嵇曉燕,姚志鵬,楊 凱,陳亞男,王 正,安新國(guó)

基于MSLSTM-DA模型的水質(zhì)自動(dòng)監(jiān)測(cè)異常數(shù)據(jù)報(bào)警

嵇曉燕1*,姚志鵬1,楊 凱1,陳亞男1,王 正2,安新國(guó)2

(1.中國(guó)環(huán)境監(jiān)測(cè)總站,北京 100012；2.北京金水永利科技有限公司,北京 100012)

提出一種基于多元堆疊長(zhǎng)短時(shí)記憶網(wǎng)絡(luò)-差值分析(MSLSTM-DA)模型對(duì)地表水質(zhì)異常數(shù)據(jù)進(jìn)行報(bào)警的方法.該方法首先建立MSLSTM模型對(duì)水質(zhì)指標(biāo)數(shù)據(jù)進(jìn)行預(yù)測(cè),再基于預(yù)測(cè)結(jié)果的殘差分布建立DA模型,并確定各個(gè)指標(biāo)的數(shù)據(jù)異常閾值,當(dāng)實(shí)測(cè)數(shù)據(jù)與預(yù)測(cè)數(shù)據(jù)差值大于閾值時(shí)進(jìn)行數(shù)據(jù)報(bào)警.以長(zhǎng)江流域監(jiān)測(cè)斷面的水質(zhì)數(shù)據(jù)進(jìn)行了方法有效性驗(yàn)證.結(jié)果表明,構(gòu)建的預(yù)測(cè)模型對(duì)５個(gè)指標(biāo)的MAE、MAPE均值比BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)模型降低21.0%,17.8%,比LSTM模型降低16.8%,17.9%.皮爾遜系數(shù)均值比BP神經(jīng)網(wǎng)絡(luò)、LSTM模型的分別高5.9%,4.4%.5個(gè)指標(biāo)共檢出水質(zhì)異常數(shù)據(jù)37條,其中34條經(jīng)人工判斷確實(shí)存在有異常,報(bào)警準(zhǔn)確率高達(dá)91.9%.

堆疊長(zhǎng)短時(shí)記憶網(wǎng)絡(luò)；差值分析；水質(zhì)異常報(bào)警

地表水水質(zhì)自動(dòng)監(jiān)測(cè)站作為監(jiān)測(cè)地表水水質(zhì)現(xiàn)狀、及時(shí)預(yù)警潛在環(huán)境風(fēng)險(xiǎn)的重要基礎(chǔ),自1999年至今已建設(shè)了近3000個(gè)水質(zhì)自動(dòng)監(jiān)測(cè)站,形成了覆蓋我國(guó)十大流域和主要湖庫的水質(zhì)自動(dòng)監(jiān)測(cè)網(wǎng)絡(luò).自動(dòng)監(jiān)測(cè)站水質(zhì)數(shù)據(jù)作為衡量國(guó)家地表水水質(zhì)狀況的基礎(chǔ)數(shù)據(jù),實(shí)時(shí)發(fā)現(xiàn)異常數(shù)據(jù),從而預(yù)判水環(huán)境面臨的風(fēng)險(xiǎn)具有重要意義[1].存在異常的水質(zhì)數(shù)據(jù)很大程度上是由于環(huán)境污染造成的,因此實(shí)現(xiàn)水質(zhì)異常數(shù)據(jù)的報(bào)警對(duì)實(shí)現(xiàn)水環(huán)境早期污染的預(yù)警預(yù)報(bào)有重要的支撐作用[2-4].

