WOA-BiLSTM-Attention Model for Short-term Power Load Projection

Aiming at the periodicity, nonlinearity and certain volatility of power load, this study employs historical power load data and the defining characteristics of its influencing factors for the purpose of multivariate short-term power load prediction. Subsequently, an integrated model (WOA-BiLSTM-Attention) is introduced, utilizing the whale optimization algorithm to synergize the attention mechanism and the bi-directional long short-term memory (BiLSTM) network. This innovative approach serves as a novel methodology for short-term power load forecasting. Initially, the utilization of BiLSTM involves power load prediction, wherein both historical and prospective power load data, along with their respective influencing factors, are harnessed to comprehensively capture the inherent temporal intricacies of the dataset. Following this, the Attention mechanism comes into play, acquiring the ability to independently discern the significance attributed to diverse features and chronological segments. By prioritizing attributes intricately associated with the target, it adeptly manages the extended interdependence intrinsic to the data, thus culminating in an augmented predictive accuracy. Finally, Whale Optimization Algorithm (WOA) is utilized to find the optimal neural network parameters to reduce the model stochasticity and time costs. The utilization of electricity load data from Panama serves as an illustrative case for the conducted experiments, revealing that the WOA-BiLSTM-Attention model attains an elevated level of prediction accuracy compared to single neural network methods such as BP, LSTM, BiLSTM, etc., corresponding to an R
²
of 0.93 and an RMSE of 29.49.