Method for predicting a log during drilling of a well

Abstract

The invention relates to a method for predicting a log beyond the depth of the drill bit of a well. Based on at least one log training curve acquired in a training well of depth larger than that of the drill bit, which curve is fitted to a curve of the log acquired in the well, a model for predicting the log in the well is constructed by means of a supervised machine-learning method. For each training curve, a compound training curve formed by the segment of the training curve fitted to the curve of the well up to the drill bit, and by the segment of the training curve fitted to the other training curves beyond the drill bit, is constructed. The values of the well log are predicted beyond the drill bit depth by applying the model to the compound training curves.

Description

[0001]METHOD FOR PREDICTING A LOG WHILE DRILLING A WELL Technical field The present invention relates in particular to the field of geosteering, which aims to guide the drilling operations of a well in an underground formation, based on the results of measurements (generally logging and seismic measurements) carried out in real time. More specifically, the invention relates to a method for predicting the values ​​of a log in a well beyond the drill bit, simultaneously with the drilling of this well, in particular with a view to guiding the drilling of the well beyond the depth of the drilling bit of the well, based on the prediction of the values ​​of the log. Geosteering consists of orienting the direction of a wellbore in real time, from an initial plan, determined from a model of the structures of the underground formation constructed upstream, and from logging measurements carried out during drilling.Thus, in the process of drilling a borehole, geosteering consists of adjusting the position of the borehole (inclination and azimuth angles) on the fly, to reach one or more geological targets. In the field of oil exploration, the objective is generally to orient the direction of the well towards an area of ​​the underground formation containing hydrocarbons. In mature areas, geosteering can be used to maintain a wellbore in a particular section of a reservoir in order to minimize gas or water breakthroughs and maximize the economic production of the well. Geosteering requires the upstream construction of a model of the underground formation, also called a geological model. These models, well known and widely used in the oil industry, must be representative of the boundaries of the geological layers as well as the rocks constituting these layers.This type of subsurface model is most often represented on a computer, by a mesh or grid, generally three-dimensional, each mesh being characterized by one or more petrophysical properties (facies, porosity, permeability, etc.). The construction of such a model, which must notably reflect the geometry of the boundaries of geological layers, requires at least an interpretation of a seismic image, obtained following a seismic acquisition and seismic processing. Filling the meshes with values ​​also requires analysis techniques for both seismic data (for example via stratigraphic inversion) and log data (for example gamma-ray, sonic, bulk density, electrical logs or well imaging logs). Thus, obtaining a geological model is a very long step and requires the intervention of several experts.While the borehole is being drilled according to the information derived from the geological model, new geological information is collected using logging tools. Typically, a typical configuration of tools lowered into a well includes directional and inclination sensors, as well as a gamma-ray logging tool. Other typical measurements are neutron density, prospective seismic, and downhole pressure measurements. Measurements taken while drilling usually show some differences from what is expected from the model. The geological model is then updated with the new geological information, along with the well plan to reach the corrected geological targets. Prior Art The following documents will be cited during the description: - Berg, CR, & Newson, AC (2013, August).Geosteering using true stratigraphic thickness. In SPE / AAPG / SEG Unconventional Resources Technology Conference. OnePetro. - Lallier, F., Caumon, G., Borgomano, J., Viseur, S., & Antoine, C. (2009). Dynamic time warping for stochastic stratigraphic well correlation. Search and Discovery Article, 40473. - Naville, C., Serbutoviez, S., Throo, A., Vincké, O., & Cecconi, F. (2004). Seismic while drilling (SWD) techniques with downhole measurements, introduced by IFP and its partners in 1990-2000. Oil & Gas Science and Technology, 59(4), 371-403. - Sakoe, H., & Chiba, S. (1978). Dynamic programming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26(1), 43-49. - Song, F., & Toksoz, M. N. (2009). Model-guided Geosteering for Horizontal Drilling. Massachusetts Institute of Technology. Earth Resources Laboratory. - Wang, Z. (2017). Integrated geosteering workflow for optimal well trajectory.Doctoral dissertation, Memorial University of Newfoundland. Geosteering methods have experienced significant growth in recent years. A well-known document (Berg and Newson, 2013) describes a geosteering method that uses only natural radioactivity logs (called Gamma-Ray logs). This method involves comparing a Gamma-Ray log in a horizontal well with the corresponding log in a neighboring vertical well (the pilot). In this method, the log record from the horizontal well is compressed and / or inverted until the distorted record can be best matched with the record in the pilot well. Compression is necessary because relatively small stratigraphic intervals in the pilot well will be traversed over much greater distances in the horizontal well.Inversion is also necessary because drilling can cover a given stratigraphic interval from bottom to top as well as from top to bottom. The scale factor used to match the two records is directly related to the inclination of the horizontal well. However, while this method requires little input data, it relies, as in any known geosteering technique, on real log measurements, measured