Published using Google Docs
PV_Forecast_Validation_20170731
Updated automatically every 5 minutes

Validation Statistics for the PV_Forecast Service

Jamie Taylor

31st July 2017

Introduction

The PV_Forecast service uses statistical modelling techniques to forecast the generation of individual PV systems. The PV_Live National methodology is then applied to the forecast data in order to produce a forecast of the nationally-aggregated PV generation. Naturally, the error on the national forecast can be assessed by comparison with the PV_Live National results. For context, we also include a comparison with the National Grid’s own day-ahead forecast of embedded solar PV generation.

About the Data

Here we consider a full year of historical simulated PV forecasts (hicasts) using the Sheffield Solar PV_Forecast methodology in order to evaluate and validate the performance of PV_Forecast. The PV_Forecast generation hicasts use real historical weather forecasts from MeteoGroup. The weather forecasts were initiated at 7am each day (GMT) and we considered the forecast error up to a 24 hour horizon (a.k.a lead time) i.e. up to the interval ending 7am the following day.

For benchmarking purposes we have included a comparison with the NG “day-ahead” forecast of nationally aggregated embedded PV generation, which is freely available to download from the Elexon BM Reports platform[1]. This forecast is initiated by NG some time before the start of the day in question, and so the two forecasts are not technically equivalent since the NG day-ahead forecast of the midday generation has a longer lead time than the Sheffield Solar PV_Forecast. In future we will look to study how the forecast error in PV_Forecast changes with forecast lead time.

Validation Results

typical_profiles.png

Figure 1; Typical hourly PV generation profiles for the PV_Forecast service alongside PV_LIve and the NG day-ahead forecast.

forecast_vs_actual.png

Figure 2; Forecast generation against PV_Live generation for PV_Forecast and NG Day-Ahead. Red lines show the linear fit and green lines show the line y=x.

RMSE_vs_time_of_day.png

Figure 3; Root Mean Square Error (RMSE) of PV_Forecast and NG Day-Ahead forecast by time of day.

mape_by_time_of_day_summer.png

Figure 4; Mean Absolute Percent Error (MAPE) by time for PV_Forecast and NG Day-Ahead during the months May to August (inclusive).


[1] https://www.bmreports.com/bmrs/?q=foregeneration/dayaheadwindnsolar