Electrical load forecasting based on variable T-distribution and dual attention mechanism

Accurate load forecasting is essential for power system stability and grid dispatch optimization. However, this task is challenging due to the inherent instability and volatility of the load sequence. To address this problem, this paper proposes a novel load forecasting model that integrates periodicity detection, the variable t-distribution, and the dual attention mechanism. Periodicity detection has been incorporated into the self-attention mechanism for the first time, identifying the most significant period in the raw load sequence. Subsequently, the raw load sequence undergoes processing using empirical wavelet transform, resulting in a series of subsequences. A feature attention mechanism is then employed to extract relevant input features. Furthermore, a novel variable t-distribution distance matrix is introduced into the temporal self-attention mechanism, enhancing the influence of data at identical or nearby positions in other periods based on the length of the most significant period. This modification improves the capacity of the vanilla self-attention mechanism to effectively model the relationship between data at varying distances. The hyperparameters of the variable t-distribution are obtained through Bayesian hyperparameter optimization. Empirical evaluations on two datasets with distinct meteorological and load features show that the proposed model outperforms baseline models across all metrics.