Development and Application of Dual Attention-based Short-term Forecasting Intelligent Model for Industrial Electricity Consumption

Industrial power consumption accounts for more than 70% of the total in China, it is of great practical significance to forecast industrial electricity consumption for power regulation and maintenance of the balance between supply and demand. Based on actual data, this study devotes to building an Attention-based Short-term Forecasting Model for electricity consumption (ASFM). Firstly, the Ensemble Empirical Mode Decomposition (EEMD) is used to decompose the time series data into a collection of detailed sub-series. Then, Fine-to-Coarse reconstruction algorithm is adopted to generate high-frequency, low-frequency, and trend components as input variables. An Encoder-Decoder equipped with dual-stage Attention mechanisms is used to grasp spatio-temporal characteristics of the input data. A two-phase convolution network is built to further extract the short-term mutation characteristics. The final forecasting result could be calculated by results fusion. In order to verify the performance of the proposed ASFM, we use three actual data-sets for comparing. The results indicate that the ASFM realizes higher accuracy and better portability than ARIMA, ELM and LSTM.