Short-term residential electricity consumption forecast considering the cumulative effect of temperature, dual decomposition technology and integrated deep learning

Abstract

At present, the electricity market reform has entered a deep area, electricity consumption forecasting has become increasingly important, accurate electricity consumption forecasting provides a reference basis and decision-making support for power dispatching and market transactions, and residential power consumption prediction can help users choose appropriate power suppliers and power supply programs according to the market situation, and provide the reliability and economy of power consumption. Residential electricity consumption is complex, affected by many factors and prone to significant noise disturbances, which results in electricity consumption data that is characterized by non-stationary, intermittent and erratic fluctuations. Therefore, using a single model is challenging to accurately predict household electricity consumption. To meet this challenge, this paper designs the structure of this three-step residential electricity consumption forecasting. At the first stage, we first analyzed cumulative effects of temperature on residential electricity consumption, and an hourly temperature correction model combining meteorological factors is constructed, the adjusted hourly temperature data are then entered into the predictive model. We have developed a Variable Modal Decomposition (VMD) data decomposition technique optimized for non-governmental organizational models, which improves the problem of subjectivity in parameter setting in traditional VMD, thus enhancing the performance and accuracy in data decomposition. In the second stage, developed a BiLSTM-AM based integrated deep learning model and dynamically adjusted the weights of the influencing factors by introducing an Attention Mechanism (AM) to enhance the stability of the model, also predict multiple IMF components obtained after the NGO-VMD decomposition, respectively. In the third stage, the training residuals of the BiLSTM-AM model are used as target variables to correct the prediction error in BiLSTM-AM using the XGBoost regression model. A variety of model configurations were constructed using actual data from a coastal province in southern China, and the computational results show that the integrated prediction model exhibits excellent stability and accuracy.