GHI vs DNI vs DHI: Formula and Which One to Use

Last updated: November 2026

Three irradiance numbers show up in every solar dataset, and they are not interchangeable. Pick the wrong one and your yield estimate is off before the simulation starts. Here is what each one is, in one sentence each:

• GHI (global horizontal irradiance) is the total sunlight on a flat, horizontal surface.
• DNI (direct normal irradiance) is the direct beam, measured on a surface held at right angles to the sun’s rays.
• DHI (diffuse horizontal irradiance) is the sunlight scattered by the sky and clouds, measured on a horizontal surface.

A common shortcut is to pull a single GHI number from a free database. People then skip the decomposition step, which splits GHI into direct and diffuse light. They also ignore the difference between horizontal irradiance and what actually hits a tilted array. Each of these shortcuts adds error. The error carries through the whole energy yield calculation.

This guide covers the physics of GHI, DNI and DHI. It also covers the closure equation with a worked example, transposition models, the main data sources, and reference data for 20 European cities. It ends with how irradiance uncertainty feeds into P50 and P90 yield estimates. Transposition models convert horizontal readings into the light that hits a tilted array. P50 is the yield expected to be met or exceeded in 50% of years. P90 is the yield expected to be met or exceeded in 90% of years. The city data and the Munich example come from the European Commission’s PVGIS (details under each table). For related reading, see European Solar Incentives and solar insolation: a designer’s reference.

What you will learn:

• The four types of solar irradiance (GHI, DNI, DHI, POA) and how they relate mathematically
• How to calculate irradiance on a tilted surface using decomposition and transposition models
• A worked example for a 30-degree array in Munich, with numbers from PVGIS
• How PVGIS, NSRDB, Solargis, NASA POWER and Meteonorm compare
• Reference irradiance data for 20 European cities with expected yields
• How P50 and P90 are calculated from yield uncertainty
• How solar design software and solar proposal software use irradiance data from input to energy yield

What Is Solar Irradiance and Why It Matters for System Design

Solar irradiance is the power of sunlight per unit area, measured in watts per square meter (W/m²). Solar irradiation, also called insolation, is irradiance added up over time. You measure it in kWh/m² over a day, month, or year. The Sandia PV Performance Modeling Collaborative puts it as “irradiance is to power as insolation is to energy”. Design software needs both. It uses irradiance for power at one moment and irradiation for energy yield estimates.

One peak sun hour is 1 kWh/m² of insolation, the energy you would get from 1,000 W/m² for one hour. A location receiving 4.0 kWh/m²/day is therefore said to have 4.0 peak sun hours a day. Module ratings are quoted at a fixed reference irradiance. That is why Standard Test Conditions matter when you compare datasheets.

Irradiance vs. Insolation

When someone says “Munich gets 1,211 kWh/m² a year”, they are stating annual GHI insolation (the PVGIS 5.3 figure for central Munich, used in the tables below). When a datasheet says “1,000 W/m²”, that is irradiance at a single moment. Always state the time period and the plane the number refers to. A GHI value is for a horizontal plane. A DNI value is for a plane that faces the sun. Neither is what a tilted roof receives.

How Irradiance Determines kWh/kWp

Specific yield (kWh/kWp) is the energy a system makes for each kWp of panel rating. A kWp is one kilowatt of peak panel power. Yield rises with the irradiance reaching the panels, but not in a straight line. Temperature, light spectrum, angle of incidence, inverter losses and cable losses all change the result. In the PVGIS data below, Madrid’s annual GHI is 80% higher than Stockholm’s (1,779 against 990 kWh/m²), while a 1 kWp system at its optimal tilt produces 63% more energy in Madrid (1,620 against 991 kWh a year).

The 10% Rule

A 10% irradiance error produces roughly a 10% yield error. Take a 100 kWp system that would produce 1,000 kWh/kWp. The error is 10,000 kWh a year. It repeats every year of the system’s life. So it matters most where the yield estimate sets the financing or the guarantee.

You need to understand solar angles before you work with irradiance data. The angle at which sunlight strikes a surface decides how much energy that surface receives. The zenith angle, tilt angle, and azimuth all feed directly into the irradiance equations below.

