what does an error of 0.19 mean?

It is the typical size of a miss divided by the size of the thing being measured. So 0.19 means the typical miss was about 19 percent of the average value. Lower is better. In this study 0.12 in the dry season was the best figure and 0.50 for the wet season direct beam was the worst.

do I need a sensor on the roof?

No, and that is the point. The study used ordinary weather records, then fed the predicted sunlight into PVLib with the specs of a real panel. No pyranometer sat on the roof. The trade is that nobody has yet checked these forecasts against a real meter either.

can I use the same method outside Ibadan?

The paper does not show that. It covers one set of coordinates at the University of Ibadan. You would need local weather records and a fresh run for any other town, and you should measure your own error there before quoting anyone a number.

why send the first messages by hand?

Because you are buying information, not saving time. Typing the message yourself every evening for a month shows you what the operator does with it, which words confuse him, and whether the flag ever changes a decision. Build software after somebody pays.

Sell a day-ahead sun forecast to Nigerian mini-grid operators

The finding this rests on

Three researchers at the University of Ibadan in Nigeria and the University of Venda in South Africa asked a simple question. Can you work out how much sunlight hits the ground without a sensor on the roof?

They took hourly weather records for one spot in Ibadan, running from 2005 to 2022. Then they trained three models to turn heat, damp air, wind and cloud into a sunlight figure. A Random Forest, the plainest of the three, beat both neural networks. Over a full year its typical miss on total sunlight was about 19 percent of the average value. The forecasts then went into PVLib, a free Python tool, with the specs of two real panels, to give an energy figure.

The seasons split hard. The error was 0.12 in the dry season and 0.27 in the wet season. The direct beam in the wet season was worst, at 0.50. Read the full write-up at Weather data alone can forecast solar power in Ibadan.

We score that paper 4 out of 10. It is a preprint and it covers one set of map coordinates. The strong point is openness, because the data and the full Python code are public on GitHub. So this idea is a bet on one site’s result holding at other sites.

Who has this problem

Picture an operator who runs a small solar mini-grid for a market town near Ibadan. Two hundred shops and homes are wired to it. There are panels, a battery bank and a diesel generator for the bad days.

Every evening he makes one decision. Does he start the generator tonight, or does he trust the battery to last until the sun comes up?

Right now he decides by looking at the sky and by how the last few days went. If he starts the generator and does not need it, he burns fuel he paid for. If he does not start it and the morning is grey, the town wakes up dark and somebody stops paying their bill.

The installer down the road has a cousin of the same problem. A customer asks how much power a system will make. He answers with a national average. That average does not fit one street, and it does not fit June.

What you would sell them

Sell one message a day. At six in the evening the operator gets a short note that says how much sunlight is expected tomorrow, hour by hour, and what that means in units for his own array.

It carries one flag. Tomorrow looks normal, or tomorrow looks far below normal. That is the line he acts on when he decides about the generator.

Sell the installer a different thing from the same engine. Give him a one page yield sheet for a single address, with a dry season number and a wet season number, so he can put both in a quote.

What to charge

Test the daily message at 30 dollars a month for each site. Test the installer sheet at 15 dollars a report. Both are hypotheses, and you will only learn the truth by asking for money.

Here is the reasoning for the mini-grid price. Do not compare it to software. Compare it to diesel. Ask the operator what one night of running the generator costs him in fuel. If your monthly fee is smaller than one avoided night, he does not need a spreadsheet to decide.

The installer price has a different anchor. Ask him what one lost quote costs him. A sheet that wins one extra job pays for a year of sheets.

How you would build the first version

The parts are already on the shelf, and they are free.

Take the authors’ public code and data from GitHub. Use PVLib for the panel step, and type in the panel model the client actually has. Train the Random Forest first, because it won here and it is the easiest of the three to run. Get one site working end to end before you touch a second.

Send the daily message by hand. Type it into WhatsApp yourself at six in the evening, every evening, for a month. This is slow on purpose. You will learn what the operator does with it.

Do not build an app. Do not build a dashboard, a login page, or an interface for other developers. Do not buy a sensor. Do not sign up a second town until the first one pays.

The one-week test

Day one. Pull the public code and free weather records. Get one hourly forecast running for one Ibadan address on your own laptop.
Day two. Write the evening message by hand for that address. Keep it to four lines. Show it to one person who is not in this business and check that they understand it.
Day three. Find five mini-grid operators and three installers. Call them. Ask what they do today when they need to know about tomorrow’s sun.
Day four. Send each of them a real message or a real yield sheet for their own site, free, with no attachment and no login.
Day five. Ask all eight for the first month of money, at 30 dollars or 15 dollars. Take mobile money on the call.

By Friday you have a yes or a no. Yes means at least two people paid. No means everyone said it was interesting. Interesting is a no.

What would kill this

The study is one place. Every number here comes from a single set of coordinates at the University of Ibadan. The paper shows nothing about Kano, Lagos or a village fifty kilometres away. Your first job at any new site is to measure your own error, not to quote 0.19.
You are worst when you matter most. The error nearly doubles in the wet season, from 0.12 to 0.27, and the direct beam reaches 0.50. The wet season is exactly when the operator needs the warning. A service that is vague in June will not be trusted in December.
Nobody checked it against a real meter. The forecasts were scored against a weather database, not against a sensor in Ibadan and not against the metered output of an installed system. Your customer will judge you on his meter. That comparison has not been made yet by anyone.
Day ahead is a harder job than the paper did. The study used recorded weather to work out sunlight. To sell tomorrow, you must feed it tomorrow’s weather forecast, which carries its own errors. Your real accuracy will be worse than 0.19, and you will not know how much worse until you run it live for a month.
The forecast may change nothing. If the operator starts the generator on grey mornings anyway, your message is a nice note that saves him no fuel. Watch what he does, not what he says.
The parts are free. The code and the data are public. A customer with one clever intern can copy you. Your only durable edge is the daily habit and the local trust, which is why you send the message by hand at first.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.