Abstract: The short-term power load data has complex time series characteristics and nonlinear characteristics, and the traditional prediction model is difficult to accurately predict the short-term power load. Therefore, this paper considers the influence of the date, weather and other factors on the power load, and proposes a fusion bidirectional long-term and short-term power load. Memory Neural Network, Attention Mechanism, and Short-Term Electricity Load Prediction Model with Improved Equalization Optimizer (IEO-BiLSTM-Attention). The model uses the BiLSTM-Attention neural network to learn the internal rules and dependencies of the sequence, and uses an improved balanced optimizer to optimize the selection of hyper-parameters in the neural network model to improve the accuracy of prediction and the performance of the model. Among them, the improvement strategy of this paper will start from three aspects. First, in order to solve the problem of low initial population history caused by the high randomness of the initialization of the equilibrium optimizer, the optimal point set method is introduced to initialize the population strategy; secondly, this paper proposes the Golden Levy guidance strategy , in order to improve the defect of the population diversity reduction of the equilibrium optimizer in the later iteration. Finally, this paper proposes an adaptive keyhole imaging reverse learning mechanism to improve the search efficiency of the equilibrium optimizer. The performance of the algorithm proposed in this paper is verified by using the classical benchmark function. The experimental results show that the algorithm proposed in this paper has better optimization accuracy and convergence speed. In addition, the model proposed in this paper is applied to the power system load data set. The experimental results show that The model proposed in this paper has better accuracy and stability.