The share of renewable energy in modern power systems is increasing, causing its load to fluctuate more erratically than in conventional power systems. This volatility leads to lower accuracy of load prediction. To address this issue, this paper introduces a short-term load prediction model combining the dung beetle optimization algorithm (DBO) with optimized variational mode decomposition (VMD) and an improved whale optimization algorithm to optimize bidirectional long short-term memory (IWOA-BILSTM) neural networks. The DBO is used to optimize the VMD, the time series data is decomposed, and various feature data are classified according to the minimum envelope entropy. This enhances the decomposition effect. The fluctuation of the data is reduced by effectively decomposing the original data. Then the whale optimization algorithm is improved using a nonlinear convergence factor, adaptive weight strategy and random difference variation strategy to enhance the local and global search ability of the whale optimization algorithm. Thus an improved whale optimization algorithm (IWOA) is obtained, and it is then used to optimize bidirectional long short-term memory (BILSTM) neural networks, increasing the accuracy of model predictions. Finally, this method is tested on real load data from a location, yielding favorable results. The resulting metrics for relative root mean square, mean absolute and mean absolute percentage errors are recorded at 0.0084, 48.09, and 0.66%, respectively. These outcomes verify the effectiveness of the proposed model in short-term load prediction.