針對(duì)水質(zhì)監(jiān)測(cè)中異常數(shù)據(jù)的預(yù)警問題,許多學(xué)者做了大量的研究,也提出了多種基于不同原理的方法. 基于歷史水質(zhì)各個(gè)指標(biāo)數(shù)據(jù)之間的關(guān)聯(lián)性進(jìn)行分析,如果當(dāng)前時(shí)刻數(shù)據(jù)不滿足相關(guān)性指標(biāo)則判定為異常[5].因?yàn)閿?shù)據(jù)相關(guān)性是基于多組數(shù)據(jù)進(jìn)行計(jì)算的,對(duì)評(píng)判單一時(shí)刻的數(shù)據(jù)效果不太好,不適用于水自動(dòng)站這種具有實(shí)時(shí)性要求的預(yù)警.再者就是利用統(tǒng)計(jì)學(xué)方法基于概率進(jìn)行異常的判斷[6],概率小的數(shù)據(jù)序列發(fā)生后則定義為異常數(shù)據(jù).而基于概率進(jìn)行數(shù)據(jù)異常判斷的難點(diǎn)在于無法科學(xué)的衡量異常數(shù)據(jù)出現(xiàn)的概率,預(yù)警準(zhǔn)確率難以保證.隨著機(jī)器學(xué)習(xí)技術(shù)的快速發(fā)展,借助其強(qiáng)大的數(shù)據(jù)分析能力,也衍生出很多異常數(shù)據(jù)檢測(cè)的方法[7-9].例如利用數(shù)據(jù)聚類的方法,對(duì)當(dāng)前指標(biāo)時(shí)間序列數(shù)據(jù)進(jìn)行是否為離群點(diǎn)的檢測(cè),如果是離群點(diǎn),則定義為異常數(shù)據(jù)[10].該方法適用于水質(zhì)數(shù)據(jù)突變的異常檢測(cè),而對(duì)于趨勢(shì)性的異常檢測(cè)效果不好.通過機(jī)器學(xué)習(xí)模型對(duì)未來水質(zhì)數(shù)據(jù)進(jìn)行預(yù)測(cè),然后根據(jù)監(jiān)測(cè)值和預(yù)測(cè)結(jié)果的差值評(píng)判數(shù)據(jù)是否異常,這種方法依賴于預(yù)測(cè)模型的效果.基于傳統(tǒng)的BP神經(jīng)網(wǎng)絡(luò)算法對(duì)歷史數(shù)據(jù)進(jìn)行訓(xùn)練,預(yù)測(cè)未來一段時(shí)間的水質(zhì)指標(biāo)數(shù)據(jù)值,當(dāng)實(shí)測(cè)數(shù)據(jù)和預(yù)測(cè)值偏差較大時(shí)定義為異常數(shù)據(jù)等[11].但是上述方法忽略了水質(zhì)監(jiān)測(cè)數(shù)據(jù)序存在的時(shí)序性,因此影響預(yù)測(cè)模型的精度.

自動(dòng)監(jiān)測(cè)站水質(zhì)監(jiān)測(cè)數(shù)據(jù)作為反映水環(huán)境狀況連續(xù)變化的數(shù)據(jù)序列,具有較強(qiáng)的周期性和自相關(guān)性,而且不同監(jiān)測(cè)指標(biāo)數(shù)據(jù)間也呈現(xiàn)很強(qiáng)的相關(guān)性.基于此,本文提出一種基于多元堆疊長(zhǎng)短時(shí)記憶網(wǎng)絡(luò)-差值分析(MSLSTM-DA)模型進(jìn)行水質(zhì)異常報(bào)警的方法,通過對(duì)監(jiān)測(cè)指標(biāo)歷史時(shí)間序列數(shù)據(jù)進(jìn)行學(xué)習(xí),然后對(duì)預(yù)測(cè)殘差進(jìn)行分析確定閾值,實(shí)現(xiàn)對(duì)水質(zhì)異常數(shù)據(jù)的報(bào)警.

