The grid inertia magnitude measures the frequency stability of the system, and accurate prediction of the system inertia level in advance can avoid the risk caused by low inertia. To this end, a multi-algorithm hybrid neural network model based on modal decomposition and feature fusion for short-term prediction of system equivalent inertia is proposed. Firstly, an improved complete ensemble empirical mode decomposition with adaptive noise is used to decompose the inertia of four seasons, and a new sequence is obtained by reconstructing the new inertia based on the fine composite multi-scale fuzzy entropy of each decomposed component. Secondly, the minimum redundancy maximum relevance method is used to measure the correlation between different decomposition components and different features, and a subset of highly correlated and low redundancy features is filtered out. Finally, a bidirectional long-short-term memory network model based on Bayesian optimization algorithm is used to predict different components of different seasonal inertias, and the final prediction results are accumulated. Typical examples at home and abroad are selected for testing, which verify that the proposed method can effectively balance the prediction accuracy and prediction time, and solves the problem of seasonal differences affecting the prediction results of system inertia.