水资源监测异常数据模态分解-支持向量机重构方法

完备真实的水资源监测数据是支撑数据分析与决策的基本前提。在梳理现阶段水资源监测异常数据的基础上,提出运用移动平均拟合初筛来直观辨识异常监测数据,进而选取集合模态分解对非可直观辨识异常监测数据进行挖掘的方法。将剔除异常监测值后的时序数据作为基于粒子群优化最小二乘支持向量机模型的模拟样本,并利用其恢复所剔除的异常监测数据。对水务公司日取水量监测数据的实证分析结果表明,通过移动平均拟合与模态分解可较大限度地保留含有异常数据的特征向量并实现数据的有效重构,相比传统的统计方法其具有更好的适用性;运用粒子群优化的最小二乘支持向量机可进一步提高对剔除异常值数据的拟合效果,且符合水资源监测数据的季节波动规律特征及对实际取用水状态的客观反映,据此可相对合理地达到恢复所剔除异常监测数据的目的。

Methods of Abnormal Data Detection and Recovery for Water Resources Monitoring Based on EEMD and PSO-LSSVM

The national water resources monitoring capacity building project which started in 2012 in China is an important way to improve the level of water conservancy information. It requires that the historical time-series monitoring data of water resources should be complete and reliable so that it can be used to support data analysis and decision making. The basic scenarios for monitoring abnormal data were summed up and a comprehensive model was proposed, aiming at abnormal data detection and recovery. Moving average fitting and ensemble empirical mode decomposition (EEMD) method were introduced to identify both observable and non-observable abnormal monitoring data. The particle swarm optimization based least squares support vector machine (PSO-LSSVM) was then adopted for abnormal data recovery and imputation. All above methods were tested with the daily water consumption monitoring data of water company. Results showed that the feature vector that contained exception data could be well preserved by moving average fitting and EEDM method and the effective reconstruction of water monitoring data was achieved, exhibiting better applicability than traditional statistical methods. Moreover, it can be observed that the PSO-LSSVM model had the ability to further improve the fitting results of the time-series data that excluded outliers. The fitted curve conformed to the seasonal fluctuation rule and it was consistent with the actual state of water demand. Accordingly, the objective of recovering the excluded data exception could be achieved reasonably by using this method. Furthermore, these methods can be applied to the analysis of monitoring data in other areas.