Persistence Forecast

A persistence forecast assumes that the power output or meteorological variable at time t remains constant for all future lead times t+k. For a wind farm, this means the current wind speed is projected forward unchanged. For solar, the current Global Horizontal Irradiance (GHI) is held constant. This model requires no training data, no physics, and no computation, making it the universal reference point for evaluating forecast skill.

The Forecast Skill Score quantifies a model's value by comparing its error to the persistence baseline. The formula is: Skill = 1 - (RMSE_model / RMSE_persistence).

• A skill score of 0% means the model performs identically to persistence.
• A positive skill score indicates improvement over the naive baseline.
• A negative skill score means the model is worse than simply assuming no change, indicating a failed model.

The accuracy of a persistence forecast degrades rapidly as the forecast horizon extends. For very short-term predictions (e.g., 5-15 minutes ahead for solar irradiance during clear sky conditions), persistence can be surprisingly competitive. However, for day-ahead forecasts, persistence becomes virtually useless because it cannot anticipate diurnal solar cycles, frontal passages, or the nocturnal low-level jet dynamics that drive wind generation.

In solar forecasting, raw persistence is often outperformed by smart persistence, which multiplies the current Clear Sky Index by the theoretical clear-sky irradiance for the target time. This accounts for the deterministic diurnal cycle of the sun.

• Raw persistence: Holds GHI constant (fails at sunrise/sunset).
• Smart persistence: Holds the cloud attenuation factor constant while the sun angle changes realistically. This provides a much tougher baseline for advanced models to beat.

Persistence serves as a critical member of the verification hierarchy for operational forecasting suites. When evaluating a new Numerical Weather Prediction (NWP) model or a Long Short-Term Memory (LSTM) network, the first diagnostic check is always: 'Does it beat persistence?' If a sophisticated machine learning pipeline cannot outperform a zero-cost naive assumption, it indicates a fundamental flaw in feature engineering, data leakage, or model architecture.

Persistence forecasts catastrophically fail during irradiance ramp events or wind gusts. A ramp rate, measured in W/m² per minute, quantifies sudden power changes caused by moving clouds or turbulent wind fields. Since persistence assumes a flat-line continuation, it will miss the onset, magnitude, and duration of the ramp entirely. This is precisely why Cloud Motion Vector (CMV) techniques and Temporal Convolutional Networks (TCN) are deployed to capture advection and non-linear temporal dynamics.

The definitive metric quantifying a model's improvement over the persistence baseline. Calculated as 1 - (RMSE_model / RMSE_persistence), a positive skill score indicates the model adds value beyond simply assuming tomorrow will be like today. A score of zero means the sophisticated model performs no better than the naive baseline, while negative scores signal a model that actively degrades prediction quality.

A normalization metric that isolates cloud-driven attenuation from the deterministic solar geometry. Defined as the ratio of actual Global Horizontal Irradiance (GHI) to the theoretical irradiance under a cloudless sky, this index transforms the forecasting problem from predicting raw irradiance to predicting the stochastic cloud field. Persistence forecasts applied to the clear sky index often outperform raw GHI persistence.

Unlike the single deterministic value produced by persistence, probabilistic forecasts output a full probability distribution or set of quantiles. This explicitly communicates the inherent uncertainty that persistence ignores. Grid operators use these distributions to set dynamic operating reserves, holding more backup capacity when forecast uncertainty is high and less when conditions are stable.

A strictly proper scoring rule that evaluates probabilistic forecasts by measuring the integrated squared difference between the forecast cumulative distribution function (CDF) and the empirical observation. Unlike simpler metrics, CRPS simultaneously assesses both calibration (are probabilities reliable?) and sharpness (how concentrated is the distribution?). Persistence-based probabilistic forecasts serve as the baseline CRPS reference.

The rate of change of solar irradiance over time, typically measured in W/m² per minute. This metric quantifies the sudden power fluctuations caused by moving clouds that persistence forecasts fundamentally fail to capture. A persistence model predicts zero ramp rate by definition, making it useless for the ramp event detection critical to grid stability. Specialized ramp forecasting models must significantly outperform this naive baseline.

A recursive Bayesian algorithm that optimally estimates a dynamic system's state from noisy measurements. In forecasting, it serves as an adaptive post-processor that corrects systematic biases in real-time. Unlike static persistence, a Kalman filter continuously updates its error covariance matrix as new observational data arrives, blending the persistence assumption with live sensor feedback to produce a dynamically corrected estimate.