Research Of Time Series Missing Values Imputation Method In Ecological Monitoring Stations

Today,ecological data are generally collected and transmitted by wireless sensor networks in large-scale ecological monitoring stations,realizing real-time and remote observations across geographically distributed regions.The wireless sensory data often inevitably suffer missing values that occur continuously in a larger range due to the unreliability of system and instrument failures.It is different from the traditional manually-collected datasets with the data losses being sparsely distributed.Until now,however,researchers have little knowledge about infilling continuous missing values in time-series datasets of ecological stations.Hence,how to infill this type of missing values efficiently and accurately is urgent issue in ecological monitoring stations.In the sensor networks and ecological monitoring stations,the data are generally in the form of time series.Essentially,this issue can be reckoned as multi-step prediction of time series.This paper firstly investigates eight typical time series missing values imputation methods,and compares their performances with different missing scales and different fluctuations.On this basis,we make improvements on radial basis function(RBF)neural networks with a good imputation performance.In order to avoid overfitting or underfitting due to the inappropriate parameters setting,this paper proposes a novel parameters learning approach,called EasyRBF,to efficiently determine the parameters of the RBF neural network model.The key ideas of EasyRBF are(1)separating the optimizations of RBF network parameters in stages,and(2)leveraging two delicately nested PSO(particle swarm optimization)procedures to simultaneously solve the critical RBF network parameters,different than traditional PSO-based learning schemes.Extensive numeric experiments are conducted over Australian ecological monitoring stations datasets,to evaluate the infilling performance of EasyRBF.The results demonstrate that EasyRBF can achieve higher accuracy of infilling the soil dataset with large-scale and continuous missing,in comparison with other typical infilling approaches.