on the fly, which are strictly valid only up to the drill bit. Also known is the document (Wang, 2017) which concerns an integrated geosteering methodology to determine an optimal well trajectory, combining geostatistical algorithms, logging measurements while drilling, and reservoir simulation methods. More specifically, the objective of this integrated technique is to explore a relatively fast update process for the exploration of oil fields without complex structures.It presents a procedure for using geostatistical realizations to control the drilling process and optimize the well trajectory in relation to production performance. The statistical realizations of the reservoir are updated with the in-drilling log data throughout the drilling process. However, as in all known geosteering techniques, this method relies on actual log measurements, which are only available down to the depth of the drill bit. We also know the document (Naville et al., 2004), which concerns seismic while drilling. This technique allows, among other things, to image in front of the drilling tool using the noise of the drill bit as a seismic source. It is therefore possible thanks to this technique to know, upstream of the drill bit, changes that could impact the initial well plan.However, this method has the disadvantage of requiring at least a seismic recording system, on the one hand echoes coming from geological formations in front of the tool, and on the other hand vibrations of the drill bit itself in order to eliminate in post-processing the signature of this seismic source to refine the imaging. This technique is therefore more cumbersome to implement than methods based on simple logs. We also know the document (Song and Toksoz, 2009) which describes a method which is similar to seismic during drilling but which uses real seismic sources instead of noise from the drill bit. This technique therefore has the disadvantage of requiring both seismic sources and sensors in the field, but also requires seismic processing which can be significant. The present invention makes it possible to overcome these disadvantages.Indeed, the method according to the invention aims to predict values ​​of a log beyond the depth of the drilling bit of a well, values ​​which can be directly exploited by conventional geoguiding methods, instead of the values ​​actually measured from the log but up to the depth of the bit only. The predictions in front of the bit can make it possible to identify, in advance and during drilling, future changes in lithology (for example the transition from the clay cover to the reservoir) which are important information which can be immediately used as input data for any geoguiding method.Summary of the invention The present invention relates to a method for predicting, during the drilling of a well in an underground formation, the values ​​of a log beyond the depth of the drill bit of said well, in particular with a view to guiding the drilling of said well beyond the depth of said drill bit of said well according to said prediction. Said method is characterized in that at least the following steps are carried out: A. acquiring a learning curve of said log as a function of depth in at least one training well, by means of a logging tool, said depth of said training well extending beyond the depth of the drill bit of said well; B. acquiring a curve of said log as a function of depth in said well, by means of said logging tool, said second curve being acquired up to said depth of the drill bit of said well; C.for each of said learning curves of said log, said learning curve of said log acquired in said training well is matched up to said depth of the drill bit of said well with said curve of said log acquired in said well, by means of a method for searching for similarity between signals recorded at different scales; D. a model is constructed for predicting said log in said well by means of a supervised machine learning method which is trained on a learning base comprising said learning curves of said log matched up to said depth of the drill bit of said well with said curve acquired in said well, as well as said curve of said log as a function of depth in said well; E.optionally, if a plurality of learning curves of said log have been acquired, said learning curves are matched with each other by means of said method of searching for similarity between signals recorded according to different scales; F. for each of said learning curves, a learning curve composed of said log is constructed, formed on the one hand by the portion of said learning curve of said log matched with said curve of said log up to the depth of said drill bit, and on the other hand by the portion of said learning curve, optionally matched with said other learning curves when a plurality of learning curves of said log have been acquired, for depths beyond said depth of said drill bit; and G.predicting said values ​​of said log in said well at least beyond said depth of said drill bit of said well by applying said model for predicting said log in said well constructed in step D) to said compound learning curves. According to one implementation of the invention, said method for searching for similarity between signals recorded at different scales may be the dynamic time warping method. According to one implementation of the invention, said log may be a natural radioactivity log. According to one implementation of the invention, said supervised machine learning method may be chosen from the following list: the random forest method, extremely random decision trees, and gradient boosting. According to one implementation of the invention, steps B) to G) may be repeated for a succession of depths of said drill bit of the well.The invention further relates to a method for geoguiding the drilling of a well, in which, for each depth of a succession of depths of said drill bit of said well: a) values ​​of a log beyond said depth of the drill bit of said well are predicted by means of the method for predicting, during the drilling of a well in an underground formation, the values ​​of a log beyond the depth of said drill bit of said well as described above; b) a