With the rapid development of the smart grid, the power system has been driven toward intelligent and efficient development. In the process of predicting power load, problems such as unreasonable decomposition of load data and low prediction accuracy have been observed. A BiLSTNet combined prediction model based on optimized quadratic decomposition was proposed. Firstly, the load data were decomposed and reconstructed using improved complete ensemble empirical modal decomposition with adaptive noise (ICEEMDAN) to reduce the overall noise effect of the data. Secondly, the variational mode decomposition (VMD) was applied for the secondary decomposition and reconstruction. To optimize the decomposition parameters of VMD, the enhanced sand cat swarm optimization (ESCSO) was utilized, in which the Kent chaotic map was introduced to initialize the algorithm population, and both the vertical-horizontal search strategy and adaptive t-distribution perturbation were adopted to enhance global search and the ability to jump out of the local optimum. Finally, the reconstructed data were fed into BiLSTNet hybrid model for prediction. Experimental results demonstrate that the proposed method achieves better decomposition results and prediction accuracy on the Australian electricity load dataset than other traditional hybrid models, showing higher practical value.