Renewable power forecasting is the practice of predicting, in advance, how much electricity a wind or solar plant will produce over the coming minutes, hours, and days. Because the fuel for these plants is the weather rather than a controllable fuel supply, output cannot simply be dialled up when needed, and grid operators and plant owners must instead anticipate it. This guide explains the forecast horizons that matter, the mix of weather models, plant history, and machine learning that produces a forecast, the error metrics used to judge one, and how live plant data is fed back to keep the forecast honest.
Power forecasting in one line: Renewable power forecasting predicts the future electrical output of wind and solar plants across horizons ranging from minutes ahead to days ahead. It combines numerical weather prediction, historical plant behaviour, and statistical or machine-learning models that map forecast weather onto expected megawatts. These forecasts drive market bids, reserve scheduling, and curtailment decisions, and their accuracy is tracked with error metrics such as mean absolute error and root-mean-square error.
Renewable forecasts are organised by how far ahead they look, because a different decision depends on each horizon and a different technique works best. The shortest is the nowcast, covering the next minutes to a couple of hours, which is dominated by what is happening right now: the wind speed currently hitting the plant, the cloud edge visible on a sky camera or satellite image, and the plant's own recent output. Over such short spans, persistence and rapid pattern-tracking beat a slow weather model, so nowcasting leans heavily on live measurements rather than on a forecast computed hours ago.
The intra-day horizon, from a couple of hours out to the rest of the trading day, and the day-ahead horizon, covering tomorrow, are where numerical weather prediction takes over. Here the plant's near-term measurements matter less and the evolving weather pattern matters more, so the forecast is anchored to the output of large atmospheric models. Day-ahead forecasts are the ones that feed the wholesale electricity markets in which most plants must declare tomorrow's expected generation, and errors at this horizon translate fairly directly into imbalance costs when the plant produces more or less than it promised.
Because different horizons serve different decisions, a plant typically runs several forecasts at once and blends them. A short-term correction from live data can be layered onto a longer-term weather-driven forecast so that the near hours reflect current conditions while the outer hours reflect the broader pattern. Understanding which horizon a number came from is essential, because a day-ahead forecast and a fifteen-minute-ahead forecast for the same hour can and should differ.
At the base of most forecasts sits numerical weather prediction, in which physics-based models simulate the atmosphere on a grid and predict wind, irradiance, temperature, and cloud cover over the coming days. These models are run by meteorological agencies and commercial providers, and their raw output is a weather forecast for the region, not a power forecast for a specific plant. Turning one into the other requires local knowledge, because the wind that a grid cell predicts is not exactly the wind that a particular turbine at a particular height on a particular ridge will feel, and the irradiance over a region is not exactly what a specific tilted panel array will capture.
That translation from weather to power is where a plant's own history and machine learning come in. By studying how the plant has actually responded to past weather, a model learns the site-specific relationship between forecast conditions and delivered megawatts, absorbing effects that raw physics misses: terrain, wake losses between turbines, panel soiling, temperature derating, and the plant's own control behaviour. Modern forecasting stacks often ensemble several weather models and several statistical mappings together, because averaging diverse forecasts usually beats trusting any single one, and machine-learning layers are used to correct systematic biases the physics leaves behind.
None of this makes the weather deterministic, so good forecasting increasingly means probabilistic forecasting: not just a single expected output but a range with confidence attached. A forecast that says output will most likely be a given value but could plausibly fall within a stated band is far more useful for scheduling reserves than a bare point estimate, because it tells the operator how much backup to hold. The spread of that band is itself a forecast product, widening when the atmosphere is unpredictable and narrowing when conditions are settled.
A forecast is only worth what its track record says, so forecast quality is measured continuously against what the plant actually produced. The two most common metrics are mean absolute error, the average size of the miss regardless of direction, and root-mean-square error, which squares the misses before averaging and so penalises large errors more heavily than small ones. Both are usually expressed relative to the plant's capacity so that forecasts for plants of different sizes can be compared. Bias, the tendency to systematically over- or under-predict, is watched separately, because a forecast that is right on average but always high in the afternoon still causes trouble that a single error number can hide.
The raw material for both making and grading forecasts is the live plant data flowing through the SCADA system: the actual power at the meter, the measured wind speed or irradiance, the temperature, and how many turbines or inverters are actually available. Feeding this measured output back to the forecaster closes the loop. The nowcast uses the latest output directly, the statistical models retrain on the growing history of forecast-versus-actual pairs, and any systematic bias the models develop is caught and corrected because yesterday's error is visible today. A forecast cut off from live measurements drifts; one fed a steady stream of ground truth stays calibrated.
This is why a cloud SCADA platform is a natural home for the data side of forecasting. When meter output, resource measurements, and availability status all stream into one hosted historian, a forecasting provider can pull a clean, continuous ground-truth signal to train and correct against, and the plant's own team can trend forecast against actual to hold the provider accountable. Merobix serves this pattern of gathering field measurements into a single live record that many parties can read; and while its core focus is oil and gas, the same need to reconcile a prediction against streamed real-world output recurs across power, water, and industrial operations.
A day-ahead forecast predicts tomorrow's output, usually to declare expected generation into a wholesale market that clears the day before delivery. An intra-day forecast covers the hours remaining in the current day and is refreshed as new weather and plant data arrive, so it can correct the day-ahead view. Day-ahead relies mostly on weather models, while intra-day and shorter forecasts increasingly blend in live measurements.
Measurement tells you what the plant is producing now, but grid operators and markets need to know what it will produce hours or a day ahead so they can schedule other generation and reserves in advance. Because wind and solar depend on weather rather than a controllable fuel, that future output must be predicted rather than commanded. Forecasting fills the gap between what can be measured now and what must be planned for later.
There is no single universal number, because forecast error depends heavily on the horizon, the plant, and the local climate, and it is normally reported relative to plant capacity. Errors grow with the forecast horizon, so a nowcast is far more accurate than a day-ahead forecast, and a settled climate is easier to predict than a stormy one. The meaningful comparison is a given forecast against a naive baseline such as persistence, and against alternative providers on the same site.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.