康俊鋒,譚建林,方 雷,肖亞來(lái)

XGBoost-LSTM變權(quán)組合模型支持下短期PM2.5濃度預(yù)測(cè)——以上海為例

康俊鋒1,譚建林1,方 雷2*,肖亞來(lái)1

(1.江西理工大學(xué)土木與測(cè)繪工程學(xué)院,江西 贛州 341000；2.復(fù)旦大學(xué)環(huán)境科學(xué)與工程系,上海 200433)

為進(jìn)一步提高PM2.5濃度預(yù)測(cè)的精度,基于XGBoost和LSTM進(jìn)行改進(jìn)得到變權(quán)組合模型XGBoost-LSTM(Variable).過(guò)對(duì)預(yù)測(cè)因子進(jìn)行相關(guān)性分析,得到其它大氣污染物和氣象因素對(duì)PM2.5濃度的影響,確定最優(yōu)PM2.5濃度預(yù)測(cè)因子,再將預(yù)處理后數(shù)據(jù)集輸入LSTM模型和XGBoost模型分別進(jìn)行預(yù)測(cè),采用基于殘差改進(jìn)的自適應(yīng)變權(quán)組合方法得到最終預(yù)測(cè)結(jié)果.結(jié)果表明,污染物變量的相對(duì)重要性高于氣象因子變量,其中當(dāng)前PM2.5和CO濃度的相對(duì)重要性較高,而平均風(fēng)速和相對(duì)濕度重要性較低.XGBoost-LSTM(Variable)模型的RMSE、MAE和MAPE值為1.75、1.12和6.06,優(yōu)于LSTM、XGBoost、SVR、XGBoost-LSTM(Equal)和XGBoost-LSTM(Residual)模型.分季節(jié)預(yù)測(cè)結(jié)果表明,XGBoost-LSTM(Variable)模型在春季預(yù)測(cè)精度最好,而夏季預(yù)測(cè)精度較差.模型預(yù)測(cè)精度高的原因在于其不僅考慮了數(shù)據(jù)的時(shí)間序列特征,又兼顧了數(shù)據(jù)的非線性特征.

LSTM；XGBoost；PM2.5預(yù)測(cè)；變權(quán)組合

社會(huì)經(jīng)濟(jì)的快速發(fā)展導(dǎo)致PM2.5等空氣污染問(wèn)題日益突出[1-2],對(duì)PM2.5等空氣污染物濃度進(jìn)行精準(zhǔn)預(yù)測(cè)和提前預(yù)警具有重要意義.PM2.5濃度預(yù)測(cè)模型主要包括以CAMQ[3](通用多尺度空氣質(zhì)量模型)模式、WRF-Chem[4](區(qū)域大氣動(dòng)力-耦合模型)模式和NAQPMS[5](嵌套空氣質(zhì)量預(yù)報(bào)模式系統(tǒng))模式等為代表的機(jī)理模型,以多元統(tǒng)計(jì)理論、灰色預(yù)測(cè)模型(G,M)[6-7]、多元線性回歸模型[8]等為代表的統(tǒng)計(jì)預(yù)報(bào)模型,以及以徑向基神經(jīng)網(wǎng)絡(luò)(RBF)、反向傳播神經(jīng)網(wǎng)絡(luò)(BP)、支持向量機(jī)(SVM)等神經(jīng)網(wǎng)絡(luò)發(fā)展到基于深度學(xué)習(xí)模型的神經(jīng)網(wǎng)絡(luò),如:基于深度信念網(wǎng)絡(luò)(DBNs)、長(zhǎng)短期記憶神經(jīng)網(wǎng)絡(luò)(LSTM)等[9-12].

隨著機(jī)器學(xué)習(xí)技術(shù)的發(fā)展,有研究采用歷史氣象數(shù)據(jù)或歷史污染數(shù)據(jù),利用支持向量回歸模型[13]、隨機(jī)森林[14-16]、BP人工神經(jīng)網(wǎng)絡(luò)[17]以及LSTM網(wǎng)絡(luò)[18]等單機(jī)器學(xué)習(xí)模型,預(yù)測(cè)實(shí)時(shí)PM2.5濃度[19]、未來(lái)短期[20-21]和長(zhǎng)期PM2.5濃度[14,22-23]及PM2.5濃度的空間變異[17]等.有研究通過(guò)構(gòu)建多個(gè)單機(jī)器學(xué)習(xí)模型進(jìn)行PM2.5濃度預(yù)測(cè)比較,LSTM網(wǎng)絡(luò)在處理非線性時(shí)序數(shù)據(jù)方面性能高效并且有更好的泛化能力[24],XGBoost模型預(yù)測(cè)精度優(yōu)于其他單機(jī)器學(xué)習(xí)模型[25].為進(jìn)一步提高PM2.5濃度預(yù)測(cè)精度,有學(xué)者開(kāi)始嘗試組合多個(gè)機(jī)器學(xué)習(xí)模型來(lái)預(yù)測(cè)PM2.5濃度.宋國(guó)君等[26]和李建更等[27]分別建立了基于時(shí)間序列分解的SVR組合預(yù)測(cè)模型、Liu等[28]構(gòu)建了DBN、LSTM網(wǎng)絡(luò)和多層神經(jīng)網(wǎng)絡(luò)(MLP)的三模型組合模型.雖然組合預(yù)測(cè)模型相較于單機(jī)器學(xué)習(xí)模型可以提升和改善模型預(yù)測(cè)精度[29],但已有的組合模型研究都只是簡(jiǎn)單的將一個(gè)模型預(yù)測(cè)結(jié)果輸入另一模型進(jìn)行二次預(yù)測(cè),或者將多個(gè)模型的預(yù)測(cè)結(jié)果進(jìn)行簡(jiǎn)單求和.其特點(diǎn)類似一種“機(jī)械組合”,兩種或多種組合模型之間未發(fā)生真正的“化學(xué)反應(yīng)”.

