How it works
Generation forecasts start from numerical weather prediction, the global and regional models that weather agencies and private providers run several times a day. A forecasting model converts predicted irradiance, cloud, wind speed and temperature into power for each plant, using its layout, power curve and recent behaviour. Machine-learning models trained on the plant's own history correct systematic errors. Close to delivery, live SCADA data, satellite images of cloud and sky cameras sharpen the next few hours.
Forecasts are issued for each settlement block and revised as delivery approaches. Error is measured per block, often as mean absolute error as a share of capacity. A day-ahead solar forecast might miss by 5% to 10% of capacity on average, less on clear days and more on cloudy ones. Wind is harder to forecast. Price and load forecasts follow a similar pattern, using market history, weather and calendar effects. Many forecasts are now probabilistic, giving a range such as P10 to P90 rather than one number.
Why it matters
The schedule a plant declares and the bids it places rest on the forecast. Every MWh of error becomes a deviation or imbalance charge, or a missed sale. For a battery, the price forecast decides when to charge and discharge. Gains of a point or two of capacity in accuracy show up directly in settlement. Across a portfolio, aggregating plants reduces error, because their misses partly cancel.
A worked example
A 100 MW wind farm declares 60 MW for a 15-minute block and generates 48 MW. The miss is 12 MW, or 12% of capacity, which is 12 × 0.25 = 3 MWh of energy. Suppose imbalance is charged at $40/MWh, so the block costs 3 × 40 = $120. A year has 365 × 96 = 35,040 blocks. Missing by that much in a third of them, 11,680 blocks, costs 11,680 × 120 = $1,401,600. Halving the average miss saves about $700,000.
Illustrative figures.
Across markets
In India, wind and solar plants must forecast and schedule in 15-minute blocks, with a limited number of revisions a day, and pay deviation charges outside a tolerance band. In several US markets, the system operator runs its own central forecast of wind and solar and dispatches against it. In Europe, the balance responsible party forecasts and trades intraday to close its position before gate closure.