To solve the problem of intermittent missing sensor data in wind pressure measurements， a method for predicting the wind pressure time series on the surface of structures by combining the improved empirical mode decomposition algorithm （IEMD） with bidirectional long short-term memory （BiLSTM） networks is proposed. First of all， intrinsic mode functions （IMFs） are obtained through the adaptive decomposition of wind pressure time series using the IEMD method with soft sifting stopping criterion. The sample entropy is adopted to reconstruct IMFs and obtain subsequences. Then， BiLSTM networks are established and trained with these subsequences， while the prediction is also conducted， in which the Bayesian algorithm is employed to optimize the hyperparameters of neural networks. Finally， a case study of wind load prediction based on the wind tunnel test data of low-rise buildings is conducted to validate the proposed model. The results show that， compared with the traditional prediction models （e.g. multi-layer perceptron， BiLSTM）， the prediction model based on IEMD and BiLSTM presents higher accuracy and calculation efficiency， which is capable of predicting Gaussian and non-Gaussian wind pressure signals.