羅浩郭盛勇包為民
摘要:大壩變形預(yù)報對大壩運行安全評估起著至關(guān)重要的作用。傳統(tǒng)模型預(yù)報精度不夠、模擬效果不穩(wěn)定;若大壩變形數(shù)據(jù)有異常值時,傳統(tǒng)機器算法模型識別和處理異常值的靈活性很小,導(dǎo)致預(yù)報結(jié)果有偏差。為了解決這些問題,首次將隨機森林算法運用到大壩變形監(jiān)測領(lǐng)域,將大壩測點根據(jù)隨機森林相似性矩陣分成若干個子集,針對每一個子集建立隨機森林預(yù)測模型,分區(qū)建立預(yù)測模型更符合工程實際情況。選取拱壩變形作為研究對象,驗證所建模型的適用性。結(jié)果表明,根據(jù)隨機森林的相似性矩陣對大壩各測點的分區(qū)情況符合物理和工程實際意義,對各分區(qū)子集測點利用隨機森林模型建立的預(yù)測模型,與支持向量機、BP神經(jīng)網(wǎng)絡(luò)模型相比,預(yù)測結(jié)果精度較高、模型穩(wěn)定性好,為大壩變形監(jiān)測提供了新思路。
關(guān)鍵詞:拱壩變形;監(jiān)控模型;監(jiān)測點分區(qū);隨機森林;變形預(yù)測
中圖分類號:TU196.1文獻標志碼:A文章編號:
16721683(2016)06011606
Random forest model and application of arch dam′s deformation monitoring and prediction
LUO Hao1,2,GUO Shengyong2,BAO Weimin1
1.College of Water Resources and Hydrology,Hohai University,Nanjing 210098,China;
2.Yalong River Hydropower Company Ltd,Chengdu 610051,China)
Abstract:Dam deformation prediction plays an important role in the safety assessment of dam operation.Traditional models lack forecasting precision and the simulation effect is not stable enough.Besides,if abnormal values of dam deformation exist,traditional machine algorithm model lacks the flexibility of dealing with these abnormal data,which will lead to the deviation of the forecasting results.In order to solve these problems,random forest algorithm was introduced to the field of dam deformation monitoring for the first time.Similarity matrix of random forest was applied to divide dam deformation monitoring points into several parts.Random forests prediction model was established for each part,which will avoid the defects of traditional models such as modeling of single point or using the same model for all deformation monitoring points.Establishing forecasting model for different parts of dam was more in line with engineering practice.Deformation data of one arch dam was analyzed and the feasibility of random forest model was verified.The results showed that partition of dam deformation points based on similarity matrix of random forest conformed to the physical and engineering practical significance.Compared with support vector machine and BP neural network model,the prediction model of random forests for each part had the higher prediction precision and stability,which provided a new approach in the area of dam safety monitoring.
Key words:arch dam deformation;monitoring model;partitions of monitoring points;random forests;deformation prediction
國內(nèi)外普遍將大壩變形監(jiān)測[12]作為主要的監(jiān)測項目,大壩受各種復(fù)雜因素的影響,變形值是反映其運行狀態(tài)的最直觀的表征。根據(jù)大壩變形的原型觀測資料建立準確的預(yù)測模型,對大壩位移進行預(yù)測,能及時發(fā)現(xiàn)大壩的異常變化,采取措施防止事故發(fā)生。因此,大壩變形預(yù)報對大壩運行的安全評估起著至關(guān)重要的作用。目前應(yīng)用較多數(shù)學(xué)模型主要包括統(tǒng)計模型[23]、確定性模型[45]和混合型模型[56],這些模型在一定程度上可以揭示監(jiān)測值和影響量之間定性和定量關(guān)系,但由于影響大壩位移的因素復(fù)雜,傳統(tǒng)的方法受變量多重共線性的影響或模型參數(shù)的選取不恰當,使得模型精度下降。近年來,一些學(xué)者將新興的機器算法如人工神經(jīng)網(wǎng)絡(luò)[78]、遺傳算法[9]、蟻群算法[1011]、支持向量機[1213]等算法建立大壩監(jiān)控模型,[JP2]但這些監(jiān)控模型的研究和應(yīng)用尚未達到完善的程度,每種方法都存在一定程度上的優(yōu)缺點。另外,由于大壩具有整體性,布置在壩體和壩基的各測點之間存在差異性和關(guān)聯(lián)性,目前位移監(jiān)控模型還是以單測點為主,單測點位移監(jiān)控模型存在很大的局限性,不能反映大壩整體位移變化情況;多維多測點模型較單測點位移模型更符合工程實際情況,但由于多測點位移監(jiān)控模型[14]中待定參數(shù)較多,要達到一定的變形分析和預(yù)報精度,對原型觀測數(shù)據(jù)要求較高,給建模造成很大難度,在實際工程中的應(yīng)用并不廣泛。隨機森林(Random Forest,RF)[15]算法是由Breiman在2001年提出的一種新的機器學(xué)習技術(shù),隨機森林模型能有效地分析非線性、具有高度共線性和相互影響的數(shù)據(jù),不需要提前假定模型的數(shù)學(xué)形式,該算法在在生物學(xué)[1617]、土壤學(xué)[1819]、醫(yī)學(xué)[20]等領(lǐng)域已經(jīng)得到了一定的應(yīng)用,但在大壩安全監(jiān)測領(lǐng)域應(yīng)用幾乎沒有。此外,相似性矩陣是隨機森林算法的重要的分析工具之一,嘗試利用隨機森林算法的相似性矩陣來表征大壩各位移監(jiān)測點之間的相似性關(guān)系,基于這種相似關(guān)系,將大壩測點分區(qū),分別對各區(qū)建立隨機森林回歸預(yù)測模型。隨機森林算法預(yù)測精度高、對于異常值的處理和噪聲方面具有很大的優(yōu)勢,不易出現(xiàn)過度擬合的線性,能有效處理復(fù)雜變量間的共線性問題,該算法為大壩安全監(jiān)控提供了一種新思路。