GHI vs DNI vs DHI: The Four Types of Solar Irradiance

Solar radiation reaching the ground splits into separate parts. Each one behaves differently. Mixing them up leads to design errors.

GHI: Global Horizontal Irradiance

GHI is the total solar radiation received on a horizontal surface. It is the number solar databases report most often. It is the starting point for most PV yield calculations. GHI combines the direct beam from the sun and the diffuse light scattered by the atmosphere.

Across the 20 European cities in the table below, annual GHI runs from 990 kWh/m² in Stockholm to 1,800 kWh/m² in Athens. Seville, which is not in that table, comes to 1,904 kWh/m² in the same PVGIS run. Also see: Spain net metering.

DNI: Direct Normal Irradiance

Direct normal irradiance (DNI) measures only the direct beam of sunlight, on a surface perpendicular to the sun’s rays. It is the main measure for tracking systems and concentrating solar power (CSP). Trackers turn the panels to follow the sun. CSP uses mirrors to focus sunlight. For fixed-tilt PV, DNI still matters because it sets the direct light that reaches your tilted array.

DNI varies more across locations than GHI. In the table below it runs from 988 kWh/m² in London to 2,131 kWh/m² in Madrid, which is more than double. Clouds block the direct beam far more than they block diffuse light.

DHI: Diffuse Horizontal Irradiance

Diffuse horizontal irradiance (DHI) is the radiation received on a horizontal surface from the whole sky, not counting the direct beam. It includes light scattered by clouds, aerosols (tiny particles in the air), and air molecules. On an overcast day, DHI is nearly all you get.

DHI is the steadiest of the three. In the 20-city table it ranges only from 472 kWh/m² (Stockholm) to 617 kWh/m² (Athens). What changes is its share of GHI. It is 31% in Madrid, 35% in Rome, 47% in Munich and 54% in London.

POA: Plane of Array Irradiance

POA is the total irradiance reaching your array’s surface, with tilt and direction taken into account. It is GHI translated onto your array geometry. You usually calculate POA instead of reading it from a standard database. A pyranometer, a sensor that measures sunlight, can measure it on site if it lies flat against the array.

Getting from GHI to POA is the main technical step, covered in the next section. You split GHI into direct and diffuse parts. Then you project each part onto the tilted surface.

The Closure Equation: GHI = DNI x cos(zenith) + DHI

The relationship that ties the three together is:

GHI = DNI x cos(theta_z) + DHI

Here theta_z is the solar zenith angle, the angle between the sun and the vertical. The direct beam is measured on a plane facing the sun. So you multiply it by cos(theta_z) to project it onto the horizontal plane. The diffuse part is already measured on the horizontal plane. So you add it as it is.

Worked example (illustrative inputs). Take a clear moment with the sun 30 degrees from vertical, DNI = 900 W/m² and DHI = 120 W/m²:

GHI = 900 x cos(30°) + 120 = 900 x 0.8660 + 120 = 779.4 + 120 = 899.4 W/m², or about 899 W/m².

You can run the same relation backwards. Given GHI = 899.4 W/m² and DHI = 120 W/m², DNI = (899.4 - 120) / cos(30°) = 779.4 / 0.8660 = 899.9. That is about 900 W/m² after rounding.

Component Comparison Table

The ranges come from the table in the city reference section below.

How to Calculate Solar Irradiance for Your Site

Getting from a GHI number to the irradiance on your tilted array takes three steps. They are decomposition, transposition, and correction. Each step adds assumptions and possible error.

Step 1: GHI Decomposition

Most databases give GHI as the main value. Some also give DNI and DHI. When they do not, you use a decomposition model to split GHI into direct and diffuse parts. The open-source library pvlib includes the common ones. Its documentation describes them as follows.

• Erbs: estimates the diffuse fraction from the clearness index Kt (GHI divided by the sunlight that would reach a horizontal surface at the top of the atmosphere) using a relationship fitted to measured data. Then it works out DHI = diffuse fraction x GHI and DNI = (GHI - DHI) / cos(zenith). The inputs are GHI, zenith angle and day of year.
• Boland (also written BRL, for Boland-Ridley-Lauret): estimates the diffuse fraction with a logistic function (an S-shaped curve) of the clearness index.
• DISC: converts GHI to DNI using relationships fitted to measured data, between the global and direct clearness indices.
• DIRINT: a change to DISC that uses GHI over a time series and, if you have it, dew point temperature.