1 模型建立

1.1 MSLSTM模型

圖1 MSLSTM模型結(jié)構(gòu)示意

長(zhǎng)短時(shí)記憶神經(jīng)網(wǎng)絡(luò)(LSTM)模型是一種具有時(shí)間循環(huán)結(jié)構(gòu)的神經(jīng)網(wǎng)絡(luò)模型,廣泛應(yīng)用于具有時(shí)間屬性的數(shù)據(jù)序列的預(yù)測(cè)分析.在地表水環(huán)境監(jiān)測(cè)中,同一斷面的水質(zhì)監(jiān)測(cè)數(shù)據(jù)具有時(shí)間維度上的因果聯(lián)系,適合采用長(zhǎng)短時(shí)記憶神經(jīng)網(wǎng)絡(luò)進(jìn)行預(yù)測(cè)分析[12-14].多元LSTM(MLSTM)是指模型輸出的變量個(gè)數(shù)多于一個(gè).對(duì)于自動(dòng)監(jiān)測(cè)站水質(zhì)數(shù)據(jù)而言,因?yàn)楸O(jiān)測(cè)指標(biāo)較多,如果僅對(duì)單一指標(biāo)進(jìn)行建模,一是模型無法獲取指標(biāo)數(shù)據(jù)間的關(guān)聯(lián)性;二是在實(shí)際的異常數(shù)據(jù)檢測(cè)應(yīng)用中,多個(gè)模型并行運(yùn)行會(huì)導(dǎo)致計(jì)算效率較低.堆疊LSTM(SLSTM)模型是具有多個(gè)隱藏LSTM層的模型,每個(gè)層包含多個(gè)存儲(chǔ)單元.該模型加大了模型深度,可以提取到水質(zhì)監(jiān)測(cè)數(shù)據(jù)序列中更深層次的特征,同時(shí)也增強(qiáng)了模型的預(yù)測(cè)能力[15-17].

本文采用的MSLSTM模型中包括4個(gè)LSTM層,其中添加BatchNorm層可以加速模型的訓(xùn)練, Dropout層有助于提高模型的泛化能力[18].模型輸出的變量包括溶解氧(DO)、高錳酸鹽指數(shù)(IMn)、氨氮(NH3-N)、總磷(TP)以及總氮(TN)等5個(gè)指標(biāo).相應(yīng)的模型結(jié)構(gòu)見圖1.

1.2 DA模型

考慮到地表水環(huán)境的復(fù)雜性,影響水質(zhì)的因素較多,因此水質(zhì)預(yù)測(cè)數(shù)據(jù)與實(shí)測(cè)數(shù)據(jù)會(huì)存在合理偏差.差值分析(DA)模型是對(duì)水質(zhì)指標(biāo)的實(shí)際監(jiān)測(cè)值與模型預(yù)測(cè)值的差值序列進(jìn)行分析.當(dāng)預(yù)測(cè)模型達(dá)到一定的預(yù)測(cè)精度后,模型的輸出與實(shí)際值的差值序列應(yīng)該是在0值上下波動(dòng),并關(guān)于0值呈現(xiàn)近似對(duì)稱分布.基于此本文構(gòu)建了一種基于迭代法計(jì)算差值閾值的流程.首先計(jì)算監(jiān)測(cè)值和模型預(yù)測(cè)中差值的絕對(duì)值序列D,初始上限值U設(shè)置為0,置信水平設(shè)置為0.9,計(jì)算序列D中不大于上限值的占比,當(dāng)該占比大于0.9時(shí),則得到閾值;如果小于等于0.9,則增加初始上限值,直到得到該預(yù)測(cè)模型差值的閾值.DA模型結(jié)構(gòu)見圖2.

圖2 DA模型結(jié)構(gòu)

2 異常數(shù)據(jù)檢測(cè)方法

基于MSLSTM-DA模型進(jìn)行水質(zhì)異常數(shù)據(jù)報(bào)警的流程包括構(gòu)建模型數(shù)據(jù)集、訓(xùn)練MSLSTM模型、建立DA模型、模型效果評(píng)估以及異常數(shù)據(jù)實(shí)時(shí)報(bào)警.

