Forecast accuracy measures summarize differences between predicted and observed generation. They help teams compare forecasts, but a single metric cannot answer every operational question.
Mean absolute error, or MAE, describes the average size of absolute errors. Because errors are not signed in the average, over- and under-forecasting do not cancel each other out.
Root mean square error, or RMSE, also summarizes error size. Squaring errors before averaging gives larger misses more influence, which can make RMSE useful when large errors carry greater operational concern.
Normalized RMSE, or nRMSE, scales RMSE against a reference quantity. The chosen denominator changes how the result should be read. A comparison should state that denominator rather than treating every nRMSE value as equivalent.
Metric choice should follow the decision. A portfolio team may care about overall error, while a grid or trading workflow may also need to inspect particular regions, horizons, ramp periods or error direction.
The comparison is only as useful as its inputs. Match forecast and actual timestamps, define the observed period, check missing or corrected measurements, and state the baseline used for comparison.
When reviewing a provider, ask for the technology, region, horizon, sample period, data selection, aggregation method and metric definition behind each result. Without this context, a benchmark cannot show how a forecast will behave for your assets.
Calculate the error metrics
Let pᵢ be predicted power and aᵢ measured power at the same timestamp, for n observations. MAE = (1/n) × Σ |pᵢ − aᵢ|. RMSE = √[(1/n) × Σ (pᵢ − aᵢ)²]. Both have the same unit as the power values, such as MW.
For a capacity-normalized comparison, nRMSE = RMSE / P_rated × 100%. State this denominator explicitly. A result normalized by average generation or an observed range describes something different and should not be compared directly.
A small worked example
For an illustrative three-observation example, predictions of 10, 20 and 30 MW against measurements of 12, 18 and 35 MW produce absolute errors of 2, 2 and 5 MW. MAE is 3 MW and RMSE is √11, approximately 3.32 MW. With a rated capacity of 50 MW, capacity-normalized RMSE is approximately 6.63%. These are invented teaching values, not enercast results.
Define the evaluation before collecting results
Fix forecast issue times and lead times before calculating scores. Do not substitute forecasts issued after the operating decision. Align timestamps, units, meter intervals and daylight-saving treatment. Explain how missing observations, outages and curtailment are handled, and keep the same valid observations for every provider or baseline.
Inspect results by horizon, asset and operating condition alongside the overall score. Portfolio aggregation can hide large site-level errors. Comparing a forecast with persistence or another agreed baseline helps show whether the forecast adds value for the task being evaluated.