Concrete placing productivity using a novel neural network design

For planning purposes an accurate estimate of the productivity for insitu concreting operations is desirable. Various methods have been investigated to model such productivity; this paper will consider a novel neural network architecture the Twin Nested Recurrent Network (TNRN), in the search for increasingly accurate predictions. Neural networks are trained by providing them with a set of data from historic projects. Providing a larger training data set generally increases the accuracy of the output. In conventional neural network modelling, this training set is based upon the outcomes of a series of individual operations and thus is limited. The approach to be used here would expand this training set by considering the productivity of each individual concrete delivery used in each operation thus creating a much larger training set of data from which the neural network can learn. Further, by preparing the data in this way, an original architecture can be developed for the model. This architecture is a twin-loop nested recurrent network. The network is presented with each delivery for a given operation, and then with the productivity of the operation itself. Initial results are encouraging: this approach allows a much greater flexibility and a significant reduction in the sensitivity of changes to the network training functions.