Electricity load forecasting is the practice of predicting how much electricity a site, building or distribution area will need, hour by hour or day by day, using past consumption, production plans, the calendar and weather data. It lets you manage power purchasing, contracted capacity, shift planning and the energy budget without surprises. Set up properly, a forecast is not just a chart but a decision input used every day.
Knowing your consumption versus predicting it
Many businesses see their electricity bill at the end of the month and never see their consumption at all. Those with an energy monitoring system are a step ahead: they know how much each line drew yesterday. Knowing how much the site will draw at 2 pm tomorrow, how close it will come to its contracted capacity, or where next month's bill will land, is a different job.
The answers change real decisions. An industrial site trading on the market in its own name, or its supplier, has to anticipate the next day's hourly requirement and plan its purchases accordingly. The planning team wants to know which hour to move the press line to so it avoids peak rates. Finance wants to base the annual energy budget on something firmer than "last year plus inflation".
You cannot forecast what you do not measure, so most projects begin with factory energy monitoring. Once metering is in place, the same data becomes the foundation of a forward-looking model.
The cost of leaving it unsolved
Electricity demand in Türkiye is growing, and industry accounts for a very large share of it. According to the Ministry of Energy and Natural Resources information centre, Türkiye's electricity consumption rose 2.1% to 360.9 TWh in 2025, and the National Energy Plan expects it to reach 455.3 TWh by 2030.
Industry accounted for 45% (130,271 GWh) of Türkiye's net electricity consumption in 2023. (TEDAŞ, 2024 Türkiye Electricity Distribution Sector Report)
Generation is changing too. The same Ministry source shows that in 2025 wind supplied 10.9% and solar 10.5% of Türkiye's electricity. As weather-dependent generation grows, hourly prices become more volatile, and getting tomorrow's requirement wrong becomes more expensive. The global picture is similar: the IEA's Electricity 2026 report forecasts average annual demand growth of 3.6% over 2026–2030.
For a business, operating without a forecast shows up in three places: imbalance charges, exceeding contracted capacity, and producing during expensive hours. Breaching reactive energy limits can be added to the list; we cover that in our article on the reactive power penalty.
Why electricity load forecasting matters for businesses in Türkiye
In Türkiye's electricity market, consumers whose annual consumption exceeds a threshold set by the regulator (EPDK) are classed as eligible consumers and can choose their supplier. Suppliers, and large consumers trading in their own name, bid for the next day's hourly requirement on the day-ahead market run by the market operator EPİAŞ. When actual consumption deviates from the quantity bought, the difference is settled as an imbalance, usually at a less favourable price.
A site on a fixed-price contract does not see this risk directly, but the supplier prices it in. A business that can reliably predict its load profile negotiates supply terms with stronger evidence. For sites on a two-part tariff, most of them connected at medium voltage, exceeding the contracted capacity triggers a capacity overrun charge under the applicable tariff terms; a short-term forecast can warn you before the overrun happens.
Electricity load forecasting methods by horizon
There is no single "right model"; the question is what the forecast is for. Once the horizon and purpose are set, the choice of method narrows naturally.
| Horizon | Typical use | Key inputs | Suitable methods |
|---|---|---|---|
| Very short (15 min – a few hours) | Capacity overrun alerts, load shifting | Live meter data, machines running | Moving averages, gradient-boosted trees |
| Short (next day, hourly) | Market nominations, shift planning | Production plan, calendar, weather forecast | Regression, gradient-boosted trees, LSTM |
| Medium (weeks – months) | Energy budget, supply contracts | Order book, season, maintenance calendar | Seasonal models (SARIMA), scenario analysis |
| Long (a year or more) | Capacity, transformers, solar investment | Growth plans, new lines, efficiency projects | Scenario models, expert assumptions |
Measuring accuracy with a single percentage can mislead. At sites with very low night-time load, the classic mean absolute percentage error (MAPE) exaggerates small deviations; weighted absolute percentage error (WAPE) or mean absolute error in kWh is more meaningful. Always compare the model with a simple benchmark, such as "same day and hour last week"; if the model cannot beat it, the complexity is not justified.
Six steps to set up electricity load forecasting
- Write down the purpose and horizon. Market nominations, overrun alerts or budgeting? Each needs a different resolution and a different accuracy metric.
