Yarın Türkiye Ne Kadar Elektrik Tüketecek? Gün Öncesi Talep TahminiHow Much Electricity Will Turkey Use Tomorrow? Day-Ahead Demand Forecasting

Electricity cannot be stored; every hour, as much has to be generated as is consumed. That is why the next day's consumption is forecast a day ahead, hour by hour. In this study I built such a forecasting model from 85,584 hours of Turkey's consumption data and tested it the way it would be used in real life, each day with only what was known up to that day.

The question and the rule

The target is Turkey's total electricity consumption for each of the 24 hours of tomorrow (day D). The rule: when forecasting day D, the model may see consumption only up to the end of day D−2. The day in between is left out because the forecast has to be delivered a day in advance.

As the error measure I used the mean absolute percentage error (MAPE): the average, over hours, of how far the forecast is from the actual value in percent.

Data

• Consumption: The real-time consumption series on the EPİAŞ Transparency Platform, January 1, 2017 – October 6, 2026. 85,584 hours with no missing or duplicated hour. Turkey has not used daylight saving time since 2016, so there is no clock-shift problem either.
• Temperature: Hourly temperature for Istanbul, Ankara, Izmir, Adana and Erzurum from the Open-Meteo archive.
• Calendar: Public and religious holidays; from these I derived holiday eves, the day after, bridge days and the month of Ramadan.
• Official forecast: The Load Forecast Plan published on the same platform, the official hourly forecast for the next day. I used it as a benchmark.

What drives consumption?

I looked at the series before modelling; a good forecasting model has to know what moves consumption.

Hour of the day and day of the week

On working days consumption bottoms out at 05:00, climbs quickly after 08:00 and peaks at 11:00. On weekends the morning rise is slower and the peak comes at 20:00. On Sundays consumption is 12% below average.

Temperature

The curve is U-shaped. Consumption is lowest between 14 and 18 °C (−8%); as it gets hotter the air-conditioning load arrives, and above 26 °C consumption is 16% above the yearly mean. The rise in the cold is much smaller (+3%), because most heating runs on natural gas, not electricity. The record in the series is also a summer afternoon: July 28, 2025, at 14:00, 59,504 MW.

Holidays

On religious holidays consumption is 28% lower than on a normal day; 30.1% on the first day and 23.5% on the fourth. Even the eve is 19% lower. Because these holidays move about 11 days earlier every year, seasonality cannot capture the effect; the calendar has to be given to the model explicitly.

The series also carries a trace of the pandemic: consumption in April 2020 was 15% lower than in the same month a year earlier, and 17% lower in May 2020.

Model

I used gradient-boosted decision trees (scikit-learn, HistGradientBoostingRegressor). What the model is given:

• Past consumption: The same hour 2, 3, 7 and 14 days earlier, plus the level and trend of the last week.
• Calendar: Hour, day of week, day of year, holiday type, which day of the religious holiday, eve, bridge day, Ramadan.
• Holidays in the past: How many of the reference days were holidays? Without this, in the week after a holiday the model takes the holiday's low consumption for “normal”.
• Temperature (in the second model): The temperature of the forecast day and its difference from the last week.

Backtest

I did not test the model with a single train and test split. From January 2022 to September 2026, at the start of every month I retrained it on the data known up to that day and had it forecast only that month: 57 months, 1,734 days, 41,616 hours. I did not change the model settings after looking at test results.

Results

• The simple baseline is not a hard bar to clear, but it is a necessary one. “The same hour last week” gives an error of 5.2%. The model brings that down to 2.1%.
• On 77% of days the daily error is below 2%. The median daily error is 1.27%; a small number of bad days pull the mean up.
• Temperature cuts the error by a fifth, but that figure was computed with the actual temperature of the forecast day. Below I look at how optimistic that is.

The real difference is on holidays

On normal days the model halves the baseline's error. On religious holidays the baseline is off by 38%, because it copies last week's normal day onto the holiday; the model stays at 4%. The hardest day is the eve: people work half a day, and it happens only twice a year.

The chart shows two things. The baseline is far too high during the holiday and far too low in the week after, because then it copies the holiday week. The model follows the actual values in both; but on the two days before the holiday (May 25–26) it, too, stays high. The reason is below.

How much does temperature help?

Temperature helps most in June and July: the air-conditioning load is very sensitive to the weather. In mild months such as May and October the two models are close.

To see where in between, I deliberately added error to the temperature: a random offset for each day, constant through that day. A model trained on data before 2025 gave these results for 2025–2026:

Even if the weather forecast is off by 2 °C a day on average, using temperature beats not using it. At 3 °C the advantage is gone. Next-day weather forecasts are usually more accurate than that, so temperature is worth keeping in the model.

Comparison with the official forecast

EPİAŞ publishes an official hourly forecast for the next day, every day: the Load Forecast Plan. I put the two side by side over the same 41,592 hours, against the same actual values.

On this data the model had a lower error than the official plan even without temperature, and was more accurate on 71% of days. The result points the same way in each of the five years.

How sure is the forecast?

An interval is often more useful than a single number: “tomorrow at 14:00, consumption will be between these two values with 80 percent probability.” For this I trained the same model twice more, to predict the 10% and 90% quantiles. A good interval should contain 80% of the actual values.

The raw interval came out too narrow: only 65% of the actual values fell inside. The model is more confident than it should be. To correct this, each month I widened the interval using the test errors of the previous twelve months (a conformal method). Coverage rose to 81%; the mean width of the interval is 5.6% of consumption.

The correction fixes the average but not every day type: on holiday eves coverage is 59%. The days on which the model is most wrong are the days on which it knows least that it might be.

The days the model got most wrong

On all nine of the worst days the model forecast consumption too high, and they fall into three groups:

• Unforeseeable events. The planned power cuts to industry during the natural gas shortage of January 2022, and the earthquakes of February 2023. No model that looks at past consumption can know these in advance.
• Holidays that are not in the calendar. When a religious holiday is extended to nine days by administrative leave, the working days in between become holidays in practice. The calendar I used does not contain them.
• Overlapping holidays. May 1, 2022 was both a holiday eve and a public holiday; the data had never contained such a day before.

The second group can be fixed: administrative leave is usually announced weeks ahead and can be added to the calendar by hand. That is the first improvement to make.

Limitations

Conclusion

Most of a good demand forecast came not from the choice of model but from three things: setting the target correctly (a ratio to the last week instead of megawatts), giving the calendar to the model explicitly, and testing the model the way it will be used, each day with only what was known up to that day.

The mean error is not a good summary. The model is off by less than 2% on most days; the real risk is concentrated in a few days a year, and most of those are about the calendar. For someone using the forecast, a more useful question than “what is the average error” is “on which days should I not trust it”.

Data: EPİAŞ Transparency Platform (real-time consumption and load forecast plan), Open-Meteo (temperature). Test period January 1, 2022 – September 30, 2026.