A short pvlib example with Erbs, using illustrative inputs of GHI = 600 W/m², zenith = 40° and day of year 172:

from pvlib import irradiance

out = irradiance.erbs(ghi=600, zenith=40, datetime_or_doy=172)
print(out["dni"], out["dhi"], out["kt"])

With pvlib 0.16.1 this returns DNI = 426.3 W/m², DHI = 273.5 W/m² and Kt = 0.593. Check it against the closure equation. 426.27 x cos(40°) + 273.46 = 426.27 x 0.7660 + 273.46 = 326.54 + 273.46 = 600.0 W/m², which is the GHI you started with.

If your data source already provides DNI and DHI with GHI, use those values. They come from the provider’s own method, tuned to its satellite data. They will usually beat a generic decomposition model applied afterwards.

Step 2: Transposition Models

Once GHI is split into DNI and DHI, you project each part onto your tilted array. The total POA irradiance has three parts:

POA = POA_beam + POA_diffuse + POA_reflected

The beam part is a geometric projection of DNI onto the tilted surface, using the angle of incidence. The ground-reflected part uses an albedo factor. Albedo is the share of light the ground reflects. The hard part is the diffuse light.

Isotropic Model

The simplest model assumes diffuse light arrives equally from every part of the sky:

POA_diffuse = DHI x (1 + cos(tilt)) / 2

It is fast. It treats the sky as an even source of light (the pvlib documentation describes it that way). So it leaves out the brighter sky near the sun and near the horizon, which more detailed models include.

Hay-Davies Model

The Hay-Davies model uses DNI, DHI, the irradiance above the atmosphere, and the sun and surface angles. It works out the sky’s diffuse light on a tilted surface. It splits diffuse light into two parts. One is a circumsolar part, the glow around the sun, treated like a beam from the sun’s direction. The other is an even background, called isotropic.

Perez Model

The Perez 1990 model divides the sky into zones. They are a circumsolar disc, a horizon band and an even background. Brightness coefficients fitted to measured data depend on how clear and bright the sky is. It is available in pvlib as irradiance.perez. Unless you have a reason to choose another model, Perez is a sensible default. For software options, see 7 Best PVsyst Alternatives.

Step 3: Worked Example, 30-Degree Array in Munich

Take a south-facing array tilted 30 degrees in central Munich (48.14°N, 11.58°E, 524 m). All inputs are annual averages from PVGIS 5.3 (SARAH-3 radiation data, 2005 to 2023). They were pulled through the PVGIS API on 28 October 2026. The reflected part below matches a ground albedo of 0.2.

Horizontal plane:

• Beam on the horizontal plane: 638.2 kWh/m²
• Diffuse (DHI): 573.1 kWh/m²
• GHI: 638.2 + 573.1 = 1,211.3 kWh/m²
• Diffuse share of GHI: 573.1 / 1,211.3 = 47.3%

Tilted plane (30°, due south):

• Beam: 817.5 kWh/m²
• Diffuse: 588.8 kWh/m²
• Ground-reflected: 16.2 kWh/m²
• POA: 817.5 + 588.8 + 16.2 = 1,422.5 kWh/m²

What the numbers show:

• Tilting the array raises the beam part by 28% (817.5 / 638.2 = 1.281). It raises the diffuse part by 3% (588.8 / 573.1 = 1.027).
• POA is 1,422.5 / 1,211.3 = 1.174 times GHI, so 17.4% more irradiation than a horizontal surface.
• The reflected part matches the albedo formula: POA_reflected = GHI x albedo x (1 - cos(tilt)) / 2 = 1,211.3 x 0.20 x (1 - 0.8660) / 2 = 16.2 kWh/m².

Yield. For a 1 kWp system at the same tilt with 14% system losses, PVGIS 5.3 returns 1,130 kWh a year. Divide 1,130 kWh/kWp by the 1,422.5 kWh/m² of POA irradiation. The result is 0.79. This is the implied performance ratio once PVGIS has applied its angle-of-incidence, spectral, temperature and system losses. The performance ratio compares the energy a system delivers with the energy its rated power and the sunlight received would give in an ideal case. See the performance ratio guide for how that figure is defined.

