Research On Short-Term Wind Power Forecasting Method

With the nervous aggravation of energy shortage and environmental pollution, wind power as one of the most promising clean energy has developed rapidly in recent decades. However, large-scale wind power threatens the security, stability, economic and reliable operation of the power system seriously. So, in order to provide effective guidance to the operators and reduce the adverse effects of wind power to grid, model of wind power forecasting with high accuracy is needed urgently. Generally, numerical weather prediction(NWP) or historical operation data are always used as inputs to predict wind power. But, there are some disadvantages of NWP data and serious mistake would be caused when only using historical statistics to forecast short-term wind power. Because of these, single-point and probability prediction models of short-term wind power only using historical data were studied in this paper.A new method based on non-iterative algorithm was proposed. According to this approach, no matter how long the ahead time was, neural networks for each forecast period were established all using real data as inputs. This method can reduce the accumulated error successfully and predict wind power of later 24 hours effectively, when only statistics are used.Next, a single-point prediction model for short-term wind power was put forward based on non-iterative method and divided period optimization. Through this model, the best number of input for each period was found out respectively and optimal weights for two methods based on historical speed data and power data were calculated completely. Then, this model was applied to analyze an example wind farm. Experimental results show that the method of divided period optimization presents characteristics of different times and greatly improves prediction accuracy of every period. And the model based on non-iterative method and divided period optimization has higher accuracy obviously when it is used to predict short-term wind power.Because single-point forecasting methods can’t provide probabilistic information, a probability prediction model for short-term wind power based on non-iterative method and RBF-quantile regression was established. According to this model, historical power data was decomposed into a series of sequences by wavelet transform. Then, probability forecasting for each component was calculated through RBF-quantile regression. And finally, distribution forecasting of wind power was obtained using Latin Hypercube Sampling method. The results show that this model has smaller error and better probabilistic evaluation indexes. So, it can be well applied in single point and probability prediction. The prediction results has high reference value.