Ultra-short-term Wind Power Forecast Considering Correlation Characteristics Of Multi-source Information

Wind energy is an important renewable energy.Affected by the randomness,volatility,and intermittence of wind energy,wind power presents a high degree of uncertainty,which makes large-scale wind power grid connection brings severe challenges to power system security and power quality.Therefore,accurate power prediction is one of the bases to ensure the stable operation of power system.However,a large number of ultra-short-term wind power forecasting studies focus on time series extrapolation modeling,and the accuracy often decreases sharply with the increase of time horizons.This paper studies the impact of multi-source information on ultra-short-term wind power forecasting performance and its modeling methods.Firstly,the correlation analysis of time series information and weather information is carried out through three strategies of filtering,wrapper,and embedding method.Take a large wind farm in Northeast China as an example,the correlation relationship of multi-source information is quantitatively and qualitatively analyzed as the theoretical basis of multi-source information modeling.Secondly,facing the trend of complication and black box property of wind power forecasting modeling,the interpretability of machine learning is taken as the starting point.The local interpretable model-agnostic explanations algorithm is introduced.And combining the high dimensionality of input information and the fuzziness of modeling mapping of wind power prediction,a new index is proposed to measure the input-output function relationship of local samples.The proposed index is tested and a risky model in practical application is defined.Thirdly,through the transparency of the training results of the black box model,the oriented improvement strategy of ultra-short-term wind power forecasting is explored.For the input information of wind power modeling,based on the principle that the data-driven results should be consistent with physical mechanism,a new feature selection method is proposed on the basis of three feature selection strategies: filtering,wrapper,and embedding method.The prediction error can be further reduced by 0.36%-4.69% on the basis of 17.23%-28.48%.For the four kinds of modeling strategies for ultra-short-term wind power forecasting,they are recombined from the view of the model training,and a model method of time-varying switching output mechanism is proposed.The combined framework constructed by the switching modeling method can robustly improve the prediction accuracy in the gated recurrent unit,radial basis function,and error back propagation neural network.Finally,in order to further quantify the uncertainty information of ultra-short-term wind power forecasting,the probabilistic prediction modeling strategy is explored.Based on the point prediction results,combined with the inertia operation characteristics and kinetic equation of wind turbine,a multi-scene division method for error is proposed,and the method is verified by both parameterized and non-parametric modeling methods at multi-time scale and confidence level,respectively.The proposed method is verified better than the traditional method which is based on the predicted value partition or not.In addition,for the ultra-short-term wind power probability prediction,there is no significant difference of the interval coverage for the same method under different time horizons,but the interval bandwidth decreases with the decrease of time horizon.It is the same as the accuracy change of point prediction under different horizons,which also shows that the high accuracy of point prediction accuracy can effectively improve the performance of probability prediction.