A multi-energy loads forecasting model based on dual attention mechanism and multi-scale hierarchical residual network with gated recurrent unit

Multi-energy loads forecasting (MELF) is crucial for the effective management of integrated energy systems (IES) and the balance between energy supply and demand. Nevertheless, a complex coupling relationship exists between multi-energy loads, and they are also influenced by external factors such as meteorological conditions, calendar information, and random user behaviors. Moreover, existing methods are usually difficult to capture the characteristics of multi-energy loads with obvious regularities, which limits the prediction accuracy. To address these challenges, we design a MELF method based on a dual attention mechanism, multi-scale hierarchical residual network, and gated recurrent unit (GRU), referred to as the DAM-MSHRN-GRU method. First, we design a dual attention mechanism that allocates suitable weights to various time points and input features, mitigating the impact of time and external factors on prediction accuracy. Next, we develop a multi-scale hierarchical residual network to extract both short-term load fluctuations and long-term periodic load characteristics, enhancing the forecasting capability during periods of significant load volatility. MSHRN uses depthwise convolution residual blocks with different kernel sizes to convolve the exogenous features of multi-energy loads one by one to capture the regularities of multi-energy loads. Finally, we utilize GRU to capture the temporal patterns of the load. Simulation results demonstrate that the DAM-MSHRN-GRU obtains higher forecasting accuracy compared to existing models.