Short-Term Residential Electricity Forecast Based on Hybrid Method Inspired by Gate Strategy

Residential electricity consumption forecasting is of great significance to electricity dispatching and balance. However, electricity consumption series have high non-linear, non-stationary and random characteristics, making it difficult to obtain satisfactory forecasts. This paper proposes a hybrid method motivated by the multivariate function derivation mathematical idea. First, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is adopted to decompose the complex original series into several relatively simple and stationary sub-sequences to further reduce the forecasting difficulty. Second, since the bidirectional long short-term dependencies extracted by bidirectional gated recurrent unit (BiGRU) are not all favorable, we add a new gate mechanism to select bidirectional hidden states and propose a gate-augmented BiGRU (GA-BiGRU) to forecast the above sub-sequences. Third, instead of linear or fixed weight combination, a gate-augmented combination way (GAC) is designed to integrate the above forecasts via dynamic weights. The proposed hybrid method, namely CEEMDAN+(GA-BiGRU)+GAC, not only makes full use of the original data but also improves forecasting performance via the gate strategy. Two real-world residential hourly electricity consumption forecasting cases shown that the proposed method is superior to multiple single and hybrid methods in terms of four performance metrics.