Research On Wind Power Short-term Prediction Based On Improved RNN

Wind energy resources are volatile and uncertain,so large-scale integration of wind turbines will have a great impact on power system scheduling,but also lead to the generation of wind farm abandonment and power limitation.Therefore,how to effectively complete the work of wind power prediction is very important for wind power Grid-connected and effective utilization of wind power.In order to improve the short-term prediction accuracy of wind power,a combined quadratic decomposition model based on Wavelet Packet Decomposition(WPD) and Complete Ensemble Empirical Mode Decomposition Adaptive Noise(CEEMDAN) is proposed to preprocess historical wind power data.Firstly,the model decomposes the historical wind power data into WPD by three-layer wavelet packet decomposition,and obtains a series of high-frequency and low-frequency component subsequences.Secondly,all high-frequency component subsequences are decomposed into CEEMDAN by self-adaptive complete empirical mode decomposition,and a series of intrinsic mode functions(IMF)are obtained.Finally,all subsequences are normalized as training data and test data of the subsequent prediction model.Then,a short-term wind power prediction model based on improved Gated Recurrent Unit(GRU) neural network is proposed.GRU neural network is a new type of Recurrent Neural Network(RNN),which can overcome the problem that traditional RNN can not deal with long-distance dependence.Firstly,the scaled exponential linear units(SELU) activation function is used to optimize the GRU neural network.Then,the Adaptive Computation Time(ACT) algorithm is used to optimize the improved GRU neural network,and a self-regulating multi-layer wind power prediction model is established.Finally,the TensorFlow distributed machine learning framework is used to train and test the model.Experiments show that the proposed model has better prediction accuracy than the benchmark model.