A k-nearest neighbor attentive deep autoregressive network for electricity consumption prediction

Xihe Qiu, Yajun Ru, Xiaoyu Tan, Jue Chen, Bin Chen, Yun Guo

Abstract: Electricity is vital in daily life and crucial for sustainable economic development. Accurate forecasting of energy consumption ensures efficient electricity system operation and supports strategic decision-making for energy distribution. Current time-series methods neglect the influence of neighboring regions’ electricity consumption and the varying impact levels caused by multiple factors on the target area. Therefore, we propose the k-nearest neighbor attentive deep autoregressive network (KNNA-DeepAR) model, which combines a k-nearest neighbor approach with an attentive deep autoregressive network, to achieve precise short-term electricity consumption predictions. By extracting informative features from historical time-series data, we incorporate electricity consumption data from the k regions closest to the target area as additional variables. Leveraging the attention mechanism, we assign varying weights to each variable to capture their interdependencies. Experimental results on a public dataset of electricity loads in fourteen U.S. regions demonstrate the superiority of our model. Compared to state-of-the-art time-series models, our model achieves higher predictive accuracy and exhibits significant potential as an effective approach for accurately predicting electricity consumption and other time-series tasks.