2.1 構(gòu)建模型數(shù)據(jù)集

利用臨近值填充法對(duì)水質(zhì)監(jiān)測(cè)指標(biāo)歷史數(shù)據(jù)集中缺失的數(shù)據(jù)進(jìn)行填充,形成預(yù)處理數(shù)據(jù)集.并根據(jù)監(jiān)測(cè)數(shù)據(jù)的時(shí)間順序以7:2:1的比例將預(yù)處理數(shù)據(jù)集分割為訓(xùn)練、驗(yàn)證與測(cè)試數(shù)據(jù)集,并根據(jù)訓(xùn)練數(shù)據(jù)集中各個(gè)指標(biāo)的最大、最小值對(duì)這3個(gè)數(shù)據(jù)集的對(duì)應(yīng)指標(biāo)進(jìn)行Min-Max歸一化,將數(shù)據(jù)轉(zhuǎn)換到01之間.

2.2 訓(xùn)練MSLSTM模型

根據(jù)訓(xùn)練數(shù)據(jù)集對(duì)MSLSTM模型進(jìn)行訓(xùn)練,訓(xùn)練迭代次數(shù)設(shè)置為10000次,神經(jīng)網(wǎng)絡(luò)優(yōu)化器選擇Adam,模型最后一個(gè)全連接層的激活函數(shù)選擇sigmoid.通過實(shí)時(shí)判斷驗(yàn)證數(shù)據(jù)集的預(yù)測(cè)效果提前結(jié)束訓(xùn)練,防止過擬合.在模型訓(xùn)練完成后,分別將訓(xùn)練、驗(yàn)證數(shù)據(jù)集作為模型的輸入數(shù)據(jù)得到相應(yīng)的模型預(yù)測(cè)值.

模型訓(xùn)練完成后,需要對(duì)模型預(yù)測(cè)的精確度進(jìn)行評(píng)估.本文采用平均絕對(duì)誤差(MAE)、平均絕對(duì)百分比誤差(MAPE)、皮爾遜相關(guān)系數(shù)對(duì)模型進(jìn)行評(píng)價(jià).

2.3 建立DA模型

計(jì)算水質(zhì)監(jiān)測(cè)指標(biāo)監(jiān)測(cè)值和模型預(yù)測(cè)值之間的差值,并對(duì)差值序列進(jìn)行關(guān)于0值對(duì)稱分布檢驗(yàn).如果通過檢驗(yàn),則建立DA模型,根據(jù)置信水平得到水質(zhì)異常區(qū)間以及水質(zhì)異常閾值.否則更改模型結(jié)構(gòu)和參數(shù)再次進(jìn)行模型的訓(xùn)練.

2.4 模型效果評(píng)估

針對(duì)預(yù)測(cè)數(shù)據(jù)集計(jì)算模型預(yù)測(cè)值,根據(jù)預(yù)測(cè)值和監(jiān)測(cè)值之間的差值以及得到的水質(zhì)異常數(shù)據(jù)閾值,進(jìn)行異常數(shù)據(jù)的檢測(cè),當(dāng)檢測(cè)出來的數(shù)據(jù)的確為異常數(shù)據(jù)時(shí),說明模型可以應(yīng)用于實(shí)際的水質(zhì)異常數(shù)據(jù)檢測(cè).否則的話則重新設(shè)置差值分析模型的置信水平.

2.5 異常數(shù)據(jù)實(shí)時(shí)報(bào)警

獲取當(dāng)前時(shí)刻水質(zhì)自動(dòng)監(jiān)測(cè)數(shù)據(jù)的實(shí)時(shí)監(jiān)測(cè)值,并計(jì)算模型對(duì)該時(shí)刻水質(zhì)的預(yù)測(cè)值,當(dāng)實(shí)測(cè)值和模型預(yù)測(cè)值得差值超過異常閾值時(shí),確認(rèn)當(dāng)前時(shí)刻水質(zhì)監(jiān)測(cè)值為異常數(shù)據(jù)并報(bào)警,否則為正常數(shù)據(jù).