此外,由于PM2.5濃度變化既受氣象因素影響,也受空氣污染物影響[30-31],但已有基于機(jī)器學(xué)習(xí)的PM2.5濃度變化預(yù)測(cè)研究大都只采用氣象數(shù)據(jù),或者只采用污染物濃度歷史數(shù)據(jù)來(lái)進(jìn)行PM2.5濃度預(yù)測(cè),預(yù)測(cè)精度受限.因此,本研究嘗試將氣象數(shù)據(jù)、空氣污染物數(shù)據(jù)和PM2.5濃度歷史數(shù)據(jù)結(jié)合,在分析空氣污染物和氣象因素對(duì)PM2.5濃度影響基礎(chǔ)上,設(shè)計(jì)了一種基于殘差賦權(quán)[32]改進(jìn)的自適應(yīng)賦權(quán)方法,構(gòu)建XGBoost模型和LSTM網(wǎng)絡(luò)變權(quán)組合模型,對(duì)未來(lái)1h短期PM2.5濃度進(jìn)行預(yù)測(cè),以期為環(huán)境監(jiān)測(cè)部門及社會(huì)公眾提供預(yù)警及精準(zhǔn)預(yù)測(cè).

1 研究材料與方法

1.1 研究區(qū)域與數(shù)據(jù)

上海市(30°40′～31°53′N,120°52′～122°12′E)位于中國(guó)東部沿海的長(zhǎng)江三角洲地區(qū),是典型的特大型城市,面積約6340km2,地形起伏小,屬于亞熱帶季風(fēng)氣候,其空氣質(zhì)量一直引人關(guān)注.本研究選取上海市10個(gè)環(huán)境監(jiān)測(cè)站點(diǎn)(圖1)2017年1月1日～10月31日逐小時(shí)歷史空氣質(zhì)量濃度數(shù)據(jù)和氣象數(shù)據(jù)(共7297組)數(shù)據(jù),其統(tǒng)計(jì)性描述如表1所示.其中,近地面PM2.5濃度等空氣質(zhì)量數(shù)據(jù)來(lái)自于生態(tài)環(huán)境部部空氣質(zhì)量實(shí)時(shí)發(fā)布系統(tǒng)(http://106.37.208.233: 20035/),氣象數(shù)據(jù)來(lái)自于歐洲中期天氣預(yù)報(bào)中心3km′3km再分析數(shù)據(jù)(https://www.ecmwf.int/).

圖1 研究區(qū)域

表1 上海市空氣質(zhì)量數(shù)據(jù)和氣象數(shù)據(jù)統(tǒng)計(jì)性描述

1.2 研究方法

1.2.1 XGBoost XGBoost(Extreme Gradient Boosting)是一種集成的樹(shù)模型,是GBDT(Gradient Boosting Decision Tree)的改進(jìn)boosting算法,具有訓(xùn)練速度快、預(yù)測(cè)精度高等優(yōu)點(diǎn)[33].XGBoost集成了多棵分類回歸樹(shù)(CART)以彌補(bǔ)單棵CART無(wú)法滿足預(yù)測(cè)精度的不足,預(yù)測(cè)結(jié)果等于所有CART的得分總和[34].模型表示為:

XGBoost通過(guò)對(duì)代價(jià)函數(shù)進(jìn)行二階泰勒展開(kāi),使用一階和二階導(dǎo)數(shù),在訓(xùn)練集上可以更快收斂,有效提高訓(xùn)練速度,并且將正則化項(xiàng)加到損失函數(shù)上,可以降低模型的復(fù)雜度和過(guò)擬合的風(fēng)險(xiǎn).

1.2.2 LSTM LSTM(Long Short-Term Memory)是RNN(Recurrent Neural Network)的改進(jìn)模式,由Hpchreiter等[35]在1997年提出,采用LSTM層替換了傳統(tǒng)的隱藏層,通過(guò)引入輸入門、輸出門、遺忘門三種“門”結(jié)構(gòu)實(shí)現(xiàn)信息的有效篩選和長(zhǎng)期記憶.LSTM內(nèi)部結(jié)構(gòu)如圖2所示:

圖2 LSTM模型結(jié)構(gòu)

計(jì)算公式如下:

1.2.3 變權(quán)組合預(yù)測(cè)模型 本文構(gòu)建三個(gè)組合模型:XGBoost和LSTM等值賦權(quán)組合模型XGBoost-LSTM(Equal)、XGBoost和LSTM殘差賦權(quán)組合模型XGBoost-LSTM(Residual)以及XGBoost和LSTM變權(quán)組合模型XGBoost-LSTM (Variable).

(1) 單機(jī)器學(xué)習(xí)模型構(gòu)建單機(jī)器學(xué)習(xí)模型的優(yōu)劣決定組合模型的預(yù)測(cè)精度和性能,設(shè)置合理有效的超參數(shù)對(duì)于提高組合模型的預(yù)測(cè)性能和收斂速度具有重要意義[36].基于前人研究模型參數(shù)設(shè)置[37-38]對(duì)LSTM網(wǎng)絡(luò)超參數(shù)進(jìn)行設(shè)置,最終模型網(wǎng)絡(luò)層數(shù)為2,學(xué)習(xí)率設(shè)置為0.001,激活函數(shù)設(shè)置為Tanh,優(yōu)化算法選用Adam算法,迭代訓(xùn)練次數(shù)設(shè)置為100次,并設(shè)置學(xué)習(xí)率衰減為50次削弱為10%.

