趙和松1，王圓圓2，孫愛民3

　　(1.水利部信息中心，北京 100053)

　　(2.北京金水信息技術發展有限公司，北京 100053)

　　(3.河海大學計算機與信息學院，江蘇 南京 211100)

　　摘　要：為提高取水預測數據的準確性，針對現有部分取水數據異常且難以進行人工判別的問題，提出一種基于ARIMA模型與3σ準則的取水異常檢測方法。分析每個取水點每年的日取水量的時間序列數據，使用時間序列的ARIMA模型和高斯分布的3σ準則判斷日取水量是否為異常值;通過時間序列分解算法分析異常值附近取水點的趨勢，判斷異常值附近是否存在其他未檢測出的異常值，給出異常值的參考修正值。對所提模型在帶異常標簽的通用時間序列數據集上進行實驗，通過評價指標混淆矩陣驗證模型可行性，并將模型在水利部門取水數據集上進行實驗，結果表明：模型可有效檢測取水數據中的異常值并修正其值，對取水異常的原因進行分析有助于改進取用水的采集方法，提高取水監測數據的質量。

　　關鍵詞： 取水異常檢測;機器學習;ARIMA模型;3σ準則;時間序列分解算法

　　A water intake anomaly detection method based on ARIMA model and 3σ criterion

　　ZHAO Hesong1，WANG Yuanyuan2，SUN Aimin3

　　(1.Information Center, Ministry of Water Resources，Beijing 100053，China;

　　2. Beijing Jinshui Information Technology Development Co.，Ltd.，Beijing 100053，China;

　　3. School of computer and information，Hohai University，Nanjing 211100，China)

　　Abstract: In order to improve the accuracy of water intake prediction data, a water intake anomaly detection method based on ARIMA model and 3σ criterion is proposed. Some existing daily water intake data is abnormal and difficult to be manually distinguished. This paper analyzes the time series data for each water intake point, and then applies the ARIMA model of time series and the 3σ criterion of Gaussian distribution to check the outliers. The decomposition algorithm is used to analyze the trend of time series near outliers, and distinguish undetected outliers and determine their values. Experiments on the proposed model on a universal time series dataset with anomalous labels are carried out. The feasibility of the model is verified through the evaluation index confusion matrix. The results show that the model can effectively detect the outliers in water intake data and provide reference values. The analysis of the causes of water intake anomalies is helpful to improve the monitoring of water intake and improve the data reliability.

　　Key words: water intake anomaly detection;machine learning;ARIMA model;3σ criteria; decomposition algorithm of time series