Objective Aiming at the typical characteristics of natural gas load data such as noise difficulty in capturing regularity and feature redundancy a multi-feature natural gas load forecasting method based on secondary decomposition and bidirectional long short-term memory BiLSTM model is proposed. Methods First random forest RF was used to rank the importance of features and important associated features were screened out. Second singular spectrum analysis SSA was employed to decompose and reconstruct the original data extract the trend and periodic information of the data to form components and separate the high-frequency components. Then variational mode decomposition VMD was used to conduct secondary decomposition on the high-frequency components to generate individual modal components. These modal components were fed into the BiLSTM model for prediction. Finally the prediction results of each component were fused to obtain the final load point prediction result. Results Simulation results showed that compared with other traditional models the proposed secondary decomposition and combination model had a better prediction effect and higher prediction accuracy. Conclusion The newly built model has good interpretability and rationality and can be effectively applied to natural gas load forecasting providing a reliable theoretical basis and effective technical support for decisionmakers to reasonably make natural gas supply plans.