Abstract: 【Objective】 Based on the meteorological data related to water demand forecasting of winter wheat, a water demand forecasting model with fewer parameters was constructed to improve the robustness of water demand forecasting,provides a more reliable method for forecasting water demand based on meteorological information. 【Methods】 Meteorological data of Qitai County in recent five years were selected, and the water requirement of winter wheat calculated by Penman-Monteith formula was approximately the real water requirement. Four variables including average temperature, wind speed, humidity and precipitation were taken as input parameters. The water requirement of winter wheat was forecasted, and the prediction of CNN-BiLSTM was compared with that of LSTM, BiLSTM and other 6 models. 【Results】 The results showed that when a few parameters were fed into BP, RNN, LSTM, improved BiLSTM and CNN-BiLSTM models to predict water demand, the prediction effect of BP neural network was poor. In the model evaluation, CNN-BiLSTM showed an R² improvement of about 14% over LSTM and a MSE reduction of about 3.8. 【Conclusion】 CNN-BiLSTM model is more accurate in predicting wheat water demand.