Intelligent forecasting in Wind Energy Applications

Various relatively simple statistical models such as linear regression [1] and neural networks [2] have been used to forecast wind speeds, power output, and are used in prognostics e.g. tracking condition characteristics in wind turbines. There is an urgent need to develop systems which not only perform good forecasts, but compare forecast accuracies across a range of models, and dynamically select the best performing algorithm around an operating point [e.g. best performance in the last 4 hours].
This ‘competition’ between models [3] has been highly successful in other applications (economics, asset management, environmental science) but is relatively new to the wind sector. Researchers at Strathclyde have made highly successful strides in inter-model competition and corroboration for condition assessment of key infrastructure such as transformers [4]. This framework will be developed and tuned to the specific needs of the wind power forecasting problem, driving down operational cost in several key ways:
• More accurate wind power forecasting reduces the need for holding reserve generation to mitigate wind power fluctutation, thus reducing operational costs and overall cost to the consumer
• More accurate wind power forecasting enables better foresight of grid congestion on the network and facilitates more efficient dispatch of generation, reducing constraint payments and reducing overall cost to the consumer
• Benchmarking and corroboration across forecasting algorithms enables an improved understanding of algorithm performance under specific operational conditions, enabling model performance tuning, increased forecast accuracy which feeds through to the points above.