PROPHET

Forecast of Photovoltaic Production using WRF-NWP variables

In the Objective tab the AC power measurements are displayed for the selected day and PV system (when available).

The forecast procedure predicts this time series using a set of predictors extracted from the NWP-WRF model published at the Meteogalicia server using the meteoForecast package. The set of predictors are defined with different scenarios that can be selected in the Predictors tab.

Once the day to be predicted, the PV system, the training method and the scenario have been chosen, define the length of the train time series (number of days of recent past measurements - method "previous" - or selected from database - methods "ks" and "kt") in the Result tab and click on the Forecast button.

Values are shown in % of nominal power.

The experimental data used in this tool belongs to 5 PV plants. Data recording started on April 17th, 2008, although the database for this tool is restricted to the period comprised between January 8th, 2009 and December 29th, 2010. The PV plants have been equipped with a monitoring system that records the power generated by each inverter every 5 seconds. However, the forecast procedure works with hourly averages of these measurements because the WRF-NWP provides hourly outputs. The figure shows the hourly averages of the AC power measurements.

See Table 3 of the paper PV Power Forecast Using a Nonparametric PV Model, Solar Energy 2015, M. Pinho Almeida, O. Perpiñán, L. Narvarte.
The figure displays the variables extracted with the meteoForecast package from the WRF-NWP model available at the Meteogalicia server, and several spatial indexes. These variables and indexes are the predictors used to forecast the AC power time series.

Indexes

point: value at the location of interest.

IDW: interpolated value (Inverse Distance Weighting) at the location of interest.

TRI: Terrain Ruggedness Index.

TPI: Topographic Position Index.

rough: Largest inter-cell difference of a central pixel and its surrounding cells.

timeDesv: Standard deviation of consecutive predictons for the same hour.

Values are shown in % of nominal power.

Legend

The red line represents the hourly AC power measured at the plant (when available).

The black line represents the prediction of the median (quantile 0.5).

The shaded area represents the prediction interval comprised between the quantiles 0.1 and 0.9.

Training methods

previous: This method selects days immediately before the day to be predicted.

kt: This method selects days with the lowest absolute difference between the clearness index of the day to be predicted and the clearness index of each day included in the database.

ks: This method selects days with the lowest Kolmogorov-Smirnov distance between the empirical distribution function of the irradiance forecast for the day to be predicted and the empirical distribution function of the irradiance forecast for each day included in the database.

  • Previous AC power measurements from a PV plant are collected.
  • Past and actual NWP-WRF forecasts for a set of variables (solar radiation, cloud cover, temperature, wind speed, etc.) published by Meteogalicia are downloaded.
  • Each NWP-WRF variable is processed to extract information about the value at the location of interest and its relation with the surrounding locations and previous forecasts. Several indicators are extracted:
    • Value at the location of interest (point).
    • Interpolated value (IDW - Inverse Distance Weighting) at the location of interest and its surrounding cells.
    • Spatial variability indexes (TRI, TPI and roughness, Wilson et al 2007, Marine Geodesy 30:3-35) of the cells surrounding the location of interest: TRI stands for Terrain Ruggedness Index, which is defined as the mean difference between a central pixel and its surrounding cells; TPI stands for Topographic Position Index, which is defined as the difference between a central pixel and the mean of its surrounding cells; Roughness is the largest inter-cell difference of a central pixel and its surrounding cells.
    • Temporal variability index (timeDesv), given by the standard deviation of several consecutive forecasts for the location of interest.
    • Hourly clearness index (kt).
  • The time series of processed forecasts and measured AC power are divided into two time series: train and test. The train time series comprises past values of both WRF variables and AC power, whereas the test time series contains only present WRF variables from the NWP model (forecasts)
  • A Quantile Regression with Random Forests is trained with the train time series. A prediction of the median (quantile 5) and a confidence interval (quantiles 1 and 9) are produced with the test time series.