Abstract: The Bayesian statistical theory was adopted to improve traditional neural network algorithms, and constraints representing network structural complexity were introduced to the network objective function in order to avoid over-fitting the networks and enhance the generalization ability. The improved networks were applied to strip thickness prediction in Jigang 1700 mm mill, and the prediction result is superior to that of traditional neural networks in forecasting accuracy, training time and network stability. Then, the Bayesian neural networks were used to predict the plasticity coefficient of strips. Finally, the real-time forecasts of exit thickness and plasticity coefficient of strips were synthetically utilized in the thickness control system of hot strip rolling to improve strip quality further.