An Attention-based CNN-GRU Model for Resident Load Short-term Forecast

Under the background of electricity market, accurate short-term forecast of residential load can effectively strengthen demand side management. However, the randomness and uncertainty of residential electricity consumption increase the difficulty of prediction. The traditional prediction models are facing with some problems that it is difficult to effectively extract high-dimensional features and important information is easy to lose when the length of the time series is too large. Based on these, the paper established an Attention-based CNN-GRU model for resident load short-term forecast. In this model, CNN plays a role in extracting the features of historical power consumption data, GRU neural network combined with Attention mechanism is used to model these extracted features dynamically. Real data of individual household electricity consumption is taken as an example, and simulation experiments are carried out in three scenarios of univariate single step, multivariate single step and multivariate multi-step prediction respectively. The results of the experiment demonstrate that the proposed model achieves reasonable accuracy and robustness.