Research On Long Term Load Forecasting And Power Load Detection Of Smart Grid Based On Deep Learning

Electric power industry is the basic industry of national development,which profoundly affects the sustainable development of national economy.With the continuous development of modern smart power grid technology,power grid construction scale is increasing,which brings more complexity for power system data collection and data analysis tasks.Power data is characterized by strong timing,high data density,high collection cost,and more nonlinear representation.At present,power data analysis has gradually become a key task in power system,because enterprises need to carry out long-term capacity planning for power generation and transmission equipment,and monitor the abnormal behavior of power load in real time.Therefore,accurate prediction and detection of power load data has become the premise to ensure the safe and reliable operation of power system.With the advent of the era of big data and artificial intelligence,deep learning models have been used to solve the problem that traditional models cannot be fully trained in power load data analysis tasks due to the insufficient feature learning ability of the model.However,there is still a lack of mature models and solutions for power load forecasting and load detection business.The research is facing the following two challenges.First,for the electricity load forecasting business,many models are easy to ignore the potential feature correlation,periodicity and trend attributes inside the data in the long sequence forecasting task,which will cause insufficient feature learning,resulting in the performance of the model will decline sharply with the growth of the data prediction length,and the prediction accuracy will decrease.Secondly,for the electricity load detection business,many models rely too much on labeled data samples for supervised learning,which leads to the high training cost of the model and the low inference efficiency.On the other hand,the model does not deeply learn the general knowledge inside the features,which makes the model’s generalization ability insufficient.In view of the above challenges,this thesis proposes the corresponding solutions around the two key businesses under the smart grid power system: long-term load forecasting and electricity load detection tasks.The main research work is as follows:Firstly,in order to solve the problem of low prediction accuracy caused by insufficient feature learning in many long sequence load forecasting models in recent years,this thesis proposes a long-term load forecasting model based on hierarchical decomposition Self-attention mechanism encoder(LTSNet).The core idea is to apply a tree-like hierarchical architecture,and use stacked residual self-attention blocks to gradually learn the potential nonlinear feature representation of the data,extract highorder sparse features in depth,and introduce time slice feature interaction in the training process to enhance the covariance stability of the data between the temporal neighborhood,and finally perform long sequence prediction through a forward generative decoding.Through this design,the model has more stable performance in capturing the correlation,trend and periodicity of features,and enhances the accuracy of long sequence prediction.Secondly,in view of the problems such as high cost of model training and poor mobility caused by overreliance on supervision training for power load detection tasks in recent years,this thesis proposes a power load representation detection model(TSCNet)based on comparative learning in order to effectively utilize the value of unlabeled data and learn transferable knowledge in data.By introducing time domain comparison prediction and context comparison prediction,this model learns the similarity and difference between unlabeled data samples.The core idea is to form positive and negative sample pairs of source samples through two data enhancement methods with different strategies and frequencies.A stacked GRU(Recurrent Unit)network with attention mechanism is used to extract the potential timing representation of different samples,so as to conduct subsequent contrast Loss training.The contrast losses used in this thesis are Info NCE Loss and NT-Xent loss.After the training of the baseline model is complete,The weight transfer can be used to finetune the model of the downstream task,accelerate the model training speed and improve the accuracy of the model classification.Finally,this thesis aims at the above two tasks: ETT(electric Transformer Temperature,power transformer temperature),Electricity(Household consumption data set),SGCC(State Grid Corporation of China,Extensive comparison experiments with several classical baseline models have been conducted on power time series data sets such as State Grid.The experiments prove that LTSNet and TSCNet can improve the performance and efficiency of data mining more effectively in the two tasks of long series load prediction and power load detection respectively.