圖3 MSLSTM-DA模型水質(zhì)異常數(shù)據(jù)檢測(cè)流程

3 實(shí)例驗(yàn)證

3.1 數(shù)據(jù)集說明

選擇位于長(zhǎng)江流域某一斷面2019-01-01至2020-12-31的水質(zhì)自動(dòng)站監(jiān)測(cè)數(shù)據(jù),共4386條.因?yàn)镈O的監(jiān)測(cè)頻率為1次/h,而其他指標(biāo)為1次/4h,構(gòu)建模型數(shù)據(jù)集需要統(tǒng)一時(shí)間粒度,因此DO選擇和其他指標(biāo)相同的監(jiān)測(cè)時(shí)刻數(shù)據(jù)(表1).

表1 數(shù)據(jù)條數(shù)表

水質(zhì)監(jiān)測(cè)數(shù)據(jù)具有時(shí)間屬性,前后時(shí)刻的數(shù)據(jù)呈現(xiàn)很強(qiáng)的相關(guān)性,因此選擇利用臨近值填充的方法對(duì)缺失的指標(biāo)數(shù)據(jù)填充,不會(huì)影響水質(zhì)數(shù)據(jù)序列的時(shí)序性.按照時(shí)間順序?qū)㈩A(yù)處理數(shù)據(jù)集按照7:2:1的比例分割為訓(xùn)練、驗(yàn)證、測(cè)試數(shù)據(jù)集(表2).

表2 模型數(shù)據(jù)集

當(dāng)前監(jiān)測(cè)時(shí)刻以及前11個(gè)監(jiān)測(cè)時(shí)刻的5指標(biāo)數(shù)據(jù)作為模型的輸入數(shù)據(jù).當(dāng)LSTM這類神經(jīng)網(wǎng)絡(luò)在學(xué)習(xí)時(shí)間序列數(shù)據(jù)的規(guī)律時(shí),因?yàn)闀r(shí)間序列存在自相關(guān)性,神經(jīng)網(wǎng)絡(luò)會(huì)將前一時(shí)刻的數(shù)據(jù)當(dāng)作當(dāng)前時(shí)刻的預(yù)測(cè)值,導(dǎo)致最后的預(yù)測(cè)序列出現(xiàn)時(shí)間延遲現(xiàn)象.因此本文將當(dāng)前時(shí)刻水質(zhì)數(shù)據(jù)減去前一個(gè)監(jiān)測(cè)時(shí)刻的5指標(biāo)水質(zhì)數(shù)據(jù)的差值數(shù)據(jù)作為模型的輸出數(shù)據(jù).

3.2 MSLSTM模型結(jié)果分析

模型訓(xùn)練完成后,對(duì)訓(xùn)練、驗(yàn)證和測(cè)試數(shù)據(jù)集進(jìn)行預(yù)測(cè)結(jié)果的評(píng)估.精確度評(píng)估表見表3.

表3 MSLSTM模型精確度評(píng)估表

表4 三種模型精確度評(píng)估表

由精確度評(píng)估結(jié)果(表3)可以看出,指標(biāo)MAE對(duì)于各指標(biāo)的數(shù)據(jù)范圍而言,都比較低;指標(biāo)的MAPE均在20%以下,說明數(shù)據(jù)預(yù)測(cè)的精確度達(dá)到了80%以上;皮爾遜相關(guān)系數(shù)均大于0.8,說明預(yù)測(cè)結(jié)果序列和監(jiān)測(cè)值序列的線性趨勢(shì)關(guān)系較強(qiáng).從測(cè)試數(shù)據(jù)集的模型預(yù)測(cè)值和監(jiān)測(cè)值的對(duì)比曲線(圖4)可以看出,各個(gè)指標(biāo)的預(yù)測(cè)值和監(jiān)測(cè)值的曲線貼合度較高.為進(jìn)一步說明本文構(gòu)建的預(yù)測(cè)模型的有效性,選擇BP神經(jīng)網(wǎng)絡(luò)[19]、LSTM模型[20]作為對(duì)照模型.

由三種模型在測(cè)試數(shù)據(jù)集上的對(duì)比結(jié)果(表4)可以看出,本文構(gòu)建的預(yù)測(cè)模型的5個(gè)指標(biāo)的MAE、MAPE均值比BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)模型降低21.0%,17.8%,比LSTM模型降低16.8%,17.9%.皮爾遜系數(shù)均值比BP神經(jīng)網(wǎng)絡(luò)、LSTM模型的高5.9%,4.4%,表明本文構(gòu)建的預(yù)測(cè)模型對(duì)各個(gè)指標(biāo)的預(yù)測(cè)精確性比較高.