利用Scikit-learn提供的網(wǎng)格搜索(GridSearch)方法[39]對(duì)XGBoost模型的超參數(shù)尋優(yōu),模型參數(shù)最終設(shè)置為:max_depth=4,learning_rate=0.1,n_ estimators=200,subsample=0.7,colsample_bytree=0.85, silent=True,ganma=0.2.

(2)組合模型賦權(quán)組合模型精度與單機(jī)器學(xué)習(xí)模型的賦權(quán)有直接關(guān)系,賦權(quán)方法常見(jiàn)的有固定賦權(quán)與自適應(yīng)賦權(quán),其中固定賦權(quán)以等值賦權(quán)和殘差賦權(quán)法最為常見(jiàn)[40].

等值賦權(quán)將單模型賦予相同的權(quán)重,而殘差賦權(quán)組合模型表達(dá)為:

(3)改進(jìn)的變權(quán)組合模型賦權(quán)方法本文使用基于殘差賦權(quán)改進(jìn)的自適應(yīng)賦權(quán)方法的變權(quán)[41]方法構(gòu)建了XGBoost-LSTM(Variable)模型.對(duì)于單機(jī)器學(xué)習(xí)模型在基于式(9)得到所有時(shí)刻殘差賦權(quán)的權(quán)重基礎(chǔ)上改進(jìn),計(jì)算最優(yōu)值,使用前時(shí)刻權(quán)重平均值對(duì)本時(shí)刻模型進(jìn)行初始賦權(quán),即:

對(duì)于時(shí)刻,假設(shè)基于式(9)和式(11)得到各單機(jī)器學(xué)習(xí)模型權(quán)重后,計(jì)算該時(shí)刻組合模型的預(yù)測(cè)值與真實(shí)值的誤差絕對(duì)值分別為e,t、e,t,則有:

比較e,t和e,t值的大小,如果e,t<e,t則該組合模型用新的權(quán)重()代替原來(lái)的權(quán)重(),否則模型權(quán)重保持不變.

1.2.4 組合預(yù)測(cè)模型構(gòu)建流程 組合預(yù)測(cè)模型構(gòu)建流程如圖3所示,包括數(shù)據(jù)預(yù)處理、單機(jī)器學(xué)習(xí)模型和變權(quán)組合預(yù)測(cè)模型構(gòu)建以及模型評(píng)價(jià)分析.

圖3 XGBoost-LSTM(Variable)模型預(yù)測(cè)流程

(1)數(shù)據(jù)預(yù)處理得到的原始數(shù)據(jù)集進(jìn)行預(yù)處理,主要包括數(shù)據(jù)清洗、缺失值填充和歸一化處理,本研究缺失值采用缺失前后數(shù)據(jù)均值補(bǔ)充.

(2)單機(jī)器學(xué)習(xí)模型構(gòu)建數(shù)據(jù)集按照訓(xùn)練集:測(cè)試集=9:1比例劃分后,在訓(xùn)練集上分別訓(xùn)練LSTM網(wǎng)絡(luò)和XGBoost模型,確定模型最優(yōu)超參數(shù),保存訓(xùn)練模型.將測(cè)試集分別輸入模型,得到各單機(jī)器學(xué)習(xí)模型預(yù)測(cè)結(jié)果.

(3)變權(quán)組合預(yù)測(cè)模型構(gòu)建采用前文所示賦權(quán)方法確定各單機(jī)器學(xué)習(xí)模型的權(quán)重,計(jì)算得到組合模型最終預(yù)測(cè)結(jié)果.

(4)模型評(píng)價(jià)分析根據(jù)模型評(píng)價(jià)指標(biāo)比較模型預(yù)測(cè)能力,分析模型預(yù)測(cè)效果.

1.2.5 評(píng)價(jià)指標(biāo) 本研究采用常見(jiàn)的評(píng)價(jià)指標(biāo)均方根誤差(RMSE),平均絕對(duì)誤差(MAE),平均絕對(duì)百分比誤差(MAPE)以及相關(guān)系數(shù)(2)進(jìn)行模型精度比較,指標(biāo)的計(jì)算公式如下所示:

2 結(jié)果與討論

2.1 預(yù)測(cè)結(jié)果的影響因子分析

2.1.1 空氣污染物因子對(duì)PM2.5濃度的影響 PM2.5與其它空氣污染物之間存在著物理化學(xué)層面的相互轉(zhuǎn)化或者在傳輸過(guò)程之間產(chǎn)生相互影響[14],因此,對(duì)研究區(qū)PM2.5濃度與其它污染物變量之間進(jìn)行了相關(guān)性分析.

如圖4所示,分析PM2.5與其它大氣污染物(CO,NO2,O3,PM10,SO2)之間的相關(guān)性,可以發(fā)現(xiàn), PM2.5與各污染物之間均存在一定的相互關(guān)系.其中PM10、CO與PM2.5之間的相互關(guān)系極強(qiáng),而O3與PM2.5之間的相關(guān)性最低,所以可以忽略O(shè)3對(duì)于PM2.5的影響,這與前人[14]的研究結(jié)果相同.

綜上分析,將SO2、NO2和CO 3個(gè)污染物變量作為預(yù)測(cè)模型的輸入,其中與PM2.5有極強(qiáng)相關(guān)性的PM10未加入到輸入變量集中,因?yàn)榻?jīng)過(guò)實(shí)驗(yàn)分析PM2.5與PM10相關(guān)性過(guò)高導(dǎo)致產(chǎn)生冗余,從而導(dǎo)致精度降低.

