POWER GENERATION FORECASTING ENGINE
A wind farm in Europe that is supplying power to the electric grid was having a high margin of error in its output prediction, that was resulting in them having to pay a high penalty to the grid. (They were operating on a 30% error rate. Errors of greater than 10% incurred a penalty that increases with the magnitude of the error.) They needed an improved weather and wind prediction model, so as to reduce the error in their power output prediction.
Recognizing the limitations of single global models for weather, our data scientists decided to look at multiple models in combination with hyper local weather sensors at the wind farm and combine this with other climate phenomena. Farm topology and efficiency were also incorporated to develop an ensemble of models that would deliver higher accuracy.
- Our data scientists used data from three different sources:
- Prediction output from global 3rd party weather models like GFS and BOLAM.
- Hyper local observations collected from sensors at the wind farm
- Data from global climatology phenomenon like El Niño (raw weather information at a global level)
- We used machine learning techniques on these data sets and made multiple base models.
- Multiple base models were combined to create a hybrid model to reduce variability of the individual base model to arrive at a final prediction of the wind speed and wind direction with greater accuracy.
- These were then used along with data on the topology of the wind farm and historical power output data from the farm to predict the power output from the farm.
- The machine learning continuously learns for past data and assigns appropriate weight to each model hence, improving the prediction accuracy each time.
TECHNIQUES, TECHNOLOGIES, TOOLS
ARIMA, Neural Networks, Time Series Forecasting, Deep Learning
The forecasting engine was able to predict power generation with greater accuracy. This brought down the error rate from 30% to 16%, resulting in a 70% reduction in the potential penalty.