3.3 建立DA模型

計(jì)算驗(yàn)證數(shù)據(jù)集的預(yù)測(cè)值,并繪制預(yù)測(cè)值與指標(biāo)監(jiān)測(cè)值差值的直方圖,并進(jìn)行閾值計(jì)算.

圖5 5個(gè)水質(zhì)監(jiān)測(cè)指標(biāo)的差值分布

Fig.5 Difference distribution charts

從圖5可以看出,5個(gè)指標(biāo)的差值序列主要集中在0值附近,并關(guān)于0值近似呈現(xiàn)對(duì)稱分布.

3.4 異常數(shù)據(jù)報(bào)警驗(yàn)證

基于DA模型得到的數(shù)據(jù)異常閾值,對(duì)測(cè)試數(shù)據(jù)集進(jìn)行異常數(shù)據(jù)檢測(cè).5指標(biāo)共檢測(cè)異常數(shù)據(jù)37條,其中IMn6條,NH3-N12條,TP8條,TN 11條,DO并未發(fā)現(xiàn)異常數(shù)據(jù).異常數(shù)據(jù)見圖6.

對(duì)檢測(cè)出的異常數(shù)據(jù)進(jìn)行人工核查,發(fā)現(xiàn)34條數(shù)據(jù)存在異常的可能性較大,異常數(shù)據(jù)檢出正確率達(dá)到91.9%,并且不存在異常數(shù)據(jù)的遺漏情況.說明此方法可以應(yīng)用于水質(zhì)自動(dòng)監(jiān)測(cè)數(shù)據(jù)管理的異常數(shù)據(jù)報(bào)警.

4 結(jié)論

4.1 本文提出的多元堆疊長(zhǎng)短時(shí)記憶網(wǎng)絡(luò)模型,可消除異常值以及缺失值對(duì)預(yù)測(cè)模型的影響,可實(shí)現(xiàn)地表水水質(zhì)數(shù)據(jù)的預(yù)測(cè),精確度達(dá)到了80%以上,為進(jìn)一步提升水環(huán)境質(zhì)量提供了依據(jù).

4.2 通過預(yù)測(cè)數(shù)據(jù)的殘差分布建立差值分析模型,得到判定水質(zhì)數(shù)據(jù)異常的閾值.基于該閾值以及預(yù)測(cè)數(shù)據(jù)對(duì)當(dāng)前時(shí)刻水質(zhì)監(jiān)測(cè)數(shù)據(jù)進(jìn)行異常判定,預(yù)警準(zhǔn)確率可達(dá)90%以上.

[1] 邢夢(mèng)林,王瀟磊,鄭 瑤,等.淺談水質(zhì)自動(dòng)監(jiān)測(cè)系統(tǒng)的建設(shè)及應(yīng)用研究 [J]. 環(huán)境科學(xué)與管理, 2013,(10):40-42.

Xing M L, Wang X L, Zheng Y, et al. Discussion on construction and application of water quality automatic monitoring system [J]. Environmental Science and Management, 2013,(10):40-42.

[2] 嵇曉燕,劉廷良,孫宗光,等.國(guó)家水環(huán)境質(zhì)量監(jiān)測(cè)網(wǎng)絡(luò)發(fā)展歷程與展望 [J]. 環(huán)境監(jiān)測(cè)管理與技術(shù), 2014,26(6):1-4,8.

Ji X Y, Liu T L, Sun Z G. The Development of water environmental quality monitoring network in China [J]. The Administration and Technique of Environmental Monitoring, 2014,26(6):1-4,8.

[3] 計(jì) 紅,韓龍喜,劉軍英,等.水質(zhì)預(yù)警研究發(fā)展探討 [J]. 水資源保護(hù), 2011,27(5):39-42.