圖4 PM2.5與其他大氣污染物的相關(guān)性分析

表2 PM2.5濃度與氣象因素相關(guān)系數(shù)

2.1.2 氣象因素對(duì)PM2.5濃度的影響 氣象因子也是影響PM2.5濃度的一個(gè)重要因子,已有大量學(xué)者證明PM2.5濃度與風(fēng)速、風(fēng)向、濕度、氣壓、氣溫等因素之間具有密切關(guān)系[8,42-43].對(duì)研究區(qū)PM2.5與氣象因素進(jìn)行皮爾遜相關(guān)性分析,結(jié)果如表2所示.PM2.5與氣象因子存在一定的相關(guān)性,其中PM2.5與氣壓和風(fēng)向呈正相關(guān)關(guān)系,與氣溫、風(fēng)速、邊界層高度、相對(duì)濕度和降水量呈負(fù)相關(guān)關(guān)系.

在本研究中,氣象因子作為輔助變量進(jìn)行PM2.5濃度預(yù)測(cè),因此氣象因子所有變量均加入本文實(shí)驗(yàn)中.

2.2 變量重要性分析

利用訓(xùn)練好的XGBoost模型對(duì)輸入變量的重要性進(jìn)行評(píng)價(jià),如圖5所示,對(duì)于未來(lái)1h PM2.5濃度預(yù)測(cè),變量重要性結(jié)果為污染物變量大于氣象變量重要性,其順序?yàn)楫?dāng)前PM2.5濃度、CO濃度、SO2濃度、NO2濃度、降水量、邊界層高度、風(fēng)向角度、平均氣溫、平均氣壓、平均風(fēng)速、相對(duì)濕度.污染物變量中當(dāng)前PM2.5濃度值和和CO濃度值重要性相對(duì)較高,而SO2、NO2濃度值重要性相對(duì)較低.氣象因子變量中,降水量和邊界層高度較為重要,平均風(fēng)速和相對(duì)濕度的重要性相對(duì)較低.

圖5 變量重要性分析

2.3 短時(shí)預(yù)測(cè)分析與對(duì)比

為了驗(yàn)證改進(jìn)的組合模型XGBoost-LSTM (Variable)精度,選擇XGBoost、SVR、LSTM、XGBoost-LSTM(Equal)、XGBoost-LSTM(Residual)模型進(jìn)行對(duì)比實(shí)驗(yàn).不同模型預(yù)測(cè)值與實(shí)際值的對(duì)比如圖6～7所示.

由圖6可知,PM2.5濃度值實(shí)際值處于15～80ug/ m3時(shí),各模型預(yù)測(cè)值和實(shí)際值的擬合度均較高,而對(duì)于實(shí)際值小于15ug/m3和大于80ug/m3的擬合效果均較差.單機(jī)器學(xué)習(xí)模型的擬合效果劣于組合模型的擬合效果,組合模型中,改進(jìn)的變權(quán)組合模型與實(shí)際值的擬合效果最好,起伏程度更加接近PM2.5濃度變化的實(shí)際趨勢(shì),偏差較小.

由圖7可知,組合模型的預(yù)測(cè)精度優(yōu)于單機(jī)器學(xué)習(xí)模型和傳統(tǒng)賦權(quán)方法組合模型預(yù)測(cè)精度.其中,改進(jìn)的組合模型XGBoost-LSTM(Variable)的MAE、MAPE和RMSE值相較于XGBoost-LSTM(Equal)模型分別提升了27.3%、22.9%、32.7%,相較于XGBoost-LSTM(Residual)分別提升了20.6%、19.7%、15.1%,表明改進(jìn)的變權(quán)組合方法具有更高的預(yù)測(cè)精度.

圖6 各模型預(yù)測(cè)和實(shí)測(cè)結(jié)果對(duì)比

圖7 不同模型實(shí)測(cè)值與預(yù)測(cè)值

2.4 不同季節(jié)預(yù)測(cè)分析

研究區(qū)屬于典型亞熱帶季風(fēng)性氣候,季節(jié)氣候存在明顯差異,并且不同季節(jié)具有不同的污染物來(lái)源.因此針對(duì)不同季節(jié)選取典型月份進(jìn)行預(yù)測(cè)分析,月份選取分別為春季(4月)、夏季(6月)、秋季(10月)和冬季(1月).

由圖8可知,本研究改進(jìn)的變權(quán)組合模型在春季和秋季的預(yù)測(cè)結(jié)果較好,其中春季即4月份為代表的預(yù)測(cè)精度最高,RMSE、MAE和MAPE各指標(biāo)值分別為1.65、1.23和2.81,遠(yuǎn)小于其它季節(jié)的指標(biāo)值;而在夏季和冬季的預(yù)測(cè)結(jié)果較差,其中夏季的預(yù)測(cè)結(jié)果最差,指標(biāo)值分別為7.56、6.04和15.19.對(duì)于模型不同季節(jié)典型月份預(yù)測(cè)結(jié)果分析來(lái)看,造成夏季預(yù)測(cè)結(jié)果較差原因是由于夏季強(qiáng)烈的大氣層活動(dòng),降雨頻率高以及風(fēng)速快,形成了較好的大氣顆粒物擴(kuò)散和清除的氣象條件[44].而在冬季預(yù)測(cè)結(jié)果較好是由于PM2.5濃度與影響因子的相關(guān)性更好[25,45-46].