Roof pitch often decides the tilt. Use the roof pitch calculator to convert a pitch to degrees. Then compare the result with the optimal tilt in the city table below (39° for Munich).

Temperature Corrections

Irradiance data alone does not give the full picture. Crystalline silicon modules lose output as cell temperature rises above the 25°C of Standard Test Conditions. The module datasheet gives the rate as the temperature coefficient of power. Professional solar software applies that coefficient hour by hour from air temperature, irradiance and wind speed. In the Munich run above, PVGIS reports a temperature and low-irradiance loss of 6.2%. This sits inside a total loss of 20.6%. The total also combines the 14% system loss, a 3.03% angle-of-incidence loss and a 1.55% spectral gain.

Solar Irradiance Data Sources Compared

The quality of your irradiance data sets the quality of your yield prediction. The table lists the main sources by what each operator publishes. Prices and licence terms change. Check them on the vendor’s site.

Sources: the JRC pages on using PVGIS 5 and data sources and calculation methods, the Solargis Prospect page, the NASA POWER methodology page and the Meteonorm interpolation documentation. We could not read cost, resolution and validation figures for NSRDB, Solargis and Meteonorm from their own pages for this update. So none are quoted.

The Global Solar Atlas is a quick visual overview of solar potential when you need a first look at a site.

Irradiance Across Europe: Reference Data by City

The table gives annual irradiation and expected yield for 20 European cities. Every value is a 2005 to 2023 average from PVGIS 5.3 (SARAH-3 radiation data). It was pulled through the PVGIS API on 28 October 2026 for the coordinates of each city centre. GHI and DNI come from the monthly radiation service. DHI comes from the hourly series on a horizontal plane. Optimal tilt is the fixed, south-facing angle that PVGIS returns when asked to maximise annual yield. Yield is the annual energy from 1 kWp at that tilt, with PVGIS’s 14% default system loss. The table is sorted by GHI.

European Solar Irradiance Reference Table

Also see: solar panel ROI in Italy, Germany solar subsidies and France solar feed-in tariffs.

What the Table Shows

GHI spread. Among these 20 cities, Athens has the most GHI (1,800 kWh/m²) and Stockholm the least (990 kWh/m²). Athens receives about 82% more.

DNI spread is wider. Madrid’s DNI (2,131 kWh/m²) is more than double London’s (988 kWh/m²). Tracking and tilt optimization gain more in absolute terms where the direct beam is strong. For the latest details on Italy, see Commercial Rooftop Solar Case Study Italy.

DHI changes little, its share changes a lot. Diffuse irradiation stays between 472 and 617 kWh/m² across the table. As a share of GHI it goes from 31% in Madrid to 54% in London. It is 47% in Munich and 48% in Stockholm.

Optimal tilt is below latitude everywhere in this table. The gap runs from about 3 degrees in Madrid and Barcelona to 15 degrees in Stockholm. It widens with latitude. A “tilt equals latitude” rule of thumb overstates the best angle at every one of these sites.

The yield gap is smaller than the GHI gap. Madrid has 80% more GHI than Stockholm but produces 63% more per kWp (1,620 against 991 kWh/kWp).

The forecasting guide covers the day-to-day side of irradiance prediction. It shows how solar energy forecasting connects to these irradiance patterns, especially for grid integration and trading.

How Irradiance Data Quality Affects P50 and P90 Yield

Energy yield estimates come with a chance of being beaten. The irradiance dataset is one input to the uncertainty behind that number.

P50 vs. P90: What They Mean

P50 is the median expected annual energy yield. The system is as likely to produce more than this in a given year as less. P90 is the yield expected to be met or exceeded in 90% of years. The Solargis guide to Pxx estimates assumes annual yield follows a normal distribution, the usual bell curve. It estimates P90 by subtracting 1.282 standard deviations from P50:

P90 = P50 x (1 - 1.282 x combined_uncertainty)

Here combined_uncertainty is the standard deviation, written as a fraction of P50. Here are two recomputed examples:

• At 6% combined uncertainty: P90 = P50 x (1 - 1.282 x 0.06) = P50 x (1 - 0.0769) = P50 x 0.923, so P90 is 7.7% below P50.
• At 10% combined uncertainty: P90 = P50 x (1 - 1.282 x 0.10) = P50 x (1 - 0.1282) = P50 x 0.872, so P90 is 12.8% below P50.