direction of drilling of said well is determined for depths beyond the depth of said drill bit of said well, based at least on said values ​​of said log predicted beyond the depth of the drill bit of said well.Furthermore, the invention relates to a computer program product downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing steps C) to G) of the method as described above, when said program is executed on a computer. Other characteristics and advantages of the method according to the invention will appear on reading the following description of non-limiting exemplary embodiments, with reference to the appended figures described below. List of figures Figure 1A and Figure 1B show curves of natural radioactivity as a function of depth measured respectively in a training well and in a target well.Figure 1C illustrates a matching of the homologous events of the curves of Figures 1A and 1B detected by means of a similarity search method between signals recorded at different scales. Figure 1D and Figure 1E present the difference between the curves of Figures 1A and 1B, respectively before and after matching of the homologous events. Figure 2A, Figure 2B and Figure 2C present natural radioactivity logging curves measured respectively in a target well and in two training wells. Figure 3A presents a natural radioactivity logging curve predicted by the model according to the invention in the target well of Figure 2A, as well as a natural radioactivity logging curve actually measured in this target well. Figure 3B presents the difference between the two curves of Figure 3A.Description of the embodiments The invention relates to a method for predicting, during the drilling of a well in an underground formation, the values ​​of a log beyond the depth of the drill bit of the well, in particular with a view to guiding the drilling of the well according to the prediction. According to the invention and as will be described more precisely below, the prediction of the values ​​of a log is carried out for a well, hereinafter referred to as the "target well", at a time of advancement of the drilling tool in this well, in other words at a depth of the drill bit of the well. The objective is to predict the values ​​of a log, which has been measured up to the depth of the drill bit of the target well, beyond the depth of the drill bit of the target well.As will be described below, this prediction is made from a curve of the log measured in the target well down to the depth of the drill bit, and from at least one curve of the same log measured in another well of the subterranean formation, hereinafter called a "training well", extending beyond the depth of the drill bit. Thus, the method according to the invention aims to predict values ​​of a log in a target well, beyond the depth of the drill bit, from logs measured in training wells extending beyond the depth of the drill bit. Very preferably, the training well(s) cross the same subterranean formation as the target well. Alternatively, a training well may cross a geological analogue of the subterranean formation. Generally, a log results from a measurement made using a logging tool (or even a probe) moving in a well.The result of a well log measurement is a curve representing the variations of a characteristic of a geological formation as a function of depth in the well. Examples of well logs include gamma ray logs, sonic logs, bulk density logs, electrical logs, or well imaging logs. In a preferred embodiment of the invention, the predicted log may be a natural radioactivity log, known as a Gamma-Ray log or GR log. GR logs measure natural radioactivity in subterranean formations. They can be used to identify lithologies and to correlate depth intervals from one well to another. Sandstones and carbonates lacking clay have low concentrations of radioactive materials and give low GR responses.As the clay content increases, the GR response increases due to the concentration of radioactive materials in the clays. Predicting the gamma-ray response in a well in front of the drilling tool can therefore provide advance information on future lithology changes (e.g., the transition from a clay cover to a reservoir). This can reduce drilling risks and costs and improve well placement. Note that the method according to the invention can be applied to any logging indifferently. GR logging is preferred here because of its strong potential for geosteering.The method according to the invention comprises at least the following steps: 1) Acquisition of a logging learning curve in at least one training well 2) Acquisition of a logging curve in the target well 3) Matching each learning curve with the curve of the target well up to the drill bit depth of the target well 4) Construction of a model for predicting the logging in the target well 5) Matching the learning curves to each other (Optional) 6) Construction of a compound logging learning curve 7) Prediction of the logging in the target well beyond the drill bit depth The method according to the invention is implemented for a drill bit depth of the well.According to an implementation of the invention which will be described more broadly below, step 1) can be carried out only once, beforehand, and steps 2) to 7) of the method can be repeated as the target well is drilled, for example for a succession of depths of the drilling bit of the well. According to an implementation of the invention, in the case where several logs are measured in the target well, it is possible to repeat, at a given bit depth, steps 1) to 7) of the method according to the invention for each log of the plurality of logs that it is desired to predict, in order to take into account the plurality of predicted logs in particular to guide the drilling of the well beyond the depth of the bit. The method according to the invention comprises steps implemented by computer means (a computer), in particular steps 3) to 7).The computing means may comprise data processing means (a processor) and data storage means (a memory), as well as an input and output interface for entering data and restoring the results of the method. According to one embodiment of the invention, steps 2) to 7) may be carried out in real time during a drilling operation, so as to be able to possibly implement geoguidance for well drilling. The steps of the method according to the invention are detailed below. 