圖8 組合模型四季典型月份預(yù)測(cè)結(jié)果

2.5 討論

使用氣象數(shù)據(jù)、空氣污染物數(shù)據(jù)以及PM2.5濃度歷史數(shù)據(jù)構(gòu)建了變權(quán)組合模型.以上海為研究區(qū)域,進(jìn)行未來(lái)1h短期PM2.5濃度預(yù)測(cè).采用改進(jìn)的XGBoost-LSTM(Variable)變權(quán)組合模型,RMSE、MAE和MAPE值為1.75、1.12和6.06,遠(yuǎn)小于甕克瑞[47]提出的組合模型預(yù)測(cè)值8.901、6.774和8.862,以及Liu Hui[28]提出的集成模型4.51、2.78和7.79,是由于將時(shí)間序列預(yù)測(cè)模型中性能最好的LSTM網(wǎng)絡(luò)和非線性模型中表現(xiàn)較好的XGBoost模型以變權(quán)組合的形式進(jìn)行組合預(yù)測(cè),該模型不僅考慮了數(shù)據(jù)的時(shí)間序列特征,又兼顧了數(shù)據(jù)的非線性特征;對(duì)于短時(shí)預(yù)測(cè)分析對(duì)比結(jié)果,本研究改進(jìn)的XGBoost- LSTM(Variable)變權(quán)組合模型優(yōu)于XGBoost- LSTM(Equal)組合模型、XGBoost-LSTM(Residual)組合模型,是由于本方法考慮到XGBoost模型和LSTM網(wǎng)絡(luò)在不同時(shí)刻預(yù)測(cè)誤差不同,通過(guò)對(duì)不同時(shí)刻采取不同的權(quán)重值,充分融合XGBoost模型和LSTM網(wǎng)絡(luò)的優(yōu)勢(shì).

研究區(qū)域選擇中,不同地區(qū)的污染物組成以及氣象條件具有強(qiáng)烈的地方性特點(diǎn),因此,本研究只選取上海市作為研究區(qū)域探討模型的表現(xiàn).另外,在PM2.5預(yù)測(cè)影響因素選擇上,本研究目前只將氣象和空氣質(zhì)量污染物要素作為預(yù)測(cè)因子,未來(lái)應(yīng)該考慮土地利用變化因素、經(jīng)濟(jì)、交通、環(huán)保政策等更多合適的因素進(jìn)行預(yù)測(cè)研究,以進(jìn)一步提高PM2.5預(yù)測(cè)精度.

3 結(jié)論

3.1 模型變量重要性分析可知,污染物變量的相對(duì)重要性高于氣象因子變量重要性,其中當(dāng)前PM2.5和CO濃度相對(duì)重要性高,而平均風(fēng)速和相對(duì)濕度重要性較低.

3.2 由于組合模型不僅考慮了數(shù)據(jù)的時(shí)間序列特征,又兼顧了數(shù)據(jù)的非線性特征,因此,與單機(jī)器學(xué)習(xí)模型和其它組合模型結(jié)果相比,改進(jìn)的變權(quán)組合模型的預(yù)測(cè)結(jié)果與真實(shí)值更加接近,誤差更小,穩(wěn)定性也更強(qiáng),可以用于PM2.5濃度短期預(yù)警預(yù)報(bào).

3.3 由于季節(jié)特征等差異,改進(jìn)的組合模型在季節(jié)上的表現(xiàn)有所差異,表現(xiàn)為在春、秋季節(jié)預(yù)測(cè)效果較好,而在夏、冬季節(jié)預(yù)測(cè)結(jié)果較差.

[1] Kim Y, Manley J, Radoias V. Medium- and long-term consequences of pollution on labor supply: evidence from Indonesia [J]. IZA Journal of Labor Economics, 2017,6(1):1-15.

[2] 王庚辰,王普才.中國(guó)PM2.5污染現(xiàn)狀及其對(duì)人體健康的危害[J]. 科技導(dǎo)報(bào), 2014,32(26):72-78.

Wang G C, Wang P C. PM2.5pollution in China and its harmfulness to human health [J].Science & Technology Review, 2014,32(26):72-78.

[3] Dennis R L, Byun D W, Novak J H. The next generation of integrated air quality modeling: EPA's models-3 [J]. Atmospheric Environment, 1996,30(12):1925-1938.

[4] 周廣強(qiáng),謝 英,吳劍斌,等.基于WRF-Chem模式的華東區(qū)域PM2.5預(yù)報(bào)及偏差原因[J]. 中國(guó)環(huán)境科學(xué), 2016,36(8):2251-2259.

Zhou G Q, Xie Y, Wu J B, et al.WRF-Chem based PM2.5forecast and bias analysis over the East China Region [J].China Environmental Science, 2016,36(8):2251-2259.

[5] Qingxin W, Qiaolin Z, Jinhua T, et al. Estimating PM2.5concentrations based on MODIS AOD and NAQPMS data over Beijing?Tianjin?Hebei. [J]. Sensors (Basel, Switzerland), 2019,19(5):1207.

[6] Zhang Z, Wu L, Chen Y. Forecasting PM2.5and PM10concentrations using GMCN(1,N) model with the similar meteorological condition: Case of Shijiazhuang in China [J]. Ecological Indicators, 2020,119: 106871.

[7] Pai T, Ho C, Chen S, et al. Using seven types of GM (1, 1) model to forecast hourly particulate matter concentration in Banciao City of Taiwan [J]. Water, Air, & Soil Pollution, 2011,217(1):25-33.

[8] 方曉婷,段華波,胡明偉,等.氣象因素對(duì)大氣污染物影響的季節(jié)差異分析及預(yù)測(cè)模型對(duì)比——以深圳為例[J]. 環(huán)境污染與防治, 2019, 41(5):541-546.

