GridPassport case study: forecast that learns the home

How to get good PV forecasts.

Publicly available PV forecasting tools are useful for a rough view of tomorrow's solar energy. If you want to manage a battery, charge an EV or run appliances when solar is available, rough is not enough for Home Power Automation. GridPassport improves the forecast by learning the real installation and correcting it every 15 minutes.

21 kWpreal home installation, not a lab example

1.57 -> 0.44 kWtypical miss fell after GridPassport used live PV data

15 minforecast corrected from observed production

Abstract: public PV models are good enough to orient yourself: tomorrow looks sunny, cloudy or mixed. We tested whether that is enough for Home Power Automation. It was not. On a connected 21 kWp installation, GridPassport used the real home, inverter feedback and nowcasting to cut the typical active-slot miss from 1.57 kW to 0.44 kW and move energy bias from -13.83% to -0.17%. The point is not a better chart; it is a better input for battery control, EV charging and energy arbitrage.

Why people check a PV forecast

Most people do not look at a solar forecast because they enjoy forecast models. They want to plan the next day. Should the EV charge at noon or later? Is it worth running the washing machine during the solar window? Should the battery keep reserve for the evening, or accept solar now?

Public tools help with that first question. They can tell you whether tomorrow is likely to be a solar day. Our free online PV forecast tool is useful for that rough plan. A home energy management system needs more. It has to make small decisions all day: charge, pause, export, import, hold reserve, shift load. Those decisions happen inside the day, not once in a daily total.

The practical question is: "What will this home produce, and when?" Weather is the starting point. Two homes under the same sky can behave differently because of shading, inverter limits, dirty panels, real tilt, real orientation, cable losses and the way the installation behaved in the last few hours.

What we compared

The data comes from one GridPassport V1 home in Central Poland, with a 21 kWp PV installation and telemetry from January 1 through May 30, 2026. The broad public DayAhead baseline starts on January 16 and runs through May 30. In that 124-day public-only window, daily PV energy error was about 22%.

The strongest comparison is May 15 through May 30, 2026. That is the window where public DayAhead, GridPassport same-day and GridPassport nowcast can be compared on the same active PV slots.

A 15-minute slot miss is a power miss, not "kW in 15 minutes". If the forecast says 8 kW and the home produces 6.4 kW, the miss is 1.6 kW. Over a 15-minute slot, that is 0.4 kWh of energy. This is why the number matters for batteries and EV charging.

Public forecast vs GridPassport

The public forecast was useful. It was not accurate enough to run the home.

The same 21 kWp installation shows the difference. Public DayAhead gave a rough plan. GridPassport same-day improved it by learning the home. GridPassport nowcast improved it further by checking real PV production every 15 minutes.

Forecast layer	What it knows	Typical 15-min miss	Daily energy error	Energy bias
Public DayAhead	Weather forecast plus declared or assumed PV setup.	1.57 kW / 0.39 kWh	baseline 17.2%	-13.83% -122.2 kWh total
GridPassport same-day	Updated weather plus the learned response of the real home.	1.20 kW / 0.30 kWh	25.3% lower 12.8%	-4.21% -37.3 kWh total
GridPassport nowcast	Updated weather plus production measured from the inverter every 15 minutes.	0.44 kW / 0.11 kWh	67.3% lower 5.6%	-0.17% -1.5 kWh total

The slot miss is measured only in active PV slots. For the public forecast, 1.57 kW is about 7.5% of the 21 kWp plant size and about 0.39 kWh in one 15-minute slot. It does not mean every slot produced the same amount of power.

Public baseline

21.9-22.1%

Daily PV energy error over the broader public DayAhead window.

Nowcast result

64.2%

Active slots within 1 kW, versus 36.9% for public DayAhead.

What GridPassport does differently

Public tools usually ask for the PV size, location, orientation, tilt and sometimes losses. Those values are helpful, but they are still declarations. GridPassport treats the real production history as the stronger signal.

If the real installation behaves as if the panels are shaded in the afternoon, the model should learn that. If the roof performs differently from the declared orientation, the model should learn that too. If the weather forecast says sun, but the clouds have not cleared yet, the model should react from the last observed PV production.

