Even though it’s a potent source of renewable energy, wind is unpredictable. In any grid with sufficient wind capacity, grid managers face challenges in balancing demand and supply. Therefore, accurate tools and processes for predicting wind speed (and power) are needed to balance the grid. To facilitate this, grid managers require power producers to submit forecasts of generation. The grid managers balance demand for power with the forecast to decide on power needed from other sources like gas, coal which can be scheduled based on demand. Forecasts for wind speed and farm power output are to be given for the next day. Forecasting is therefore essential for wind farm management and growth of wind energy mix in the total supply.
In this article, we have analyzed different methodologies and noted the conclusions about their efficacy in providing an accurate forecast. To begin with, let us define some metrics based on which different forecasting methodologies are compared. These metrics are calculated based on the difference between actual value and predicted value which we will henceforth refer to as the error value. In both cases, the higher the value of the metric, the lesser correlation (or accuracy) there is between the predicted and actual values.
A few simple approaches to wind power forecasting were evaluated
This method is considered the simplest approach to wind forecasting. In this method, the predicted wind speed value for the next slot will be the actual wind speed value in the last available slot. Slot size may vary between 10 min to an hour or even greater. As a simplistic level we may choose to use the previous day average wind speed as the forecast for the next day. This will serve as a benchmark to compare improvements offered by other methods of forecasting.
Diurnal Averaging method:
In this method, wind speed for the farm was calculated in each 10minute/1 hour slot of the day. Corresponding slots of previous 7 days can be averaged and that value will be considered as the forecasted value for the next day (8th). Within this scheme, we also tested with different configurations like using 6 day averages or 3 day averages or even 15 minute slot values instead of 10 minutes slot values in order to see if we are getting a better correlation. The tabulation shows the best result we achieved.
Forecasting based on Weather agency data:
Many different institutions (at a national and international level) provide 3 hour or 6 hour wind forecast data for specific region. We utilized the publicly available data from the servers of US based National Oceanic and Atmospheric Administration (NOAA) – which is a federal agency responsible for the countries information services regarding weather and climate data. This data is freely available at the link: http://nomads.ncep.noaa.gov/ . This data is generally in a GRIB file – most commonly used file format for storing environmental & climatic data. In order to test with various data sets we took wind speed forecast data at 10m and 80m above sea level for the specific latitude corresponding with one of our 50MW farm in India. We also evaluated data from Access G- which is a paid data service from the Australian government http://www.bom.gov.au/nwp/doc/access/NWPData.shtml & the National center for medium range forecasting under the Indian government – http://www.ncmrwf.gov.in/
Forecasts were derived from all of the above methods and then compared with corresponding actual values generated at sites studied by us. Summarized in the table below are the calculated metric values for each of the methods described above:
|Forecasting Method||RMSE value||MAPE value|
|Forecast Agency data||2.8||35% to 55%|
|Persistence Method||0.85||28% to 60%|
|Diurnal Averaging Method||0.85||25% to 40%|
There are many conclusions we can draw from this analysis. Firstly, weather forecast data has a large variability from actual measured data and it could only be utilized for understanding long interval trends but has limited usability for short term forecasts. The averaging techniques used also do not yield a high correlation forecast and thus will not be effective for this purpose. There is a definite need to apply machine learning techniques which will utilize different advantages of historic data and forecast agency data to generate even more accurate forecasts.
At Algo Engines, we have explored some machine learning models for accurate wind speed forecasting. Stay tuned to the blog to find out more about converting wind speed forecasts to energy forecasts based on power curve and details of machine learning models for wind forecasting.