Skip to content

Abstract

Anti-sway control of tower cranes requires knowing where the payload will be based on the input data. This paper measures how far ahead payload sway can be predicted from its own recent history, using an open experimental dataset recorded on a laboratory Inteco tower crane. Five methods, from a zero forecast to an ensemble of feedforward networks, are compared across prediction horizons from 0.01 s to 1.0 s and across three families of crane motion, with generalisation to unseen cable lengths using leave-one-scenario-out validation. The results are as follows. One-step prediction is degenerate: linear regression on three lags reaches 0.0796°, below the 0.0879° encoder resolution, so no comparison of methods is useful at that horizon. A linear model remains adequate to roughly 0.07 s and becomes worse than predicting zero sway near 0.85 s. The dominant error at long horizons is a phase error: oscillation amplitude is reproduced to between 79% and 93% while the predicted waveform lags the true one by 10% to 17% of a period. Only the network ensemble stops degrading, having a normalised error of 0.416 between 0.5 s and 1.0 s, against 0.612 for the trees and 0.883 for linear regression. Cable lengths outside the training range worsen the predictions in every method, and the loss grows with horizon: at 1.0 s the linear model is worse than the zero forecast on extrapolation folds (1.304) and the tree ensemble loses 65%, because past its last split threshold a tree simply returns the value of the longest cable it was trained on. How far an unseen cable lies from the training range is better measured in natural frequency than in metres. The two ends of the trolley range are separated by similar distances in metres, but their natural frequencies differ by a factor of 4.8. In both families tested, predictions are more difficult at the short-cable end. The accuracy is 1.4 to 2.8 times worse while the crane is under command than during free decay, which is the regime in which a controller actually has to operate. The free-oscillation period produces an effective pendulum length: L_eff = 0.924L + 0.166, reproduced across five measurements, including three excitation types and two axes.