Research On Wind Power System Power Prediction Based On Deep Feature Mining

The demand for wind energy,as a major source of renewable energy,has being growing significantly,which in turn has driven the structural transformation of the global power system from fossil fuels to renewable energy.Due to its cost competitiveness and its advantages in reducing greenhouse gas emissions,wind energy has become a key supply component of contemporary power systems.But the intermittent,nonlinear,and wildly fluctuating characteristic of wind energy presents a variety of challenges for forecasting,planning,and efficiently integrating its supply into regional power systems to meet changes in electricity demand.At the same time,wind resource assessment is also an important prerequisite for the effective construction of wind turbines and wind farms.An accurate power prediction model can maximize the energy extraction capacity of the power supply system and reduce the cost of electricity production.In this paper,the features and laws of output power data of single-site,multi-site and system equipment in the wind power system are studied in depth,and two single-site wind power forecasting methods combined with deep learning and attention mechanism are developed for the high-frequency signal decomposition and periodic feature extraction of wind power signals.Combined with the multi-site spatiotemporal feature correlation of wind farms,the spatiotemporal prediction method for multi-site wind power is studied.Furthermore,the adaptive regression prediction model of wind power system equipment is also proposed.The main research results and work are presented as follows:To solve the problem that it is difficult for a single network model to efficiently capture the variation trends of high-frequency and low-frequency signals,the law of highfrequency and low-frequency signals of wind power is studied,and an integrated model,based on signal decomposition and high-low frequency division is proposed.Firstly,based on the variational modal decomposition method,the original multi-dimensional feature data is modally decomposed,and divided into high-frequency and low-frequency components,according to the average fluctuation amplitude.Then the encoding-decoding attention mechanism network and deep confidence network are used for prediction respectively.Finally,the above prediction results are integrated to fully extract the multidimensional time-domain and frequency-domain characteristics of the original signal,thereby the performance of the prediction model are improved.In terms of the characteristics of hidden periodicity and large amount of wind power time series data,the periodic law of short-term and long-term repetitive patterns of wind power is studied,and a dual-channel integrated prediction model based on periodic feature extraction and multi-layer attention network is proposed.Combined with the seasonal trend decomposition and the optimal periodic step search algorithm,the hidden periodic characteristics are extracted,and the original signal are decomposed into temporal and periodic components.Then the temporal attention network and the encoderdecoder attention mechanism network are used for prediction respectively.Finally,the linear regression attention model is used to fit the prediction results of the two components.The proposed model can deal with the long-term and short-term dependencies of the original data at the same time,and ensure the rapid convergence of the model training,thereby improve the prediction accuracy and stability.In order to combine the temporal characteristics and spatial characteristics of multisite wind farm,the spatio-temporal correlation law of multi-site are further studied,and two spatio-temporal combined prediction models are designed.Firstly,combined with the time series correlation of multi-dimensional features,the convolutional neural network is used to expand the data perception field to obtain more effective feature extraction capabilities.Combined with long and short-term memory networks,a multi-site time series convolution prediction model is designed.Secondly,the adjacency matrix is constructed by using the spatial features and time-series correlations of multi-sites in the wind farm,and combined with the time-series feature matrix of each site,an integrated model based on graph convolution and gated recursive units is designed.Furthermore,a multi-head graph attention mechanism is designed to learn the weights of neighbor nodes adaptively,so as to better achieve feature aggregation between adjacent sites.Based on this,a multi-site spatiotemporal graph attention convolution prediction model is designed.In the experiments,the above four proposed integrated models and basic models are analyzed and verified through two multi-site wind power data sets of wind farms in a progressive manner.The efficiency of proposed model algorim are proved through predicted error indicators,stability,convergence speed and algorithm efficiency.In order to deal with the complex data features,different feature types and large data sets,the homogeneity law of wind power system equipment is analyzed,and a power adaptive integrated regression prediction model is proposed.The Gaussian mixture clustering algorithm is used to aggregate the same type equipment with similar data characteristics to form several equipment subsets.Then,an adaptive loss function is designed to determine the optimal control parameters of the regression decision model for different equipment subsets,thereby the adaptive regression prediction model is designed.The experimental comparative analysis of wind power equipment data sets shows that the proposed regression model can adjust parameters adaptively according to the characteristics of different data sets,and has the best performance in terms of prediction accuracy,stability,and convergence speed.