1) Acquisition of a learning curve for the logging in at least one training well During this step, a learning curve for the logging of interest is acquired as a function of the depth in at least one training well, using a logging tool. This acquisition may be carried out during the drilling of the training well or after the training well has been drilled.According to the invention, a training well is a well whose depth extends beyond the depth of the drill bit of the target well (preferably beyond at least 10 meters, very preferably beyond at least 20 meters), in order to predict the logging for depths going beyond the depth of the drill bit. It is clear that a learning curve acquired in a training well according to the invention comprises values ​​of the logging for depths going beyond the depth of the drill bit of the target well. The method according to the invention requires at least one training well, preferably at least 2, very preferably at least 5. The invention is particularly configured to be relevant from the oil exploration phase, during which few wells are available. It is clear that the prediction of the logging will be all the more reliable as the number of training wells is large.It is clear that the learning curves of the log of interest can have depth scales that are distinct from each other, but also from the depth scale of the log curve measured in the target well. Indeed, the geological layers of an underground formation are rarely flat and horizontal. In addition, the thickness of the geological layers can vary laterally. Thus, the recording of a log corresponding to measured values ​​as a function of depth, a value representative of the same geological event (for example a boundary of geological layers) may not necessarily be at the same depth in one well as in another. 2) Acquisition of a curve of the log in the target well During this step, a curve of the log of interest as a function of depth is acquired in the target well, using the logging tool used in step 1).It is clear that this is not necessarily strictly the same tool, but rather any tool capable of measuring this same log. According to the invention, the log curve is acquired up to the depth of the drill bit of the target well, knowing that it is precisely desired to predict the log beyond the depth of the bit. As discussed above, the log curve recorded in the target well may have a depth scale distinct from the depth scale of the learning curves. In other words, a value representative of the same geological event (for example a boundary of geological layers) may not necessarily be at the same depth in the target well as in the learning well(s).3) Matching each learning curve with the target well curve up to the drill bit depth of the target well In this step, each learning curve is matched with the curve measured in the target well up to the drill bit depth of the well, using a similarity search method between signals recorded at different scales. More precisely, such a similarity search method between signals recorded at different scales aims to match the events of one curve with the homologous events of another curve. In other words, this step aims to match the events of a curve measured in a learning well with the homologous events of the curve measured in the target well, up to the drill bit depth of the target well.It is clear that we then obtain a learning curve whose events are scaled to the curve measured in the target well. In other words, the matching is carried out by considering the curve measured in the target well as a reference. Figures 1A to 1E illustrate the principle of this matching. More precisely, Figures 1A and 1B respectively present a curve C1 and a curve C2 of the measurement of natural radioactivity GR as a function of depth Z, the curve C1 having been measured in a learning well and the curve C2 having been measured in the target well. It can be observed that the curves C1 and C2 present homologous events, but shifted along the x-axis.Figure 1C illustrates, via gray lines, a matching of homologous events of curves C1 and C2 detected by means of a similarity search method between signals recorded at different scales (in this case the dynamic time warping method described below). It can be observed that while some homologous events were recorded at very close depths (essentially vertical gray lines), other homologous events were recorded at distant depths (oblique gray lines). Figures 1D and 1E present the difference between curves C1 and C2, respectively before and after matching. It can be observed in these figures that the difference between curves C1 and C2 is clearly less important after matching than before matching.More specifically, the standard deviation of the difference between the C1 and C2 curves decreases from 16% to less than 4% of the maximum amplitude of the natural radioactivity GR measurement, after matching. According to one implementation of the invention, the Dynamic Time Warping (DTW) method can be used to perform the matching. This method was first described in (Sakoe and Chiba, 1978). For the sake of clarity, it is possible to use a simple acoustic analogy to get a practical idea of ​​DTW. Consider two singers, one rather good and the other rather mediocre. The latter sings 'off-key' and does not respect the rhythm (sequence of time intervals between the sung notes), nor the key (absolute pitch of the notes), nor the tempo (performance speed) of the song.Faced with two recordings of the same song by these two singers, an ear, even a poorly trained one, will easily detect a certain similarity and especially several differences between these recordings. DTW not only precisely quantifies the difference between these recordings, but above all automatically matches the homologous events observed in these two recordings in time. It is also clear that this technique, although called dynamic time warping, can be applied to any signal whose abscissa axis is different