Renewable MOS
Description
Renewable MOS (Model output statistics) is a localised power forecast based on the radiation or wind forecast of multiple weather models and historical power measurements of wind turbines or solar power plants. Thereby, the model output is corrected statistically by the function developed from the measurement data (Model Output Statistics), using the weather forecasts from multiple models as a basis. The forecast is adjusted to the local station, using quality-controlled measurements from the site which needs to be available over at least 1 year and in hourly intervals.
General Process
Power forecasts can be optimized by using statistical post-processing algorithms. The improvement of the forecast is depending on several site-specific conditions. On average, an improvement of 30% (e.g. reduce MAE from 35% to 25%) is realistic. To assure the functionality of the post-processing algorithms, a quality-control procedure for the measurement data is crucial.
For each MOS training meteoblue applies the following procedure:
Delivery of measurement data in standard format
Measurement quality control
Forecast API configuration
Delivery of MOS setup report
Methodology
meteoblue algorithms use a conditional MOS, which combines a simple neural network with improved meteoblue MOS technology. The algorithms differentiate between multiple weather conditions to find the best fit of linear regression for more than 30 weather variables from multiple weather models. The resulting forecast algorithm is individually designed for the specific site and provides the best possible forecast (state of the art) via a special API set for the specific application and location.
MOS Forecast Variables
The possible forecast variables and range/intervals are:
· Solar Energy_kWh; (Availability variable is necessary to calculate kWh/kWp)
· Wind Energy_kWh; (Availability variable is necessary to calculate kWh/kWp)
· Wind Speed
Other variables can be supplied on request.
Configuration Process
The MOS power forecasting model is determined based on the forecast of radiation or wind speed. The forecast is adjusted to the local power plant by a statistical function, derived from past measurements, as follows:
Customer provides location information: position, altitude
Customer provides power plant information: Solar: Capacity (kWp), exposition (slope, facing), plant type Wind: Capacity (kWp), turbine type, turbine height, number of turbines
Customer provides measurement time series in standardised format and UTC time.
meteoblue conducts data quality control, to ensure a good return on investment.
meteoblue extracts archived weather data and finds best statistical correlation.
meteoblue implements correction factors into the API forecast system.
meteoblue provides corrected MOS power forecast with 24 updates per day.
Observation Data Format
It is recommended to provide at least one year of data in hourly time resolution to train the renewable MOS. The data should be provided in a standard format using comma separated files (.csv). Especially designed routines will be applied to detect completeness, time-shift and plausibility of the time series. Apart from production time series, minimal metadata (header information) have to be supplied. The observation data for MOS configuration should support following requirements:
standard text format: .csv oder .txt
one file for each site, that contains the full available time range
FILENAME: Site-ID_LAT_LON_timerange_variable_generationdate.csv
e.g.: FOR-224_42.75_12.86_20190523-20201218_PV_20211218.csv
The file contains different columns for SITE_ID, date and time (in UTC) and the target variables. Energy production (Energy_kWh) is always delivered with system availability (=available capacity (kWp) at a given time stamp).
Sample:
ID;date;UTC;Energy_kWh;kWp FOR-224;23.05.2020;12:00:00;0.12;87.98 FOR-224;23.05.2020;13:00:00;0.63;87.98 FOR-224;23.05.2020;14:00:00;0.04;87.98 FOR-224;23.05.2020;15:00:00;4.92;87.98 FOR-224;23.05.2020;16:00:00;11.63;87.98
Sample Files:
- Sample File Station List
- Sample File Time Series
Accuracy Improvements
The expected precision of power forecast with the MOS forecast is between 5% and 25% rMAE on an hourly basis for the 2-24 hour forecast, depending on location and year. The value of statistical post processing differs depending on location and input data (measurements). The highest value of these methods can be observed for the exact configuration of PV yield forecast. Differences in site results are shown in figure below.
Hourly mean absolute errors (rMAE) of meteoblue day-ahead (multimodel & PV-Mos) and intraday (3h-nowcast) PV yield forecast validated on 38 PV systems in Central Europe.
Solar Monitoring and Nowcasting
Description
Solar Nowcasting is the correction of radiation forecast based on real time satellite observations collected every 15 minutes. Thereby, the model radiation output of the present day is corrected in real time with the satellite derived radiation. Based on the actual cloud pattern and wind vectors a so-called cloud motion forecast is processed which optimises the forecast accuracy of the next 5 hours. Depending on the actual weather situation persistence, cloud motion and the model output are weighted differently for the time horizons of the next hours.
The solar forecast retrieved before sunrise (blue), in the morning (red) at noon (green) and after sunset (violet) of the same day
The optimised radiation nowcast will reduce short term forecast errors of all radiation and solar power variables and allows satellite monitoring of solar energy yields, because the actual satellite observation is applied on past time stamps (up to 4 days back). Thus, the solar forecast API contains only observation values after sunset and is therefore combining forecast and monitoring within the same data feed. The effects of nowcasting are depending on the accuracy of the weather model. If the day ahead forecast is correct, nowcasting has almost no impact (see Figure above), otherwise the nowcast is strongly changing the content of the API (see Figure below).
The solar forecast retrieved before sunrise (blue), in the morning (red) at noon (green) and after sunset (violet) of the same day
Nowcasting Variables
Presentlyy, the nowcasting technology is automatically implemented in the API for all locations withing sight of the Meteosat Second Generation (MSG) satellite. Thus, it covers Europe, the Arabian Peninsula, Africa, Brazil and parts of its neighbouring countries (see figure below).
The full MSG disc of Global Horizontal Radiation (GHI) at 2016-03-16 15:00 UTC (Source: KNMI)
Nowcasting Variables
The nowcasting is implemented on all radiation variables and solar power simulations.
Cloud Motion Vectors (CMV)
Based on the image of the cloud-top level, the wind vectors of the next 6 hours are extracted for the specific cloud height of each point. A cloud motion vector (wind trajectory) is calculated for each time horizon of 1-6 hours.
Solar Radiation Nowcast Based on Cloud Motion Vectors (CMV)
Based on the six wind trajectories, the referring pixel layer is chosen for each time horizon (1-6 hours). This pixel is used to calculate the expected radiation. The result is a radiation forecast for the next 6 hours. Cloud motion nowcasting works only after sunrise when the referring satellite data is available.
Persistence Solar Radiation Nowcast
For the persistence solar radiation nowcast, the radiation of the last 4 hours is used (without any cloud motion) and weighted.
Combined Solar Radiation Nowcast
For different time horizons and weather situations, the weighing of persistence, cloud motion and numerical weather prediction is changing. It is useful to combine the different approaches to minimize errors. Weather variability could be classified with the clearness index of the surrounding satellite pixels.
Expected Precision
The expected precision of radiation or power forecast with nowcasting is between 5 and 20% rMAE on an hourly basis for 0-5 hour forecast, depending on location and satellite availability. The accuracy of the nowcast is declining over the forecast horizon (see Figure 7.5). After 5 hours, the multi-model is performing better than the nowcast.
Nowcasting (red: 1h, blue:2h, orange:3h) improves performance of solar forecast service (black line). Vertical axis shows MAE (in %) of hourly forecast over a full year. Stations sorted in order of ascending error.