基于时序AR补偿RBF模型的滑坡位移预测

Landslide displacement prediction based on RBF neural network compensated by time-series AR

为了提高复杂态势下滑坡位移预测的准确性,构建了基于时序AR补偿RBF神经网络的滑坡位移预测模型。首先采用RBF(Radial Basis Function)神经网络对滑坡位移整体趋势进行逼近,获取预测残差;然后基于时序AR构建预测残差补偿器;最终将AR预测残差值与RBF逼近值进行叠加,从而实现滑坡位移预测。以隔河岩水电站进水口滑坡38期监测数据为例,采用AR补偿RBF模型进行预测。预测结果表明:相较于单一RBF神经网络,AR补偿RBF模型的预测平均相对误差由12.718%降低至4.703%,均方误差由0.232降低到了0.032;AR补偿RBF模型对滑坡位移拐点、突变点的逼近更符合实际,且具有较高的外推预测能力。

In order to improve the accuracy of the landslide displacement prediction under the complex conditions,a landslide displacement prediction model based on Radial Basis Function(RBF)neural network compensated by time-series AR was constructed.First of all,the learning and nonlinear approximate abilities of the RBF was used to extract overall trend term of the landslide displacement,then the prediction residual was obtained.After that,we constructed a prediction residual compensator based on time-series AR model.Finally,the AR predicted residual value was superimposed with the RBF approximation value to realize the landslide displacement prediction.The effectiveness of the model was verified by using the data of the landslide at water intake slope in Geheyan Hydropower Station.The results showed that after compensated by AR,the average relative prediction error was reduced from 12.718%to 4.703%comparing with the single RBF network,and the mean square error was reduced from 0.232 to 0.032.The RBF model compensated by AR is more suitable for the data series approximation at turning points and abrupt points of the landslide displacement,and has high extrapolation prediction ability.