from time, in this case whose abscissa axis corresponds to depth. The document (Sakoe and Chiba, 1978) describes the original implementation of the DTW method as follows. Let X(i) and Y(j) be two signals (here temporal, but the technique can be applied to any signal) to be matched.We define the function C(i,j) representing the cost density function of the warping for a pair (i,j) by: ^(^, ^) = ^^^( ^(^) − ^(^ )) + ^^^( ^(^ − 1, ^ − 1) , ^(^ − 1, ^), ^(^, ^ − 1) ) knowing that the first elements are given by the relations: ^(1,1) = ^^^( ^(1) − ^(1 ) ) ^(1, ^) = ^^^( ^(1) − ^(^ ) ) + ^(1, ^ − 1)) ^(^, 1) = ^^^( ^(^) − ^(1 ) ) + ^(^ − 1,1)). The Warping Function of the homologous indices i and j of the signals X and Y respectively, is defined as follows: let N be the number of samples of the signals X and Y. We assume that the pairs of points (X(1), Y(1)) and (X(N), Y(N)) are the first and last homologous points respectively. Then, starting from N. ème homologous point, the (N-1) èmepair of homologous points is the one whose cost density function corresponds to ^^^(^(^, ^ − 1), ^(^ − 1, ^ − 1) , ^(^ − 1, ^)), and so on in a recurring manner. Thus the (i-1) èmepair of homologous points is the one that corresponds to ^^^(^(^, ^ − 1), ^(^ − 1, ^ − 1), ^(^ − 1, ^) and so on up to the first pair of points (X(1), Y(1)). From the correspondence function, we determine for example a signal Y matched with the signal X. It is clear that we can use any improvement of this technique. Alternatively, we can use the method of searching for similarity between signals recorded according to different scales described in the document (Lallier et al., 2009), which concerns a variation of the DTW method applied to the correlation of geological horizons between wells.This method differs from the method of (Sakoe and Chiba, 1978) not only by the fact that the curves considered are in depth and not in time, but especially by the fact that it uses a stochastic technique of stratigraphic correlations of wells, the aim being to automatically generate several possible sets of well correlations taking into account the interpretations made along the well trajectories. In practice, the probability of correlation of each stratigraphic marker and the interval identified along the well path is calculated using: - available information on the interpreted well (e.g. paleobathymetry, lithology), - some reference correlation lines from clear first-order stratigraphic boundaries, - a sedimentological scenario on the studied area, - sedimentological concepts deemed applicable.Using these probabilities, several possible sets of well correlations are generated stochastically in a modified version of the DTW algorithm of (Sakoe and Chiba, 1978). 4) Building a model to predict the logging in the target well In this step, a model to predict the logging in the target well is built using a supervised machine learning method that is trained on a learning base comprising: - the learning curve(s) mapped, up to the said depth of the well drill bit, to the curve of the logging measured in the target well. These curves correspond to the input data of the supervised machine learning method during this training phase; - the curve measured in the target well.This curve corresponds to the output to be reproduced by the supervised machine learning method during this training phase. Thus, during this step, a supervised machine learning method is trained on the curves measured in the training wells, after these curves have been matched with the curve measured in the target well, so that the model can predict the log values ​​in the target well. This gives us a model for predicting the log in the target well. It is clear here that this is a regression problem, since we are determining values ​​of a continuous variable (the log measurement).According to one implementation of the invention, any of the following methods may be used as a supervised machine learning method: Random Forest, Extremely Random Trees, Gradient Boosting, Support Vector Machine, General Linearized Model, Generalized Additive Model or Feedforward Artificial Neural Network for Deep Learning.These methods are very simple to implement, even for a non-expert, in particular because they require only a few parameters: the maximum calculation time (typically a few minutes are sufficient, for example 200 seconds), and the distribution between the data that will actually be used for learning and the data that will be used for learning validation (typically 80% and 20% respectively). According to a preferred implementation of the invention, the supervised machine learning method can be chosen from the following list: the random forest method, extremely random decision trees, and gradient boosting. It has indeed been found that for a certain number of application examples on which the method according to the invention has been implemented, the three aforementioned methods have most often ranked among the most efficient in terms of prediction reliability.According to a highly preferred implementation of the invention, the gradient boosting method can be used. Indeed, it turned out that for the application examples on which the method according to the invention was implemented, this method was statistically the one that most often results in the best predictions. Gradient boosting uses two concepts, namely ensemble machine learning (Ensemble Machine Learning) and machine learning boosting (Machine Learning Boosting). In ensemble machine learning, the learning models are fitted to the data individually or combined into an ensemble. The ensemble is a combination of simple individual models that together create a new, more powerful model. Furthermore, machine learning boosting is a particular method of creating ensembles in the following manner.First, it starts by fitting an initial model (e.g., linear regression or decision tree) to the data. Then, the next model is built to make it more accurate to predict cases where the first model performs poorly. The combination of these two models is better than either of the two previous models taken separately. This process, called boosting, is repeated several times, with