Research On Ultra-Short-Term Wind Power Forecasting Combination Model Based On Deep Learning

With the continuous development of human society,the increasing consumption of fossil energy is accompanied by increasing energy shortages and ecological environmental pollution.It has become the consensus of all countries in the world to adjust the energy structure to reduce the dependence on fossil energy and develop renewable and cl ean energy to promote the sustainable development of society.As a kind of renewable clean energy with wide distribution and abundant resources,wind energy has a broad application prospect.However,wind power is characterized by fluctuation,intermittency and uncertainty,and the grid connection of a large number of wind turbines will bring serious challenges to the operational stability,real-time scheduling and power quality of the power system.Accurate and reliable wind power forecasting allows the power system to make reasonable scheduling plans and adjust generation tasks in advance,making full use of wind power while minimizing the security problems of the grid system caused by wind power.The current ultra-short-term forecasting of wind power has the problems of the difficulty of considering the prediction accuracy,timeliness,and applicability simultaneously.Therefore,based on a large number of wind power data combined with deep learning algorithm,this paper conducts research on ultra-short-term wind power forecasting with strong timeliness,and the main research contents are as follows:(1)For ultra-short-term prediction of wind farm power in smaller scale.The CNN-Bi GRU ultra-short-term wind power prediction model based on Attention mechanism is proposed,which takes a single wind turbine as the prediction unit,uses CNN to compress the hidden states in the Bi GRU network,and extract the spatio-temporal correlation between wind power data,which shortens the computation time and also alleviates t he problem of gradient disappearance and gradient explosion.In addition,the Bi GRU network is used to perform bidirectional learning on the relationship between wind power characteristics,and the Attention mechanism is used to assign different weights to the hidden states of the Bi GRU through learning parameter matrices and mapping weighting.The experiments show that the model has good performance in prediction performance and efficiency,and the root-mean-square error and average absolute error are controlled within 3MW,which effectively improves the accuracy of ultra-short term prediction of wind power.(2)For ultra-short-term prediction of wind power in large-scale wind power clusters.The QPSO-TCN-Bil STM wind power ultra-short-term forecasting model based on feature weight analysis and cluster dynamic partitioning is proposed.The model takes the dynamically divided class regions as prediction units,introduces analytic hierarchy process and entropy weight method to analyze the subjective and objective weight of each feature,determines the subjective and objective weight ratio by quantum particle swarm optimization algorithm,and calculates the comprehensive weight of each feature.The wind turbines with the same characteristics are classified into one class by dynamic clustering according to the period.Then establish QPSO-TCN-Bil STM models for various regions to predict their power values;Finally,the power prediction value of each region is accumulated to obtain the power prediction value of the wind power cluster.The experimental results show that the model has a good feature extraction ability,the consumption time is about three times of the whole wind farm prediction,is one twentieth of the single wind turbine prediction,can guarantee the prediction accuracy and shorten the prediction time of the model.(3)For ultra-short-term and high precision prediction of wind power clusters with complex environmental characteristics.According to the operational characteristics of wind turbines,a wind power ultra-short-term prediction model based on trend and randomness.The model first analyzes the importance of each impact feature through a random forest(RF)algorithm to reduce the error caused by the varying degrees of impact of each feature on wind power.Secondly,according to the characteristics and corresponding weights,the class area is divided,and the data is divided into historical power data and meteorological feature data based on the class region.Then the trend power prediction model is established based on historical power data,and the random power prediction model is established based on meteorological characteristic data,and the prediction results are fused using functions.Finally,the ultra-short-term power prediction value of the wind power cluster is obtained by accumulating and summing the prediction values of various regions.The root mean square error and mean absolute error values of the experimental results are lower than 2.5% of the rated power of the wind farm,and the stability index value is lower than 2.0% of the rated power of the wind farm.It shows that the model effectively combines trend power and random power,and can better capture the trend and random change characteristics of wind power.On the basis of ensuring the timeliness of prediction,it improves the prediction accuracy of wind power.This paper conducts research on ultra-short-term wind power prediction based on deep learning combination models,and the proposed model algorithm achieves certain improvement in prediction accuracy,timeliness and adaptability,which provides new ideas and methods for the research of ultra-short-term prediction of wind power.