Fang X T, Duan H B, Hu W M,et al. The seasonal differential effects of meteorological parameters on atmospheric pollutants and the prediction model comparison: a case study of Shenzhen [J]. Environmental Pollution & Control, 2019,41(5):541-546.

[9] Liao Q, Zhu M, Wu L, et al. Deep learning for air quality forecasts: a review [J]. Current Pollution Reports, 2020:1-11.

[10] 戴李杰,張長(zhǎng)江,馬雷鳴.基于機(jī)器學(xué)習(xí)的PM2.5短期濃度動(dòng)態(tài)預(yù)報(bào)模型[J]. 計(jì)算機(jī)應(yīng)用, 2017,37(11):3057-3063.

Dai L J, Zhang C J, Ma L M,et al. Dynamic forecasting model of short-term PM2.5concentration based on machine learning [J].Journal of Computer Applications, 2017,37(11):3057-3063.

[11] 鄭 毅,朱成璋.基于深度信念網(wǎng)絡(luò)的PM2.5預(yù)測(cè)[J]. 山東大學(xué)學(xué)報(bào)(工學(xué)版), 2014,44(6):19-25.

Zheng Y, Zhu C Z. A prediction method of atmospheric PM2.5based on DBNs [J].Journal of Shandong University(Engineering Science), 2014,44(6):19-25.

[12] 朱晏民,徐愛(ài)蘭,孫 強(qiáng).基于深度學(xué)習(xí)的空氣質(zhì)量預(yù)報(bào)方法新進(jìn)展[J]. 中國(guó)環(huán)境監(jiān)測(cè), 2020,36(3):10-18.

Zhu Y M, Xu A L, Sun Q. New progress for air quality forecasting methods based on deep learning [J].Environmental Monitoring in China,2020,36(3):10-18.

[13] 謝永華,張鳴敏,楊 樂(lè),等.基于支持向量機(jī)回歸的城市PM2.5濃度預(yù)測(cè)[J]. 計(jì)算機(jī)工程與設(shè)計(jì), 2015,36(11):3106-3111.

Xie Y H, Zhang M M, Yang L, et al. Predicting urban PM2.5concentration in China using support vector regression [J].Computer Engineering and Design,2015,36(11):3106-3111.

[14] 侯俊雄,李 琦,朱亞杰,等.基于隨機(jī)森林的PM2.5實(shí)時(shí)預(yù)報(bào)系統(tǒng)[J]. 測(cè)繪科學(xué), 2017,42(1):1-6.

Hou J X, Li Q, Zhu Y J, et al. Real-time forecasting system of PM2.5concentration based on spark framework and random forest model [J].Science of Surveying and Mapping, 2017,42(1):1-6.

[15] 任才溶,謝 剛.基于隨機(jī)森林和氣象參數(shù)的PM2.5濃度等級(jí)預(yù)測(cè)[J]. 計(jì)算機(jī)工程與應(yīng)用, 2019,55(2):213-220.

Ren C R, Xie G. Prediction of PM2.5concentration level based on random forest and meteorological parameters [J].Computer Engineering and Applications,2019,55(2):213-220.

[16] 夏曉圣,陳菁菁,王佳佳,等.基于隨機(jī)森林模型的中國(guó)PM2.5濃度影響因素分析[J]. 環(huán)境科學(xué), 2020,41(5):2057-2065.

Xia X S, Chen J J, Wang J J, et al. PM2.5concentration influencing factors in China based on the random forest model [J].Environmental Science,2020,41(5):2057-2065.

[17] 王 敏,鄒 濱,郭 宇,等.基于BP人工神經(jīng)網(wǎng)絡(luò)的城市PM2.5濃度空間預(yù)測(cè)[J]. 環(huán)境污染與防治, 2013,35(9):63-66.

Wang M, Zou B, Guo Y, et al. BP artificial neural network-based analysis of spatial variability of urban PM2.5concentration [J].Environmental Pollution & Control, 2013,35(9):63-66.

[18] 白盛楠,申曉留.基于LSTM循環(huán)神經(jīng)網(wǎng)絡(luò)的PM2.5預(yù)測(cè)[J]. 計(jì)算機(jī)應(yīng)用與軟件, 2019,36(1):67-70.

Bai S N, Shen X L. PM2.5Prediction based on LSTM recurrent neural network [J].Computer Applications and Software,2019,36 (1):67-70.

[19] Zhang Y, Bocquet M, Mallet V, et al. Real-time air quality forecasting, part I: History, techniques, and current status [J]. Atmospheric Environment, 2012,60(1):632-655.

[20] 段大高,趙振東,梁少虎,等.基于LSTM的PM2.5濃度預(yù)測(cè)模型[J]. 計(jì)算機(jī)測(cè)量與控制, 2019,27(3):215-219.

Duan D G, Zhao Z D, Liang S H, et al. Research on PM2.5concentration prediction based on LSTM [J].Computer Measurement & Control, 2019,27(3):215-219.

[21] Liu D, Sun K. Short-term PM2.5forecasting based on CEEMD-RF in five cities of China [J]. Environmental Science and Pollution Research, 2019,26(32):32790-32803.

[22] Huang K, Xiao Q, Meng X, et al. Predicting monthly high-resolution PM2.5concentrations with random forest model in the North China Plain [J]. Environmental Pollution, 2018,242.

[23] Mao X, Shen T, Feng X. Prediction of hourly ground-level PM PM2.5concentrations 3days in advance using neural networks with satellite data in eastern China [J]. Atmospheric Pollution Research, 2017,8(6):1005-1015.