each successive model trying to correct the shortcomings in the predictions of the previous combined models. Finally, gradient boosting is a special type of boosting in machine learning. The general principle is that the best possible next model, when combined with the previous models, minimizes the overall prediction error.The term 'gradient amplification' comes from the fact that the target results for each case are defined as a function of the gradient of the error with respect to the prediction. More precisely, each new model takes a step in the direction that minimizes the prediction error, in the space of possible predictions for each learning case. 5) Matching the learning curves to each other (Optional) This step is optional and will only be implemented if a plurality of learning curves of the logging have been acquired. During this optional step, the learning curves are matched to each other, using a method of searching for similarity between signals recorded at different scales.In other words, during this step, if we have several training wells and if a logging curve has been recorded in each of these training wells, we match the curves recorded for each of these wells. Indeed, each logging curve has its own scale, because the geological layers are not flat, horizontal, and laterally isometric. To implement this step, we can use the Dynamic Time Deformation technique described above or the technique described in the document (Lallier, 2009) cited above to match the learning curves recorded for each of the training wells.Advantageously, the matching is performed by choosing as a reference the training well with the shortest distance to the target well, preferably with the shortest distance to the target well at the drill bit depth of the target well. This allows the training curves to be matched to each other with, in principle (depending on the stratigraphy), the least difference in scale from the curve acquired in the target well.6) Construction of a compound logging learning curve During this step, for each of the learning curves, a compound logging learning curve is constructed, formed on the one hand by the portion of the learning curve matched with the curve up to the depth of the drill bit, and on the other hand by a portion of the learning curve, optionally matched with the other learning curves when a plurality of logging learning curves have been acquired, for depths beyond the depth of the drill bit.In other words, for each training well, a composite curve is constructed, the part above the depth of the drill bit corresponding to the curve measured in this training well and matched with the curve measured in the target well, and the part below the depth of the drill bit corresponding: - if only one logging training curve has been acquired: to the part of the curve measured in this training well for depths beyond the depth of the drill bit, - if several logging training curves have been acquired: to the part of the curve measured in this training well and matched with the other training curves. At the end of this step, a composite training curve is obtained for each training well, consistent with the target well above the drill bit, and consistent with any other wells below the drill bit.7) Prediction of the log in the target well beyond the depth of the drill bit In this step, the values ​​of the log in the target well, beyond the depth of the drill bit of the well, are predicted by applying the model for predicting the log constructed in step 4) to the compound learning curves constructed in step 6). In other words, the regression model constructed in step 4) is applied to the compound learning curves constructed in step 6), which include values ​​of the log beyond the depth of the drill bit of the target well, so as to predict values ​​of the log in the target well beyond the depth of the drill bit of the target well. This results in a curve of the predicted log of interest in the target well comprising values ​​at least for depths beyond the depth of the drill bit of the target well.According to one implementation of the invention, the log of interest in the target well can be predicted beyond the depth of the drill bit of the target well, but also for depths below the depth of the drill bit of the target well. This can make it possible to verify the quality of the model constructed in step 4) in the part below the depth of the drill bit of the target well, by comparing the values ​​of the predicted log with those of the measured log. According to one implementation of the invention, the log of interest in the target well can be predicted only beyond the depth of the drill bit of the target well. This makes it possible to reduce the execution time of the method according to the invention, and thus to provide a predicted log curve more quickly, which is advantageous for the purpose of geoguiding the drilling of the target well in real time.Thus, the method according to the invention makes it possible to predict the values ​​of a log in the target well at least beyond the depth of the drilling bit of the well. The method can advantageously be implemented simultaneously with the drilling of this well, in order to guide the drilling of the well beyond the depth of the drilling bit of the well, according to the prediction of the values ​​of the log. Indeed, the geoguiding methods according to the prior art use real values ​​of the log, measured up to the depth of the bit only, whereas the method according to the invention determines predicted values ​​of the log beyond the depth of the bit, which can therefore be directly exploited by geoguiding methods. According to an implementation of the invention, step 1) can be carried out only once, and steps 2) to 7) as the target well is drilled, for example for a succession of depths.This makes it possible to gradually refine the prediction of the logging for the greatest depths in the target well, as the logging measurements made in the target well progress in depth with the progress of the drilling. According to an implementation of the invention, steps 2) to 7) of the method can be repeated for a succession of depths of the drilling bit of the