基于時(shí)間序列相似性與機(jī)器學(xué)習(xí)方法的頁(yè)巖氣井產(chǎn)量預(yù)測(cè)

樊冬艷 楊燦 孫海 姚軍 張磊 付帥師 羅飛

摘要：頁(yè)巖氣井單變量產(chǎn)量預(yù)測(cè)存在較強(qiáng)的不確定性，而現(xiàn)場(chǎng)生產(chǎn)動(dòng)態(tài)數(shù)據(jù)同時(shí)包括多個(gè)相關(guān)指標(biāo)，針對(duì)如何選取合理的多變量數(shù)據(jù)對(duì)頁(yè)巖氣井產(chǎn)量進(jìn)行預(yù)測(cè)，在保證計(jì)算效率的情況下提高預(yù)測(cè)精度。頁(yè)巖氣井的生產(chǎn)動(dòng)態(tài)數(shù)據(jù)集包括日產(chǎn)氣量、日產(chǎn)水量、套壓、油壓、油嘴直徑、開井時(shí)間和溫度等，采用歐式距離和動(dòng)態(tài)時(shí)間彎曲距離對(duì)生產(chǎn)動(dòng)態(tài)數(shù)據(jù)時(shí)間序列進(jìn)行相似性度量，依據(jù)與日產(chǎn)氣量的相關(guān)度，把數(shù)據(jù)分為強(qiáng)相關(guān)時(shí)間序列和弱相關(guān)時(shí)間序列；其次，基于卷積神經(jīng)網(wǎng)絡(luò)、循環(huán)神經(jīng)網(wǎng)絡(luò)、長(zhǎng)短期記憶網(wǎng)絡(luò)和門控神經(jīng)網(wǎng)絡(luò)分別對(duì)全時(shí)間序列、強(qiáng)相關(guān)序列、弱相關(guān)序列和單變量序列進(jìn)行頁(yè)巖氣井產(chǎn)量預(yù)測(cè)；最后，以平均絕對(duì)誤差、均方根誤差和決定系數(shù)作為評(píng)價(jià)指標(biāo)，得到不同序列的誤差由小到大排序?yàn)閺?qiáng)相關(guān)序列、全時(shí)間序列、弱相關(guān)序列、單變量序列，優(yōu)選的機(jī)器學(xué)習(xí)方法為門控神經(jīng)網(wǎng)絡(luò)和長(zhǎng)短期記憶網(wǎng)絡(luò)。結(jié)果表明，采用機(jī)器學(xué)習(xí)方法結(jié)合頁(yè)巖氣井強(qiáng)相關(guān)性序列（日產(chǎn)氣量、套壓、油壓、日產(chǎn)水量）能有效降低預(yù)測(cè)誤差，提高頁(yè)巖氣井產(chǎn)量預(yù)測(cè)效果。

關(guān)鍵詞：頁(yè)巖氣井； 機(jī)器學(xué)習(xí)； 相似性； 時(shí)間序列； 產(chǎn)量預(yù)測(cè)

中圖分類號(hào)：TE 312?? 文獻(xiàn)標(biāo)志碼：A

文章編號(hào)：1673-5005（2024）03-0119-08?? doi：10.3969/j.issn.1673-5005.2024.03.013

Shale gas well production forecasting based on time sequence similarity and machine learning methods

FAN Dongyan1，2， YANG Can1， SUN Hai1，2， YAO Jun1，2， ZHANG Lei1，2， FU Shuaishi1，2， LUO Fei1

（1.State Key Laboratory of Deep Oil and Gas， China University of Petroleum（East China）， Qingdao 266580， China;2.School of Petroleum Engineering in China University of Petroleum（East China）， Qingdao 266580， China）

Abstract： Production data from shale gas wells contains multiple different dynamic variables during on-site collection， and there is uncertainty for production forecasting if only a single variable is used. It is important to choose reasonable multi-variable data to predict the output of shale gas wells， and ensure the precision accuracy and computing efficiency. In this study， a new method was proposed. Firstly， a dynamic data set can be comprehensively collected， including daily gas rate， water rate， well pressure， oil choke size， well opening time and fluid temperature. Euclidean distance and dynamic time warping were used to perform similarity testing of the production dynamic data time sequences. Based on the correlation with daily gas rate， the production data were divided into strong related time series and weak related time sequences. Secondly， based on convolutional neural network， recurrent neural network， long and short-term memory network （LSTM）and gate recurrent units （GRU）， the shale gas well production was predicted for full-time sequences， strong related sequences， weak related sequences and univariate sequences， respectively. Evaluation indicators were used to verify the methods， including average absolute error， root mean squared error and decision coefficient. The results indicate that the order of error from small to large for different sequences is the strong related sequence， the full time sequence， the weak related sequence， the univariate sequence. The preferred machine learning methods are the GRU and LSTM models. The strong correlation sequence can be used to improve the accuracy and reduce errors in shale gas well forecasting.

Keywords： shale gas well; machine learning; similarity; time series; productivity prediction

產(chǎn)量的準(zhǔn)確預(yù)測(cè)在油氣井高效開發(fā)和開采過(guò)程中至關(guān)重要，涉及整個(gè)生產(chǎn)開發(fā)歷程，包括早期資源評(píng)價(jià)、中期技術(shù)調(diào)整以及后期提高采收率措施［1］?！?br>