1. Weather changes

Refresh the forecast when the weather forecast changes.

Weather forecasts are valuable, but they are not static. GridPassport can rebuild the same-day plan when the weather input changes, instead of treating yesterday's day-ahead curve as the operating plan.

2. The roof is real

Learn shading, real tilt, orientation and losses from production.

Typed-in roof parameters are a starting point. Observed production is closer to the truth. It captures panel dirt, shading, clipping, inverter behavior and installation-specific response.

3. The day is live

Correct the forecast every 15 minutes.

If expected production and real production diverge, GridPassport recalculates the short-horizon PV forecast and the home can change how it manages the battery, EV charging and flexible loads.

Measured days

On individual days, the difference was not subtle.

Public forecasts can be directionally right and still miss the home enough to change a battery or EV decision. These three days show how the connected nowcast corrected the plan.

2026-05-17

Lower-production, cloudy day where in-home correction mattered.

Actual PV: 33.716 kWh

Public17.737 kWh / 52.61%

Same-day0.803 kWh / 2.38%

Nowcast0.796 kWh / 2.36%

Nowcast reduced daily absolute error by 95.5% versus public day-ahead.

2026-05-18

Large public under-forecast corrected by GridPassport.

Actual PV: 41.899 kWh

Public18.289 kWh / 43.65%

Same-day1.388 kWh / 3.31%

Nowcast1.199 kWh / 2.86%

Nowcast reduced daily absolute error by 93.4% versus public day-ahead.

2026-05-28

Candidate site-specific afternoon mismatch.

Actual PV: 46.088 kWh

Public14.847 kWh / 32.21%

Same-day11.073 kWh / 24.03%

Nowcast0.167 kWh / 0.36%

Nowcast reduced daily absolute error by 98.9% versus public day-ahead.

Three example days

Different days, same conclusion: public PV forecasts are not enough for control.

One day was cloudy. One had a strong public under-forecast. One had a different shape problem in the afternoon. In each case, the useful operating signal came from connecting the forecast to the real installation.

2026-05-17: Cloudy, lower-production day.Public day-ahead saw too little PV for most of the active window. GridPassport stayed close after it could use the real home signal.

Actual PV Public day-ahead GridPassport same-day GridPassport nowcast

2026-05-18: Strong public under-forecast.The public curve was directionally useful but too low around the most valuable hours. Same-day and nowcast recovered much of the operating signal.

Actual PV Public day-ahead GridPassport same-day GridPassport nowcast

2026-05-28: Different problem: wrong shape in the afternoon.Public day-ahead knew it was a solar day, but missed the home-specific shape. Nowcast pulled the operating plan back toward the actual production curve.

Actual PV Public day-ahead GridPassport same-day GridPassport nowcast

PV is only half of the plan. Load is the other half.

A PV forecast tells the home how much energy may arrive from the roof. A useful home energy system also forecasts household load. The operating signal is:

expected surplus = forecast PV - forecast home load

That surplus signal answers the questions that move devices. When will solar cover the home load? When should the battery charge or hold reserve? When should an EV charger slow down? When is it better to import before a cloudy period?

In the full January 1 through May 30 telemetry period, the home had a PV-greater-than-load window on 127 days. The average window was 6.58 hours. In the strict May 15 through May 30 comparison window, every comparable day had such a window, averaging 8.19 hours.

The takeaway

If you do not have a connected energy system, use public PV tools. They are useful half-measures. They help you plan loosely and understand whether tomorrow should be sunny or cloudy.

For Home Power Automation, they are not enough. A HEMS needs a forecast that keeps learning the real installation, keeps checking the last 15 minutes of production and keeps updating the plan for the battery, EV charger and flexible loads.

That is why GridPassport forecast is not something the system does once a day. It is part of the operating loop. Weather starts the plan. The home corrects it.

Home Power Automation

Your home already has the pieces. It needs a shared power plan.

GridPassport is being built to coordinate supported batteries, EV charging, heating and cooling using solar state, forecasts, tariff signals and household requirements.

Join founding members Read how it works