well, for example every 10m, preferably every 5m, preferably every 1m. Alternatively, steps 2) to 7) of the method can be repeated for a succession of time steps during the drilling of the target well, for example every 60 minutes, preferably every 30 minutes, preferably every 10 minutes, these durations being predefined according to the progress speed of the drilling of the well (typically between 1.5m / hour and 15m / hour).It should be noted that for the application examples on which the method according to the invention was implemented, the execution time of steps 3) to 7) was of the order of 3 minutes on an Intel(R) Xeon(R) CPU E5-1620 v3 @ 3.50GHz processor (16.0 GB RAM), which is significantly below the conventional advance time of the drilling tool over a few meters: the prediction of the logging can therefore be done almost in real time.The invention further relates to a method for geoguiding the drilling of a well, in which, for each depth of a succession of depths of the drill bit of said well: a) values ​​of a log beyond the depth of the drill bit of said well are predicted by means of the method for predicting, during the drilling of a well in an underground formation, the values ​​of a log beyond the depth of the drill bit of the well as described above; b) the direction of drilling of the well is determined for depths beyond the depth of the drill bit of the well, based at least on the values ​​of the log predicted beyond the depth of the drill bit of the well.The person skilled in the art has perfect knowledge of the means for carrying out step b) above since, in the field of geoguidance, he determines the direction of drilling of the well from actually measured log curves, whereas in the present case, it is sufficient to consider the log curves predicted according to the invention as a supplement. According to an implementation of the invention, it is possible to apply, during step b), the technique described in the document (Berg and Newson, 2013) already cited above, and which relates to a geoguidance method making it possible to use only Gamma-ray logs, but using the log curves predicted according to the invention beyond the depth of the well drilling bit and not only the actually measured log curves.According to another implementation of the invention, the technique described in the document (Wang, 2017) already cited above, which relates to an integrated geosteering methodology combining geostatistical algorithms, logging-while-drilling techniques and reservoir simulation methods, can alternatively be applied. According to this implementation, the method described in the document (Wang, 2017) can be applied using the logging curves predicted according to the invention beyond the depth of the wellbore drill bit, and not only the actually measured logging curves. According to an implementation of the invention, in step b), at least the inclination and azimuth angles of the well trajectory are determined for depths extending between 10 and 20 m beyond the depth of the drill bit, for each depth of the wellbore drill bit in the succession of depths of the wellbore drill bit.Example The characteristics and advantages of the method according to the invention will appear more clearly upon reading the application example below. More specifically, the method according to the invention was applied to an underground formation located in the offshore of Western Australia, more precisely in the North Carnarvon Basin, considered to be the main hydrocarbon basin in Australia. Figure 2A shows a log curve of natural radioactivity GR measured in a target well, noted T hereafter (curve noted T in Figure 2A). The target well crosses a 6-meter oil column in the sandstones of the Barrow Group. This well also crosses water-bearing sands, with a thin oil column, marls and shales. Two training wells are available, noted Li (i=1 to 2), which are respectively approximately 3.5 km and 8 km apart from the target well T.Figures 2B and 2C present the natural radioactivity GR logging curves measured respectively in well L1 and in well L2 (curves denoted L1 and L2 in Figures 2B and 2C). Natural radioactivity values ​​are available for depths extending at least up to 3200m: 3277m for the target well T, 3295m for well L1, and 3200m for well L2. For the purpose of illustrating the reliability of the method according to the invention, we will seek to predict the GR logging values ​​for depths extending beyond 3100m, and we will check the quality of the prediction by means of the GR logging curve actually measured in the target well T for depths ranging from 3100 to 3200m. In other words, for the purpose of validating the method according to the invention, the method according to the invention is applied considering that the depth of the drilling bit in the target well is 3100m.In accordance with step 3) described above, the GR curves of wells L1 and L2 are matched with the GR curve of target well T up to the depth of 3100m, using the method described in the aforementioned document (Sakoe and Chiba, 1978). Furthermore, the method according to the invention is implemented using the gradient boosting method as a supervised machine learning method. The training of this machine learning method is carried out on the GR curves of wells L1 and L2 matched with the GR curve of target well T up to the depth of 3100m, as well as on the GR curve of target well T. This gives a model for predicting a log (here a GR log) in target well T, in accordance with step 4) described above.Then, according to step 5) described above, the GR curves of wells L1 and L2 are matched to each other, taking as reference well L1 which is the training well closest to the target well T. According to step 6) described above, compound training curves are generated for training wells L1 and L2, formed by the portion of the GR curves of these wells up to the depth of 3100m after their matching to the target well T, and by the portion of the GR curves of wells L1 and L2 beyond the depth of 3100m after their matching to each other. The model for predicting a log in the target well constructed according to step 4) is then applied to the compound training curves