[24] 趙文芳,林潤(rùn)生,唐 偉,等.基于深度學(xué)習(xí)的PM2.5短期預(yù)測(cè)模型[J]. 南京師大學(xué)報(bào)(自然科學(xué)版), 2019,42(3):32-41.

Zhao W F, Lin R S, Tang W, et al. Forecasting model of short-term concentration based on deep learning [J].Journal of Nanjing Normal University (Natural Science Edition),2019,42(3):32-41.

[25] 康俊鋒,黃烈星,張春艷,等.多機(jī)器學(xué)習(xí)模型下逐小時(shí)PM2.5預(yù)測(cè)及對(duì)比分析[J]. 中國(guó)環(huán)境科學(xué), 2020,40(5):1895-1905.

Kang J F, Huang L X, Zhang C Y, et al.Hourly PM2.5prediction and its comparative analysis under multi-machine learning model [J].China Environmental Science, 2020,40(5):1895-1905.

[26] 宋國(guó)君,國(guó)瀟丹,楊 嘯,等.沈陽(yáng)市PM2.5濃度ARIMA-SVM組合預(yù)測(cè)研究[J]. 中國(guó)環(huán)境科學(xué), 2018,38(11):4031-4039.

Song G J, Guo X D, Yang X, et al.ARIMA-SVM combination prediction of PM2.5concentration in Shenyang [J]. China Environmental Science, 2018,38(11):4031-4039.

[27] 李建更,羅奧榮,李曉理.基于互補(bǔ)集合經(jīng)驗(yàn)?zāi)B(tài)分解與支持向量回歸的PM2.5質(zhì)量濃度預(yù)測(cè)[J]. 北京工業(yè)大學(xué)學(xué)報(bào), 2018,44(12): 1494-1502.

Li J G, Luo A R, Li X l.Prediction of PM2.5mass concentration based on complementary ensemble empirical mode decomposition and support vector Regression [J].Journal of Beijing University of Technology, 2018,44(12):1494-1502.

[28] Liu H, Dong S. A novel hybrid ensemble model for hourly PM2.5forecasting using multiple neural networks: a case study in China [J]. Air Quality, Atmosphere & amp; Health, 2020:1-10.

[29] 王學(xué)梅,王鳳文,陳 滔,等.基于組合模型的PM2.5濃度預(yù)測(cè)及其不確定性分析[J]. 環(huán)境工程, 2020,38(8):229-235.

Wang X M, Wang F W, Chen T, et al.PM2.5concentration prediction and uncertainly analysis based on a composite model [J].Environmental Engineering,2020,38(8):229-235.

[30] Wang J, Shao W, Kim J. Multifractal detrended cross-correlation analysis between respiratory diseases and haze in South Korea [J]. Chaos, Solitons and Fractals: the Interdisciplinary Journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena, 2020,135:10.1016/j.Chaos.2020.109781.

[31] Chen J, Lu J, Avise J C, et al. Seasonal modeling of PM2.5in California's San Joaquin Valley [J]. Atmospheric Environment, 2014,92:182-190.

[32] 王新民,崔 巍.變權(quán)組合預(yù)測(cè)模型在地下水水位預(yù)測(cè)中的應(yīng)用[J]. 吉林大學(xué)學(xué)報(bào)(地球科學(xué)版), 2009,39(6):1101-1105.

Wang X M, Cui W. Application of changeable weight combination forecasting model To groundwater level prediction [J]. Journal of Jilin University (Earth Science Edition), 2009,39(6):1101-1105.

[33] Dietterich T G. An experimental comparison of three methods for constructing ensembles of decision trees: Bagging, boosting, and randomization [J]. Machine Learning, 2000,40(2):139-157.

[34] Wu Y, Qi S, Hu F, et al. Recognizing activities of the elderly using wearable sensors: a comparison of ensemble algorithms based on boosting [J]. Sensor Review, 2019,39(6):743-751.

[35] Hochreiter S, Schmidhuber J. Long Short-Term Memory [J]. Neural Computation, 1997,9(8):1735-80.

[36] 郭立力,趙春江.十折交叉檢驗(yàn)的支持向量機(jī)參數(shù)優(yōu)化算法[J]. 計(jì)算機(jī)工程與應(yīng)用, 2009,45(8):55-57.

Guo L L, Zhao C J. Optimizing parameters of support vector machine's model based on genetic algorithm [J].Computer Engineering and Applications, 2009,45(8):55-57.

[37] Zhai W, Cheng C. A long short-term memory approach to predicting air quality based on social media data [J]. Atmospheric Environment, 2020,237.

[38] Chang Y, Chiao H, Abimannan S, et al. An LSTM-based aggregated model for air pollution forecasting [J]. Atmospheric Pollution Research, 2020,11(8):1451-1463.

[39] Gang L, Jingying F, Dong J, et al. Spatial variation of the relationship between PM2.5concentrations and meteorological parameters in China [J]. BioMed Research International, 2015,2015,684618.

[40] 劉 明,王紅蕾,索良澤.基于變權(quán)組合模型的中長(zhǎng)期負(fù)荷概率密度預(yù)測(cè)[J]. 電力系統(tǒng)及其自動(dòng)化學(xué)報(bào), 2019,31(7):88-94.

Liu M, Wang H L, Suo L Z. Medium-and long-term load probability density forecasting based on variable weight combination model [J].Proceedings of the CSU-EPSA, 2019,31(7): 88-94.

[41] 王新民,崔 巍.變權(quán)組合預(yù)測(cè)模型在地下水水位預(yù)測(cè)中的應(yīng)用[J]. 吉林大學(xué)學(xué)報(bào)(地球科學(xué)版), 2009,39(6):1101-1105.

