Short Term Load Forecasting Based On Temporal Convolutional Network And Multi-scale Fusion

Short-term power load forecasting is the foundation for ensuring the operation of the national power system.Accurate power load forecasting plays an important role in ensuring the stable operation of the power system.Therefore,how to reduce the error of power load forecasting has been the focus of scholars’ research.The traditional short-term power load forecasting methods have high prediction accuracy and are easy to implement,but when faced with complex,nonlinear,and non-stationary load data,the prediction effect is not ideal.Moreover,the problem of subsequence fusion that is not fully considered after modal decomposition exists.Aiming at these issues,this thesis conducts research on short-term power load forecasting by combining power load data from different dimensions.The main work of this thesis includes:(1)Building a short-term power load forecasting model VMD-TCN based on Variational Mode Decomposition(VMD)and Temporal Convolutional Network(TCN).This model combines the VMD decomposed sub-sequences with the time-domain convolutional network by utilizing the flexibility of TCN to adjust memory length and its good performance in training networks with long sequence length.By sequentially training multiple sub-sequences obtained from VMD decomposition,the model can better represent the temporal features of the data.(2)Optimizing the learning rate parameter in neural network training through Cosine Annealing(COSA)to accelerate model convergence during network training.Later,using Fully Connected(FC)to fuse sub-sequences of different time scales to construct the VMD-MTCN-COSA-FC power load prediction model,aiming to improve the prediction accuracy of short-term power load forecasting.The superiority of the model is verified by comparing it with similar methods using actual power load datasets.(3)Building a multi-dimensional short-term power load integrated prediction model TCN-MLSTM based on TCN and Multi Long Short-Term Memory(MLSTM).MLSTM is used to extract the relevant features between multi-dimensional data and power load data,while TCN is used to extract the temporal features of the data.This allows the model to fully extract the data features for multi-dimensional data and achieve better prediction accuracy.The feasibility of the integrated model is verified using actual data.