Research On Power System Short-term Load Forecasting Based On Improved Ridgelet Neural Network

With the rapid development of China's power industry and the promotion of smart grid construction,electric energy has become an indispensable source of energy for people's daily lives.The access of various distributed power sources and the grid-connected operation of the micro-grid make the complexity of the power system rise sharply,it also increases the randomness and nonlinearity of the system load,which poses a severe challenge to the accuracy and stability of power system load forecasting.Short-term load forecasting of power systems is a cumbersome and complex subject,and there are many kinds of influencing factors,and forecasting methods are also emerging.This paper analyzes the factors influencing the prediction results by studying the excellent research results of scholars at home and abroad,and elaborates the principle of neural network.Aiming at the problem of low prediction accuracy of traditional neural networks,this paper applies the ridgelet neural network to the field of short-term load forecasting,and has achieved good prediction results.However,for the load with large fluctuation randomness,the conventional ridgelet neural network has great limitations and the prediction stability is poor.So this paper improves the conventional ridgelet neural network and proposes a short-term load forecasting model based on ridgelet recurrent neural network.The inheritance layer is introduced in the conventional ridgelet neural network model to store the state information of the current moments in the hidden layer neurons of the network,and then passed to the hidden layer at the next moment,which enhances the feedback connection of the network model.The ridgelet transform function is used as the excitation function inside the hidden layer neurons,which enhances the network model's ability to optimize the nonlinear load.Finally,the particle swarm optimization algorithm is used to optimize the parameters and connection weights of the network,which enhances the ability of the network model to quickly optimize.The simulation of the example shows that the model has better prediction accuracy.In this paper,the anisotropy of ridgelet transform function and the powerful learning ability of deep neural network are used to improve the architecture of network model.And the deep ridgelet neural network prediction model with multiple layers of hidden layers is proposed.The unsupervised layer-by-layer pre-training of the network model is carried out by using the restricted Boltzmann machine learning principle,and then the particle swarm optimization algorithm is used to optimize the network parameters,the connection weights and the thresholds,which strengthens the training speed and optimization ability of the network.Through the simulation test of the example,the deep ridgelet neural network prediction model has higher prediction accuracy and prediction stability than the traditional BP neural network,ridgelet neural network and deep neural network.Based on the above research,this paper further improves the architecture of the network model and constructs a deep ridgelet recurrent neural network prediction model.Introducing the receiving layer in the ridgelet recurrent neural network into a restricted Boltzmann machine,that is to say,on the basis of the Boltzmann machine structure including the visible layer and the hidden layer,the inheritance layer is added to store and feed back the hidden layer output at the previous moment,which enhancing the dynamic characteristics of the network.And then use the limited Boltzmann machine learning principle and particle swarm optimization algorithm to double optimize the network.The actual data simulation proves that the model has stronger prediction performance than the ridgelet recurrent neural network and the deep ridgelet neural network.