The 4 percentage points of extra uncertainty cost 5.1 percentage points of P90 yield.

Combining Uncertainty Sources

Independent uncertainty sources combine by root-sum-square, not by simple addition. You square each one, add them up, and take the square root. Irradiance data is one source. Others include year-to-year variability, the transposition model, the PV model, degradation and site losses. For three sources of 3%, 4% and 2%, the combined figure is the square root of (9 + 16 + 4) = square root of 29 = 5.4%, not 9%. Those three percentages are examples, not typical values. Take the real ones from your data provider’s validation report and your simulation tool’s documentation.

Satellite vs. Ground Measurement Data

Satellite-derived data (PVGIS SARAH, Solargis, NSRDB) uses images from geostationary satellites to estimate surface irradiance. These satellites stay over one spot on Earth. The data offers steady coverage and long time series. The JRC lists conditions that can reduce its accuracy. Examples are snow mistaken for clouds and dust storms that are hard to detect. It also notes that geostationary satellites do not cover polar areas.

Ground station data (pyranometer measurements) measures irradiance directly at one point. It is the most direct measurement where the instrument is well kept. But stations are few, and a point measurement may not represent a site tens of kilometres away.

A common approach is to use satellite data as the baseline. Then compare it with the nearest good ground station. Ask your data provider for the validation statistics that apply to your region.

The generation and financial tool in SurgePV covers energy-yield and financial modeling. A guided demo shows how it treats weather data and loss assumptions.

Using Irradiance Data in Solar Design Software

In daily practice, solar designers work inside software that automates most of these calculations. What matters is knowing how the software handles irradiance data, where to step in, and how to check the output.

From Irradiance to Yield: The Calculation Chain

A modern solar design platform follows a consistent chain.

1. Location and irradiance data import. You enter the project coordinates. The software gets irradiance data from its built-in database or a source you select. You get GHI, DNI, DHI and air temperature as hourly or monthly values.

2. Array geometry. You set the array tilt, azimuth and any tracking. For rooftop systems this comes from the 3D roof model. For ground mounts you set the values directly.

3. Shading analysis. Shadow analysis tools calculate the shade that nearby objects cast on the array through the year. Shading changes the irradiance reaching each part of the array. It affects the beam and diffuse light differently. A building casting a shadow on winter mornings blocks direct beam, while diffuse light is only partly affected.

4. Transposition. The software applies a transposition model (often Perez). It converts horizontal irradiance to POA irradiance on the tilted, angled and possibly shaded array.

5. Module power calculation. The software calculates DC power for each time step from the POA irradiance and cell temperature. It uses the module’s temperature coefficient, spectral response and low-light behavior.

6. System losses. The software applies DC and AC cable losses, mismatch, inverter efficiency, transformer losses, soiling and degradation in turn. This gives the net yield.

7. Energy yield output. The result is annual energy in kWh and specific yield in kWh/kWp. Reports usually show monthly values with an annual total.

Where Designers Should Intervene

Data source selection. Do not accept a default without a check. Find out which dataset and period the software uses for your region. Compare it with a second source.

Horizon profile. If the site has far-off shading from hills or tall buildings, enter or import a horizon profile. PVGIS, for instance, has a horizon option calculated from a digital elevation model (DEM, a map of ground height). Its API reported “use_horizon: true” for the runs on this page.

Albedo. The PVGIS runs on this page use a ground albedo of 0.2 (the reflected part of the Munich example matches that value). If your site has light gravel or seasonal snow, the reflected part rises. This matters most for bifacial modules. They also collect light on the back side. Read the Bifacial Solar Panel Design Guide for a walkthrough.

Soiling assumptions. Set soiling, the loss from dirt on the panels, from local conditions and cleaning schedules. Record the value you used.

Validating Your Simulation

After a run, check the result against independent anchors.

