Research on Short-Term Load Forecasting Model Based on CEEMDAN-CNN-LSTM-AM

Abstract

On the effective running and regulation of the building power system, the accuracy of the building load forecasting plays a significant role. A short-term predictive model based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN)-Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM)-Attention Mechanism (AM) is put forward in light of the characteristics of time series, non-linearities and destabilization of load data. The initial stage is to deconstruct the primitive data up into different modal parts by using CEEMDAN approach. CNN then extracts the required values, and LSTM is used for load sequence prediction. AM is integrated into the load prediction output of LSTM to weight the variables, and a load prediction model is established. The final prediction results are achieved by aggregating all components when each modal component's prediction have been calculated. Basic construction load data was used to evaluate the CEEMDAN-CNN-LSTM-AM model, and the findings demonstrated its usefulness in short-term load prediction.