Research On Short-term Load Forecasting Of Power System Based On Deep Learning

Under the drive of the "carbon neutrality" goal,constructing a new power system with new energy as the main body is an inevitable trend in the transformation of modern power systems.With the development of smart grid technology and the widespread use of smart meters,the power system can collect massive amounts of electricity consumption data.Based on this,maintaining real-time power balance on the generation and load sides of the grid and maintaining stable grid operation,accurate prediction of future demand for electricity consumption can be made using the collected data,providing data support for load scheduling.Therefore,combining the current widely used deep learning technology with the non-linearity,time-series,and uncertainty characteristics of power loads to accurately predict power load can improve energy utilization efficiency,reduce power grid operation costs,and have important practical significance.In this research,based on multiple actual and open-source datasets,we used deep learning technology for short-term load forecasting.The research content of this article includes:(1)using the maximum information coefficient algorithm to calculate the features in the dataset,selecting the main influencing factors,and pre-processing them,including processing abnormal data and missing data and using corresponding encoding methods for different features.In addition,we introduced the method of deploying the model to edge computing devices using the Tensor Flow framework.(2)We proposed a short-term load forecasting method based on TCN-GRU-Attention.To achieve rapid feature extraction of power load sequences,we used a temporal convolutional network to extract time-series features,and used gated recurrent units as the decoder to generate the predicted load value.At the same time,we added a temporal attention mechanism to alleviate the problem of model performance decline,and used an external attention mechanism to re-encode the model input and input it into the temporal convolutional network.The Experimental analysis confirms the effectiveness of this model.(3)We proposed a short-term load forecasting method based on PSA-Autoformer.The original Autoformer model extracts time-series features by calculating sub-sequence similarity.On this basis,this article introduces probability sparse self-attention mechanism and convolutional distillation mechanism to enhance the model’s feature extraction ability and reduce its computational complexity respectively.The experimental results show that the model has high accuracy.