• Compare specific yield with reference values. Use the city table above as a sanity check for Europe. Say a simulated Munich yield is far from about 1,100 to 1,150 kWh/kWp for a well-aimed system. Then look for a data or modeling error.
• Check the performance ratio. Compare your result with the ratio your simulation tool reports. Look into any value that sits far from similar projects.
• Check the monthly shape. Solar production in northern Europe changes a lot with the seasons. A flat monthly profile points to a data error.
• Cross-reference with a second tool or dataset. Differences between two runs should be explainable from their different inputs.

SurgePV covers roof modeling, shading analysis and energy-yield modeling in one workflow, so designers do not have to move data between separate irradiance tools, CAD software and simulation engines. A guided demo shows the workflow on your own project.

Conclusion

Solar irradiance is the base of every energy yield prediction. A rough estimate becomes one you can defend through four things. They are your choice of data source, GHI decomposition, your choice of transposition model, and an honest uncertainty estimate.

Three actions to take from this guide:

1. Audit your data source. Know which dataset, version and period your software uses, and cross-reference a second source for every project that matters.

2. Check your transposition model. Confirm that your simulation software uses Perez or Hay-Davies, not the isotropic model. Record which one it is.

3. Quantify uncertainty explicitly. Show P50 and P90 with the uncertainty parts listed, not a single yield number.

SurgePV publishes this guide and sells one of the design tools it mentions. The irradiance figures come from PVGIS and the model descriptions from the sources linked in the text.

Frequently Asked Questions

Which irradiance is used for PV design?

Fixed-tilt PV design starts from GHI (global horizontal irradiance). It converts GHI to plane-of-array irradiance for the tilt and azimuth of the array. DNI (direct normal irradiance) is the main input for trackers and concentrating solar. DHI (diffuse horizontal irradiance) matters most where skies are often cloudy. In the PVGIS data used on this page, diffuse light is 31% of annual GHI in Madrid and 54% in London.

Is DNI higher than GHI?

It depends on the climate. DNI is measured on a plane that always faces the sun. So in sunny regions the annual total is larger than GHI. In PVGIS 5.3, Madrid has 2,131 kWh/m2 of DNI against 1,779 kWh/m2 of GHI. In cloudier places it can be lower. London has 988 kWh/m2 of DNI against 1,070 kWh/m2 of GHI.

What is the difference between GHI and DNI in solar energy?

GHI (Global Horizontal Irradiance) is the total radiation on a flat horizontal surface, combining direct sunlight and diffuse sky radiation. DNI (Direct Normal Irradiance) is only the direct beam, measured on a surface perpendicular to the sun’s rays. The two are linked by GHI = DNI x cos(solar zenith angle) + DHI.

How do I find solar irradiance data for my location?

The European Commission’s PVGIS gives GHI, DNI and DHI for a map coordinate, with monthly values, across Europe, Africa and most of Asia. The Global Solar Atlas gives a quick map view. Solargis and Meteonorm publish their own datasets, so check their sites for terms and pricing. State the source, version and period in every yield report. Cross-check a second dataset on any project that matters.

How does cloud cover affect solar irradiance?

Cloud blocks the direct beam, so DNI falls sharply. Part of the light still reaches the ground as diffuse radiation. So DHI is the larger share of what remains. That is why the diffuse share of annual GHI is higher in cloudy cities. It is 54% in London and 52% in Brussels and Amsterdam, against 31% in Madrid (PVGIS 5.3, 2005-2023 averages).

What is a good annual GHI for a solar installation?

There is no single cut-off, because yield depends on tilt, shading and losses as well as GHI. In the PVGIS data on this page, annual GHI runs from 990 kWh/m2 in Stockholm to 1,800 kWh/m2 in Athens. The yield of a 1 kWp south-facing system at its optimal tilt runs from 991 to 1,627 kWh a year across the same 20 cities.

Sources

• European Commission JRC, Using PVGIS 5 (version 5.3, SARAH-3 2005 to 2023). Figures in the tables and the Munich example come from the PVGIS 5.3 API (MRcalc, seriescalc and PVcalc tools), run on 28 October 2026.
• Sandia PV Performance Modeling Collaborative, Irradiance and insolation
• pvlib documentation, decomposition models
• Solargis, How to calculate P90 or other Pxx PV energy yield estimates