A Data-Driven Approach for Short-term Electricity Consumption Forecast based on Hybrid Intelligent Method

Short-term electricity consumption forecast is widely used in electricity analysis, regional planning, and energy management. Electricity consumption is affected by a variety of variable factors and has strong non-linearity and complexity. The existing studies usually hard to achieve satisfactory results in dealing with nonlinear time series data. Therefore, this study proposes a data-driven hybrid model with signal decomposition and correlation forecasting technique based on high-resolution historical data. The consumption curves in the time domain are converted to several sub-sequences of different frequency bands by the variational mode decomposition (VMD) to describe signal characteristics accurately. By taking the feature vectors and sub-sequences of consumption as the inputs, the BiGRU is used to capture the contextual information from the complete time series. The attention mechanism is used to mining useful information and assigning weights to make short-term consumption forecast. The quality of data mining has been improved, and the effect of the forecast has also become better. The experiments are carried out in data sets of Shanghai and New York respectively. The results of the experiment show that this method has better prediction performance than other standards and the latest prediction methods.