Short-term power load prediction based on VMD decomposition and LSTM-Attention optimized by improved DBO

Given the challenge that traditional methods in short-term power load prediction are impacted by the complexity of power load data, resulting in poor prediction accuracy, this paper presents an effective approach based on a ‘decomposition-reconstruction’ architecture. This method integrates the Variational Mode Decomposition (VMD), Dung Beetle Optimization (DBO) algorithm, Long Short-Term Memory (LSTM) network, and Attention (ATTENTION) mechanism. Firstly, to simplify the data, VMD is applied to decompose the load sequence. Subsequently, The LSTM is employed to capture the time-series attributes of each element. The Attention mechanism is integrated to accentuate crucial information. Moreover, the improved DBO algorithm is employed to fine-adjust the variables of the LSTM, thereby boosting the model’s performance. Experiments based on actual power load data show that, compared with traditional methods, this approach enables a more precise capture of load variation patterns. As a result, it greatly enhances the precision and credibility of short-term power load prediction.