Improvement and Assessment of Neural Networks for Structural Response Prediction and Control

In an extension of a previous paper, prediction accuracy is improved for neural networks to be used as part of an adaptive structural control system. This improvement will enable reliable predictions of performance variables such as displacements and control forces further into the future. This allows more lead time for controller adjustment should a performance variable be predicted to violate a prescribed constraint. The improved prediction accuracy is due to the use of the Levenberg–Marquardt algorithm in training the neural network and the use of a single neural network for more than one performance variable simultaneously. With these improvements, far fewer iterations (and more importantly less computer processor time) are used in the neural network training, and most importantly the prediction accuracy is greatly improved. These improved neural network predictions are then compared to other prediction methods: a polynomial fit of past data and the use of the state transition matrix.