Ji H, Han L X, Liu J Y, et al. Study on water quality early warning [J]. Water Resources Protection, 2011,27(5):39-42.

[4] 王文寶,曹 騫.水質(zhì)自動(dòng)監(jiān)測(cè)站的運(yùn)行管理與水質(zhì)預(yù)警 [J]. 環(huán)境監(jiān)控與預(yù)警, 2010,2(1):54-56.

Wang W B, Cao Q. Operation of automated water quality monitoring stations and water quality prewarning [J]. Environmental Monitoring and Forewarning, 2010,2(1):54-56.

[5] 周 純,張 苒.水質(zhì)自動(dòng)監(jiān)測(cè)異常數(shù)據(jù)的分析與利用[J]. 新疆環(huán)境保護(hù), 2012,34(4):42-44.

Zhou C, Zhang R. Analysis and application of water quality outlier automatic monitoring [J]. Environmental Protection of Xinjiang, 2012,34(4):42-44.

[6] 王 威,沈興厚,張利亞.皮爾遜Ⅲ型頻率曲線在水源地水質(zhì)預(yù)警中的應(yīng)用 [J]. 河南科學(xué), 2018,36(7):1042-1048.

Wang W, Shen X H, Zhang L Y. Application of pearsonⅢ frequency curve on water quality warning of water source [J]. Henan Science, 2018,36(7):1042-1048.

[7] Huang X. Abnormal data detection technology based on improved neural network [J]. International Journal of Applied Mathematics and Soft Computing, 2018,7(1):89-94.

[8] Liu F, Su W X, Zhao J J, et al. Abnormal data detection method based on fuzzy autoregressive hidden markov Model [J]. Journal of Residuals Science & Technology, 2016,13(8):38-45.

[9] 李若楠,王 琦,劉書明.基于典型相關(guān)系數(shù)和隨機(jī)森林的水質(zhì)預(yù)警方法 [J]. 中國(guó)環(huán)境科學(xué), 2021,41(9):4457-4464.

Li R N, Wang Q, Liu S M. Water quality warning method based on canonical correlation coefficient and random forest [J]. China Environmental Science, 2021,41(9):4457-4464.

[10] 梅 林,張鳳荔,高 強(qiáng).離群點(diǎn)檢測(cè)技術(shù)綜述 [J]. 計(jì)算機(jī)應(yīng)用研究, 2020,37(12):3521-3527.

Mei L, Zhnag F L, Gao Q, et al. Overview of outlier detection technology [J]. Application Research of Computers, 2020,37(12): 3521-3527.

[11] 史 斌,姜繼平,王 鵬.基于高頻在線水質(zhì)數(shù)據(jù)異常的突發(fā)污染預(yù)警 [J]. 中國(guó)環(huán)境科學(xué), 2017,37(11):4394-4400.

Shi B, Jiang J P, Wang P. Early warning of water pollution incidents based on abnormal change of water quality data from high frequency online monitoring [J]. China Environmental Science, 2017,37(11): 4394-4400.

[12] Palani S, Liong S Y, Tkalich P. An ANN application for water quality forecasting [J]. Marine Pollution Bulletin, 2008,(56):1586-1597.

[13] Liu J J, Zhuang H, Tie Z X, et al. Water quality multi-factor prediction model using LSTM neural network based on K-similarity noise reduction [J]. Computer Systems Applications, 2019,28(2):226-232.

[14] Li X, Peng L, Yao X J, et al. Long short-term memory neural network for air pollutant concentration predictions: Method development and evaluation [J]. Environmental Pollution, 2017,231:997-1004.

[15] 王 軍,高梓勛,朱永明.基于CNN-LSTM模型的黃河水質(zhì)預(yù)測(cè)研究 [J]. 人民黃河, 2021,43(5):86-99.

Wang J, Gao Z X, Zhu Y M. Research on Yellow River water quality prediction based on CNN-LSTM Model [J]. Yellow River, 2021, 43(5):86-99.

