Short-term Power Load Forecasting Based On Structural Improved Recurrent Neural Network

With the advent of the era of big data in power,how to mine the potential information behind massive data is of vital significance.Electric energy is unstorable,and production and consumption are carried out simultaneously,so research on electric load forecasting is necessary.Especially short-term power load forecasting,which is the basis of grid planning and decision-making.At the same time,accurate short-term power load forecast is directly related to various performance indicators of grid operation.In recent years,machine learning algorithms have been widely used for short-term power load forecasting.Long Short Term Memory(LSTM)and Gated Recurrent Unit(GRU)in Recurrent Neural Network(RNN)are tailored for time series data.Their internal gating structure can solve the problem of gradient disappearance in traditional RNN,and at the same time adjust the information flow to achieve long-term memory.Based on this,this paper also studies the internal rules of load data and related influencing factors to achieve the purpose of improving the accuracy of short-term power load forecasting.The main work of this paper is as follows:1.A multi-layer bidirectional RNN model based on LSTM and GRU is proposed.The first layer of the model is composed of forward propagation LSTM and back propagation GRU units,while the second layer is opposite.Adding Rectified Linear Unit(Re LU)before and after the multi-layer bidirectional RNN network introduces non-linear factors to enhance the model expression ability.The bidirectional structure can be used to contact information about the time before and after the data,so that the network can be fully trained.Considering the seasonal differences,experiments were performed separately on the data set,and the results show that the proposed method is beneficial to improve the forecasting accuracy.2.A comprehensive data processing method combining binary k-means algorithm and Ensemble Empirical Mode Decomposition(EEMD)is proposed.According to the characteristics of daily load variation,the load data with high similarity to the predicted dates are aggregated into one category,and then the original load sequence is processed into multiple time series components by EEMD.Calculate the sample entropy of each component and merge the components with similar sample entropy values.Finally,thecorrelation between the merged sub-components and the candidate influence factors is analyzed to construct an effective feature input,thereby ensuring the validity of the data.3.A forecasting model based on the combination of multi-layer bidirectional RNN and Deep Belief Network(DBN)is proposed.LSTM and GRU were used to construct a two-layer bidirectional forecasting network,and the weight parameters of the forecasting network were obtained by DBN unsupervised greedy pre-training.Unsupervised pre-training weights is to overcome the problem that neural networks tend to fall into local optimization due to random initialization of weights.Combined with the data processing method,the verification is performed on two public and non-public data sets.The results show that the hybrid neural network model enhances the effectiveness of the forecasting model.