Hybrid Decomposition-Deep Learning Model for Energy Load Prediction

Energy data have a complex behavioral pattern due to many factors that influence the energy market. Energy forecasting promotes the effective and reliable operation of power system management and efficient allocation of resources, especially during an economic crisis. To reduce the difficulty of load volatility, noisy measurement, and non-linearity and improve short-term power load forecasting, we proposed a novel data-driven hybrid model for predicting load datasets with different time horizons. We used complete ensemble empirical model decomposition with adaptive noise (CEEMDAN) to decompose load data into several subsequences’ intrinsic mode functions (IMFs) and a residual while dealing with the mode mixing problem. The proposed model integrated bidirectional long short term with attention (BiLSTM-AM) and deep neural network (DNN) to predict the modes. The hyperparameters of BiLSTM-AM-DNN were tuned using the Bayesian optimization method to achieve high accuracy. DNN was integrated at the end of BiLSTM-AM to extract more accurate characteristics from load data and further improve the forecast. Accurate uncertainty was estimated for risk management and planning. The proposed model achieved an empirical coverage above 95% on average. A thorough comparison was made to verify the superiority of the proposed model compared to the competing models.