- Complete your metering. Collect main meter data at least hourly, preferably every 15 minutes. Separate large loads (compressors, furnaces, chillers) with sub-meters; automatic meter reading is the most practical source of this data.
- Add the data that explains consumption. Production plans, shift calendars, public holidays, temperature and humidity, and maintenance shutdowns are the strongest variables. At sites dominated by buildings, occupancy and HVAC settings come to the fore. In hotels that variable is booked occupancy; see AI in hotels and hospitality.
- Clean the data. Never train a model before meter outages, communication gaps and breakdown days are flagged. The techniques in our anomaly detection guide are useful here for filtering out abnormal values.
- Build from simple to complex. Try the benchmark first, then regression, then machine learning models. Test every model on the same historical period in a way that mimics real use.
- Embed it in decisions and monitor it. The forecast should flow automatically into the planning screen, an overrun alert or the market nomination file. Track the error weekly and retrain the model when production patterns change.
For consumers within the smart meter scope, the hourly consumption data that feeds a forecasting model also comes from the distribution company's meter; we explain the scope and replacement timetable in our guide to the automatic meter reading obligation in Türkiye.
If you want to see the shared logic with sales forecasting, our article on AI demand forecasting applies the same method to product demand.
Where forecasting makes a difference: three scenarios
A multi-shift factory. An energy-intensive line produces an hourly forecast from the next day's production plan. The planning team sees the effect of moving heavy loads outside peak hours before production starts, and gets early warning of hours approaching contracted capacity.
A multi-building campus or hospital. HVAC load depends on outside temperature and occupancy. Combined with smart building automation (BMS), the forecast allows chillers to be staged in before the busiest hours. Uses beyond energy are covered in AI in hospital operations.
Distribution and utility operators. Load forecasting by region or feeder underpins capacity planning and maintenance scheduling. When field data is gathered through a remote monitoring IoT platform, the model uses the same data pipeline.
Load forecasts are only as reliable as the metering and asset data behind them. We explain the order in which to build that foundation on a distribution network in our energy utility digital transformation guide.
Common mistakes
The first mistake is expecting an hourly forecast from billing data alone. A monthly bill may be enough for budgeting, but hourly decisions need hourly metering. If you have no access to meter data, the project should start with the metering infrastructure.
The second is leaving the production plan out of the model. At most sites, what really drives consumption is which line runs for how many hours; if that information does not flow automatically from the ERP or planning system, the forecast breaks every time the plan changes. Our article on production scheduling and APS explains why keeping the plan digital and current matters.
The third is building the model once and walking away. A new machine, an efficiency investment or a change in shift pattern permanently alters the load profile. If the error rate is not tracked weekly and the model is not retrained at regular intervals, it loses credibility within a few months.
The fourth is turning the forecast into a report nobody uses. If it does not show up on the planner's screen, in the maintenance team's alerts or in purchasing's supply negotiations, its return cannot be measured. To define that return up front, see our article on measuring AI project ROI.
How we deliver electricity load forecasting at Digital Bridge
Rather than selling a package, we start with a needs analysis and set out scope, phases and cost in a written proposal. Within our energy and utilities solutions, the path looks like this:
- Discovery: together we map existing meters, billing history, tariff structure and the decision the forecast will serve. If data is insufficient, we complete the metering side first.
- Metering and data collection: where needed we install OSOS meter reading and energy monitoring and management infrastructure, working with our own hardware team on the device side.
- Pilot: we build the hourly forecasting model for a single site or building and compare it with actual consumption over several weeks. Through our demand forecasting and analytics service, the model is trained on your own data.
- Integration: we connect the forecast to planning, SCADA and HMI screens or the reporting layer, and add anomaly detection alerts for unusual deviations.
If your site uses SmartPass, time and attendance kept against the shift plan and the zone-by-zone list of who is on site at any moment can be assessed, as aggregate occupancy counts rather than per person, as candidate variables that explain consumption. It is wise to review this new use of the data with your legal team under KVKK, Türkiye's data protection law, beforehand.
For a broader view of AI projects, read AI in business: where to start, and find related articles under Artificial Intelligence.
Compared with actual consumption, a forecast model's expected load can also point to non-technical losses; we cover that use in our article on electricity theft detection.
Next step
Gather the last 12 months of bills and any meter data you have, and write down in one sentence which decision the forecast will serve. Those two things are enough to scope a pilot. To review your data with us, use our contact page.