Short-term Load Forecasting By Long And Short-term Temporal Networks With Attention Based On Modal Decompositions

Electricity load forecasting is of great significance for the stable operation of power systems,the reduction of carbon emissions,and the improvement of power supply reliability.In this paper,with the research objective of improving the accuracy of load forecasting,we propose an improved forecasting method from both load data set and forecasting model.To achieve this goal,the following three aspects are studied:(1)For the load dataset processing problem,this paper firstly performs data cleaning on the entire dataset,including missing data filling and outlier data processing,etc.;secondly,feature screening is performed on the multi-factor weather feature data to avoid the influence of weakly correlated and irrelevant redundancy on the prediction accuracy;finally,the complete empirical mode decomposition with adaptive white noise is used for the load sequence complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)algorithm is used to decompose the load series data into intrinsic mode functions(IMF)with different frequency ranges,which reduces the complexity of the load series data.The complexity of the load series data is reduced,and the prediction difficulty is effectively reduced.(2)In order to improve the accuracy of the short-term load prediction model,a long and short-term temporal networks(LSTNet)method based on modal decomposition is designed in this paper.Firstly,the model uses the CEEMDAN algorithm to decompose the load series data into several IMF components;then the LSTNet prediction model is constructed,which includes a convolution module,a loop module,a loop-skipping module and an autoregressive module.The convolution module is used to extract the local information in the time series,the loop module and loop-skipping module are used to capture the long-term and short-term dependencies in the time series,and the autoregressive module is used to optimize the neural network for the problem of insensitivity to linear feature identification.Finally,the IMF components are input into the LSTNet prediction model separately,and the obtained results are combined to obtain the final prediction results.The superiority of this prediction model is verified by comparative analysis and ablation analysis through case studies.(3)In order to optimize the performance of the prediction model,an optimized prediction model based on temporal pattern attention(TPA)is designed in this paper.The weights of each time step are dynamically calculated using the attention mechanism,and then the weights are reasonably assigned to improve the model prediction capability by focusing important features with high weights and eliminating redundant features with low weights.The Adam W optimizer is also used so that the prediction model can automatically adjust the learning rate and weight decay coefficient to enhance the stability of the model.Finally,the optimization performance is verified by example analysis.In summary,the long-short time series network prediction model based on modal decomposition and attention mechanism designed in this paper is improved in terms of power load data decomposition and prediction model structure,respectively,which is of practical value for short-term load prediction models with multi-factor effects and provides a new solution for load prediction research.