constructed as described above in step 6).Figure 3A shows the GRP curve which is the GR curve predicted by the model, as well as the GRM curve which is the actually measured GR curve, and Figure 3B shows the difference between these two curves. It can be observed that the predicted GRP curve is perfectly consistent with the actually measured GRM curve from 3100 m to 3200 m, i.e. up to a depth of 100 m beyond the depth to which the model was trained (depth of the drill bit in the method according to the invention). Indeed, in this depth range, the error quantified by the standard deviation of the difference between the GR curve predicted by the model according to the invention and the actually measured GR curve is less than 2% of the maximum amplitude of the actually measured GR curve. Beyond 3200 m, the errors increase significantly, and exceed 10% on average of the maximum amplitude of the actual GR measurement.Thus, the application of the method according to the invention made it possible to reliably predict a log (here a GR log) beyond the depth of the well drill bit, over a hundred meters. This is very advantageous because, for conventional drilling advancement speeds, typically between 1.5 m / hour and 15 m / hour, this corresponds to more than 6 hours of drilling to reach this limit. However, the calculation time of the method according to the invention is significantly lower, typically of the order of a few minutes. Also, it is possible to restart the method according to the invention as the drilling tool advances, and to update the model as soon as the new logging measurements are available. Thus, the updating of the log prediction can be done almost in real time.Alternatively, it may be sufficient to update the prediction model of a log and to predict a new log using this prediction model updated preferably at least every 10 m, or very preferably every 5 m of advance of the drilling tool in the target well. Thus, the present invention makes it possible to predict logging responses at depths not yet reached by drilling, typically at distances several decameters in front of the drill bit during drilling, in particular with a view to guiding drilling to such distances from these predictions, and not only from actual measurements, as according to the prior art.

Claims

Claims 1. Method for predicting, during the drilling of a well in an underground formation, the values ​​of a log beyond the depth of the drill bit of said well, in particular with a view to guiding the drilling of said well beyond the depth of said drill bit of said well according to said prediction, characterized in that at least the following steps are carried out: A. a learning curve of said log as a function of depth is acquired in at least one training well, by means of a logging tool, said depth of said training well extending beyond the depth of the drill bit of said well; B. a curve of said log as a function of depth is acquired in said well, by means of said logging tool, said second curve being acquired up to said depth of the drill bit of said well; C.for each of said learning curves of said log, said learning curve of said log acquired in said training well is matched up to said depth of the drill bit of said well with said curve of said log acquired in said well, by means of a method for searching for similarity between signals recorded at different scales; D. a model is constructed for predicting said log in said well by means of a supervised machine learning method which is trained on a learning base comprising said learning curves of said log matched up to said depth of the drill bit of said well with said curve acquired in said well, as well as said curve of said log as a function of depth in said well; E.optionally, if a plurality of learning curves of said log have been acquired, said learning curves are matched with each other by means of said method of searching for similarity between signals recorded according to different scales; F. for each of said learning curves, a learning curve is constructed composed of said log, formed on the one hand by the portion of said learning curve of said log matched with said curve of said log up to the depth of said drill bit, and on the other hand by the portion of said learning curve, optionally matched. correspondence with said other learning curves when a plurality of learning curves of said log have been acquired, for depths beyond said depth of said drill bit; and G. predicting said values ​​of said log in said well at least beyond said depth of said drill bit of said well by applying said model for predicting said log in said well constructed in step D) to said compound learning curves.

2. Method according to claim 1, wherein said method of searching for similarity between signals recorded at different scales is the dynamic time warping method.

3. Method according to one of the preceding claims, wherein said log is a natural radioactivity log. 4.Method according to one of the preceding claims, in which said supervised machine learning method is chosen from the following list: the random forest method, extremely random decision trees, and gradient amplification.

5. Method according to one of the preceding claims, in which steps B) to G) are repeated for a succession of depths of said well drilling bit. 6.A method of geoguiding the drilling of a well, wherein, for each depth of a succession of depths of said drill bit of said well: a) values ​​of a log beyond said depth of the drill bit of said well are predicted by means of the method for predicting, during the drilling of a well in an underground formation, the values ​​of a log beyond the depth of said drill bit of said well according to one of the preceding claims; b) a direction of drilling of said well is determined for depths beyond the depth of said drill bit of said well, based at least on said values ​​of said log predicted beyond the depth of the drill bit of said well. 7.Computer program product downloadable from a communications network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing steps C) to G) of the method according to one of the preceding claims, when said program is executed on a computer.

Images