Wang X M, Cui W. Application of changeable weight combination forecasting model to groundwater level prediction [J].Journal of Jilin University (Earth Science Edition), 2009,39(6):1101-1105.

[42] 曲 悅,錢 旭,宋洪慶,等.基于機(jī)器學(xué)習(xí)的北京市PM2.5濃度預(yù)測(cè)模型及模擬分析[J]. 工程科學(xué)學(xué)報(bào), 2019,41(3):401-407.

Qu Y, Qian X, Song H Q, et al. Machine-learning-based model and simulation analysis of PM2.5concentration prediction in Beijing [J].Chinese Journal of Engineering,2019,41(3):401-407.

[43] 謝 超,馬民濤,于肖肖.多種神經(jīng)網(wǎng)絡(luò)在華北西部區(qū)域城市空氣質(zhì)量預(yù)測(cè)中的應(yīng)用[J]. 環(huán)境工程學(xué)報(bào), 2015,9(12):6005-6009.

Xie C, Ma M T, Yu X X. Forecasting model of air pollution index based on multi-artificial neural network in western region of Northern China [J].Chinese Journal of Environmental Engineering,2015,9(12): 6005-6009.

[44] 劉小真,任羽峰,劉忠馬,等.南昌市大氣顆粒物污染特征及PM2.5來(lái)源解析[J]. 環(huán)境科學(xué)研究, 2019,32(9):1546-1555.

Liu X Z, Ren Y F, Liu Z M, et al. Pollution characteristics of atmospheric and source apportionment of PM2.5in Nanchang City [J].Research of Environmental Sciences,2019,32(9):1546-1555.

[45] 張淑平,韓立建,周偉奇,等.冬季PM2.5的氣象影響因素解析[J]. 生態(tài)學(xué)報(bào), 2016,36(24):7897-7907.

Zhan S P, Han L J, Zhou W Q, et al. Relationships between fine particulate matter(PM2.5) and meteorological factors in winter at typical Chinese cities [J].Acta Ecological Sinical,2016,36(24):7897- 7907.

[46] 朱媛媛,高愈霄,劉 冰,等.京津冀秋冬季PM2.5污染概況和預(yù)報(bào)結(jié)果評(píng)估[J]. 環(huán)境科學(xué), 2019,40(12):5191-5201.

Zhu Y Y, Gao Y X, Liu B, et al. Concentration characteristics and assessment of model-predicted results of PM2.5in the Beijing- Tianjin-Hebei Region in autumn and winter [J].Environmental Science,2019,40(12):5191-5201.

[47] 翁克瑞,劉 淼,劉 錢. TPE-XGBOOST與LassoLars組合下PM2.5濃度分解集成預(yù)測(cè)模型研究[J]. 系統(tǒng)工程理論與實(shí)踐, 2020, 40(3):748-760.

Weng K R, Liu M, Liu Q. An integrated prediction model of PM2.5concentration based on TPE-XGBOOST and LassoLars [J].Systems Engineering-Theory & Practice,2020,40(3):748-760.

Short-term PM2.5concentration prediction based on XGBoost and LSTM variable weight combination model: a case study of Shanghai.

KANG Jun-feng1, TAN Jian-lin1, FANG Lei2*, XIAO Ya-lai1

(1.School of Civil and Surveying & Mapping Engineering, Jiangxi University of Science and Technology, Ganzhou 341000, China；2.Department of Environmental Science and Engineering, Fudan University, Shanghai 200433,China)., 2021,41(9)：4016～4025

In order to further improve the accuracy of PM2.5concentration prediction, a variable weight combination short-term 1-hour PM2.5concentration prediction model based on LSTM network and XGBoost model was proposed.First, analyze the predictive factors, explore the influence of air pollutant factors and meteorological factors on the PM2.5concentration, to determine the best PM2.5concentration predictive factors and analysis the variable importance. Then, after data pretreatment the LSTM prediction model and the XGBoost prediction model was built respectively, and adopt the adaptive variable weight combination method based on residual improvement to obtain the final prediction result. The results show that: The relative importance of pollutant variables is higher than the importance of meteorological factors, among which the relative importance of current PM2.5concentration and CO concentration is higher, while the importance of average wind speed and relative humidity is lower. The values of RMSE, MAE and MAPE of the variable weight combined XGBoost-LSTM (Variable) model proposed in this study are 1.75, 1.12 and 6.06, which are better than LSTM, XGBoost, SVR, XGBoost-LSTM (Equal) and XGBoost-LSTM (Residual) model. The combined model predicts performance best in spring but the forecast accuracy is poor in summer. The variable weight method combination model proposed in this study effectively combines the advantages of the two models, not only considers the time series information of the data but also takes into account the nonlinear relationship between the features, and has higher prediction accuracy compared with other models.

long short term memeny (LSTM)；XGBoost；PM2.5forecast；variable weight combination model

X831

A

1000-6923(2021)09-4016-10

康俊鋒(1978-),男,江西贛州人,副教授,博士,主要從事高性能GIS算法及其在環(huán)境與土地中的應(yīng)用研究.發(fā)表論文10余篇.

2021-01-26

國(guó)家重點(diǎn)研發(fā)計(jì)劃項(xiàng)目(2016YFC08033105);國(guó)家留學(xué)基金資助項(xiàng)目(201808360065);江西省教育廳科學(xué)技術(shù)研究項(xiàng)目(GJJ150661);國(guó)家自然科學(xué)基金青年基金資助項(xiàng)目(41701462)

*責(zé)任作者, 博士, fanglei@fudan.edu.cn