Short-term Power Load Forecasting Based On Deep Neural Network

Short-term power load forecasting is an essential part of the power system management and operation,can provide reference basis for power system scheduling and operation,optimize the load distribution and power generation plan,help the power market to develop reasonable power price and supply and demand strategy,promote the fair,fair and efficient operation of the power market,for the government and enterprise forecast energy planning,by adjusting the load demand to meet the requirements of energy conservation and emissions reduction,can also be found in the power system load mutation such as abnormal situation,ensure the safe and stable operation of the power system,etc.Therefore,it is necessary to make a good job of short-term power load forecast.In order to do short-term power load prediction,improve the prediction accuracy,this paper using the neural network in processing time series data,using Gated Recurrent Unit(GRU),Convolutional Neural Network(CNN),Attention,preliminary established GRU-CNN deep neural network,and through the experiment discusses the influence of Attention in different positions in the network,this paper proposes a model called GRU-CNN-Attention.The method involves using the isolation forest algorithm to screen and correct bad load data,extracting load data through GRU in time sequence,using CNN to extract features based on the daily load rule,and introducing Attention to further improve prediction accuracy.The performance of the models(GRU,GRU-CNN,and GRU-CNN-Attention)were evaluated using the mean absolute error,root mean square error and mean absolute percentage error.The results showed that the GRU-CNN-Attention model outperformed the other models with the lowest error.To address the problem that CNN predisposes to gradient disappearance and explosion and the weak ability to model long sequences and too many hyperparameters,this paper proposing the use of Temporal Convolutional Network(TCN)with expanded causal and dilated convolution to effectively handle long time series.The paper further suggests the use of residual connections to transfer load sequence features.A new model called GRU-TCN-Attention is proposed,and several experiments are conducted to determine the appropriate expansion convolution coefficient and residual connections in TCN.The paper also compares the proposed model with historical load data and the existing GRU-CNN-Attention model in terms of prediction accuracy.The results demonstrate that the GRU-TCN-Attention model outperforms the other models in terms of prediction accuracy.To address the nonlinearity and randomness of load sequences,this paper proposes the use of Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)to transform the load data into a series of periodic components by adding Gaussian white noise.This allows for easier feature extraction by the deep neural network model.The paper compares the load data analyzed with and without CEEMDAN,and the experimental results demonstrate that the CEEMDAN model yields higher prediction accuracy.Thus,the effectiveness and feasibility of the CEEMDAN method is proven.