Research On Short-Term Wind Power Prediction Based On Multi-Step Cross-Decomposition Of Wind Power And Wind Speeds

With the dramatically increasing of wind power in power system,the inherent intermittence and volatility of wind power bring great challenges to the safe operation of power grid.Improving the prediction accuracy of wind power can help the dispatcher to adjust the dispatching plan in time,so as to reduce the impact of wind power on the grid and improve the capacity of wind power consumption.This paper studies the short-term point prediction technology and interval prediction technology of wind farm power and clustered wind power from the two aspects of feature decomposition and ensemble learning.The main contents and contributions of this paper are summarized as follows:(1)The deterministic decomposition method of multiple predicted wind speeds in wind farm is studied.When using wavelet transform to decompose multiple predicted wind speeds at different heights of wind farm,it is difficult to determine the decomposition level and wavelet function.To solve this problem,a multi-step cross-decomposition method of wind power and wind speeds based on maximum matching rate is proposed,which can directly determine the decomposition level and wavelet function of each wind speed according to the maximum matching rate.Through multi-step cross-decomposition,multiple wind speeds at different heights can be decomposed into components of different frequency bands.After introducing wind direction into each wind speed component for triangular decomposition,the input feature set of wind farm prediction model can be obtained.Through the comparative analysis with200 benchmark data sets on 20 wind farms,the effectiveness of the multi-step cross-decomposition method are verified.(2)The integrated prediction method of wind farm power is studied.When using ensemble learning method to predict wind farm power,there is still a lack of deterministic method to construct differentiated input features and differentiated targets for sub-models.To solve this problem,based on the deterministic decomposition of both the wind power and wind speeds through cross-decomposition,a cross-decomposition-ensemble-learning method is proposed to predict the power of wind farm.Firstly,the differentiated targets for the sub-models are constructed by component-wise accumulation of the power components.Secondly,the input feature set is encoded into different dimensions by stacked denoising auto encoder to construct differentiated input features for sub-models.Finally,support vector machine is used to select and integrate the sub-models.The experimental results show that the prediction accuracy of cross-decomposition-ensemblelearning method is significantly improved compared with single prediction model,traditional ensemble learning model and component-wise integration model.Compared with empirical-decomposition-ensemble-learning method,cross-decomposition-ensemblelearning method can achieve relatively high prediction accuracy with relatively low input dimensions and relatively few sub-models.(3)The ensemble learning method for cluster power are studied.When using ensemble learning method to predict clustered wind power,there is still a lack of deterministic cluster feature decomposition method and the corresponding sub-model design method.For large-scale regional cluster,there is still a lack of a deterministic hierarchical division method.To solve the above problems,a hierarchical division method for regional cluster based on reference vector is proposed,which can not only hierarchically divide the large-scale regional cluster into relatively small sub-clusters with similar scale,but also make the sub-clusters meet the requirements of predictability.After that,a deterministic two-stage cross-decomposition method is proposed to decompose the features of the sub-cluster,and the corresponding sub-model design method is given.Finally,a hierarchical integration method is proposed to predict the power of clusters of different scales.Through the comparative analysis with the benchmark methods,the effectiveness of the hierarchical integration method in cluster of any scale is verified,and the utility of the hierarchical division method in improving the accuracy of cluster power prediction is proved.(4)The clustered wind power interval prediction method based on point prediction is studied.When the interval labels is constructed with the observed value as the base value and the memory model is used to establish the mapping from the predicted value to the interval labels,certain high-frequency bands of the observed value and the predicted value may have a negative impact on the interval prediction accuracy.To solve this problem,a method based on variational mode decomposition and estimated interval is proposed to filter the predicted sequence and observed sequence of power so as to select a lowfrequency band of observed value with good predictability as the base value to construct the interval labels,as well as the corresponding low-frequency band of predicted value to be the input of interval prediction model.In order to improve the prediction accuracy,the deep minimum gated memory network is used as the interval prediction model,and an adaptive interval width adjustment algorithm with the minimum estimated interval as the initial interval width is proposed.The comparison results with the benchmark methods verify the effectiveness of the feature selection method and deep memory model proposed in this paper,as well as the rationality of using the minimum estimated interval as the initial value of interval width.