Bilkent University • CS491 Senior Design Project

Multi-Factor Generation, Revenue and Risk Forecasting for Energy Producers

Multi-factor generation forecasting, imbalance cost optimization and quarterly revenue analysis for electricity producers. Weather, plant availability, grid instructions and market prices from EPİAŞ are combined in a single model.

How it works Early results

Şah RES (105 MW) — Balıkesir

24-hour generation actuals and model comparison

Illustrative data. KGÜP: the operator's finalized day-ahead production plan, published by EPİAŞ.

Problem

Forecast error is a financial problem

As the renewable share grows, every megawatt-hour of deviation shows up on the producer's balance sheet and in what investors can see.

Problem

Imbalance penalties

Wind and solar now exceed 22% of generation. Every forecast error turns into imbalance penalties that come straight out of the producer's margin.

22%+

Wind and solar share

Problem

Delayed company data

Investors and analysts only see plant performance in KAP filings, 1-2 months after the quarter has already closed.

1-2 months

Reporting lag

Approach

What GridLens does

A bidding engine that minimizes imbalance cost, intraday correction suggestions and weekly company scorecards.

Weekly

Company scorecards

How the model works

Four inputs, one forecast

Weather
Wind speed, direction and temperature enter the model as numerical weather forecasts adjusted to hub height.
Plant availability
EAK capacity declarations and free-text PMS outage notices are parsed; unplanned mechanical failures are separated out and the capacity loss is calculated.
Grid and market
YAL/YAT dispatch instructions, the day-ahead price (PTF) and imbalance risk determine how much generation reaches the grid and at what price.
Forecast and bid
All inputs are combined in one model, which produces a net generation forecast and a day-ahead bid that minimizes the expected imbalance cost.

Weather

Wind speed, direction and temperature enter the model as numerical weather forecasts adjusted to hub height.

Weather1 of 4

Wind speed

9.4m/s

Direction

225°

Temperature

14.0°C

Wind speed forecast, next 48 hours

Plant availability2 of 4

Available capacity (EAK)92 / 105 MW

“BRS - Gear box arızası, kanat tamiri”

Outage notice: gearbox failure, blade repair

UnplannedMechanical failure−13 MW

Grid and market3 of 4

Down-regulation (YAT)

−20MW

Imbalance risk

Medium14:00-17:00

Hourly day-ahead price (PTF), normalized

Forecast and bid4 of 4

Net generation forecast with P10-P90 band

Forecast Proposed bid

Proposed day-ahead bid

1,284MWh

Optimized for

Lowestimbalance cost

System

Six components, one data model

Each component handles one input or decision. All of them read from the same EPİAŞ and weather data and feed the same forecasting model.

Weather data

Hourly wind speed, direction and irradiance forecasts, adjusted to turbine hub height.

Wind speed (100 m)

9.4m/s

Direction

225°

Temperature

14.0°C

Irradiance

412W/m²

Hour	m/s	Dir.
06:00	7.8	210°
09:00	9.4	225°
12:00	11.2	232°
15:00	8.6	240°

Source: Open-Meteo · hourly · 0-48 h ahead

Outage notice parser

Reads free-text PMS and EAK notices and classifies them, for example as an unplanned mechanical failure.

Outage notice (original, Turkish)

“BRS - Gear box arızası, kanat tamiri”

Gearbox failure, blade repair

CategoryMechanical failure

Planned / unplannedUnplanned

ComponentsGearbox, blade

Source: EPİAŞ PMS notices

Bid optimization

Optimizes the day-ahead bid against the imbalance cost in TRY rather than the forecast error in MWh.

Producing less than the bid is penalized more heavily than producing more, so the best bid is not simply the expected generation.

Inputs: generation forecast, PTF and SMF prices

Quarterly revenue estimate

Company-level generation and revenue estimates while the quarter is still running, published as a weekly scorecard.

Weekly generation, current quarterWeek 8 of 13

Actual Estimate

KAP filing

30-60days after quarter end

GridLens

Weeklyduring the quarter

Sources: EPİAŞ generation data, KAP filings

Energy assistant

Answers natural-language questions by querying EPİAŞ and model data, and cites the data behind each answer.

Why did Galata Wind produce less than expected last month?

The shortfall is broken down into reported outages, grid curtailment instructions and weather deviation, with each figure linked to its source.

EPİAŞ real-time generationPMS outage noticesGridLens forecast

Data: EPİAŞ, PMS notices, GridLens forecasts

Battery storage scenarios

Arbitrage and scenario analysis for storage, in line with Turkey's 33 GW storage target.

Hourly day-ahead price (PTF), normalized

Charge at low prices Discharge at high prices

Input: hourly PTF price profiles

Early results

First test on Şah RES

Forecast error relative to installed capacity; lower is better. The first model halved the baseline error. The next step is adding outage and grid curtailment data.

Persistence baseline

22.1%

Naive reference: each hour is forecast as the previous hour's generation.

GridLens first model

11.2%

Error cut roughly in half compared with the baseline.

Operator's KGÜP plan

7.0%

Target: close this gap by adding outage and curtailment data.

Team

The people behind GridLens

Ece Bulut

Computer Engineering

AÖ

Ahmet Fatih Öztürk

Computer Engineering

İS

İzzet Yağızefe Sungurlu

Computer Engineering

Yunus Tan

Computer Engineering

Berke İsmail Erhan Asıl

Computer Engineering

Bilkent University, Department of Computer Engineering