltage for undervoltage while inverter reactive power cuts tap operations about 30 %.
• How much does model-predictive heating and cooling save over a thermostat on a time-of-use tariff, when the model and forecast are wrong?Does MPC beat a thermostat for a heat pump on a time-of-use tariff? On real TMY3 weather with a deliberately wrong model and yesterday's weather as the forecast it pays 10.7% less in Greensboro (4.1% in cold Sand Point), saves more with more thermal mass (6.5% to 25%), and loses the gain at an 8 K forecast error; the envelope is from published values, not a fit to measured data.
• Does a DC screen find the N-1 outages an AC power flow flags?DC against AC N-1 screening on IEEE 14-, 30-, 39-, 57- and 118-bus cases (every line and transformer outage, modelled, ratings swept): at 25 % rating headroom the DC screen misses 13 to 40 % of the outages an AC power flow calls insecure; flagging at 0.8 of rating misses none, at the cost of 102 false alarms on case118 (186 outages).

All 39 Energy supply and dense energy storage projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.