[16] 劉建華,李錦程,楊龍?jiān)?等.基于EMD-SLSTM的家庭短期負(fù)荷預(yù)測(cè) [J]. 電力系統(tǒng)保護(hù)與控制, 2019,47(6):40-47.

Liu J H, Li J C, Yang L Y, et al. Short-term household load forecasting based on EMD-SLSTM [J]. Power System Protection and Control, 2019,47(6):40-47.

[17] Wu Q F, Meng X Y, Liu M Y. Prediction of rice yield via stacked LSTM [J]. International Journal of Agricultural and Environmental Information Systems (IJAEIS), 2020,11(1):86-95.

[18] 楊 青,王晨蔚.基于深度學(xué)習(xí)LSTM神經(jīng)網(wǎng)絡(luò)的全球股票指數(shù)預(yù)測(cè)研究 [J]. 統(tǒng)計(jì)研究, 2019,36(3):65-77.

Yang Q, Wang C W. A study on forecast of global stock indices based on deep LSTM neural network [J]. Statistical Research, 2019,36(3): 65-77.

[19] 張秀菊,王柳林,李秀平,等.基于BP神經(jīng)網(wǎng)絡(luò)的瀟河流域水質(zhì)預(yù)測(cè) [J]. 水資源與水工程學(xué)報(bào), 2021,32(5):19-26.

Zhang X J, Wang L L, Li X P, et al. Water quality prediction of the Xiaohe River Basin based on BP neural network model [J]. Journal of Water Resources and Water Engineering, 2021,32(5):19-26.

[20] 張貽婷,李天宏.基于長(zhǎng)短時(shí)記憶神經(jīng)網(wǎng)絡(luò)的河流水質(zhì)預(yù)測(cè)研究 [J]. 環(huán)境科學(xué)與技術(shù), 2021,44(8):163-169.

Zhang Y T, Li T H. River water quality prediction based on long short-term memory neural network [J]. Environmental Science and Technology, 2021,44(8):163-169.

Water quality alert with automatic monitoring data based on MSLSTM-DA model.

JI Xiao-yan1*, YAO Zhi-peng1, YANG Kai1, CHEN Ya-nan1, WANG Zheng2, AN Xin-guo2

(1.China National Environmental Monitoring Center, Beijing 100012, China；2.Golden Water Technology (Beijing) Ltd, Beijing 100012, China)., 2022,42(4)：1877～1883

A multivariate stacked long and short term memory network-difference analysis (MSLSTM-DA) model is proposed to alarm surface water quality abnormal data. Established the MSLSTM model to predict the water quality data, and then established the DA model based on the residual distribution of the prediction results to determine the threshold value of each indicator, and alerted the data when the difference between the measured data and the predicted data is greater than the threshold value. The validity of the method was verified using water quality data from the Yangtze River basin monitoring sections. The results showed that the mean values of MAE and MAPE for five indicators were 21.0% and 17.8% lower than those of BP neural network prediction model, and 16.8% and 17.9% lower than those of LSTM model. The mean value of Pearson coefficient was 5.9% and 4.4% higher than that of BP neural network and LSTM model. 37 abnormal water quality data were detected for the 5 indicators, 34 of which were judged to be abnormal by manual judgment, with an alarm accuracy rate of 91.9%.

stacked long and short-term memory network；difference analysis；water quality alert

X832

A

1000-6923(2022)04-1877-07

嵇曉燕(1981-),女,江蘇淮安人,正高級(jí)工程師,博士,主要研究方向?yàn)樗h(huán)境質(zhì)量監(jiān)測(cè)評(píng)價(jià).發(fā)表論文30余篇.

2021-09-18

長(zhǎng)江生態(tài)環(huán)境保護(hù)修復(fù)聯(lián)合研究項(xiàng)目(2019-LHYJ-01-0301);國(guó)家水環(huán)境監(jiān)測(cè)監(jiān)控及業(yè)務(wù)化平臺(tái)技術(shù)研究課題(2017ZX07302002)

*責(zé)任作者, 正高級(jí)工程師, jixy@cnemc.cn