To achieve net-zero carbon emissions by 2050, the demand for renewable energy is increasing exponentially, with offshore wind farms as one potential area of investment. Offshore wind farm development requires effective mapping of near subsurface for turbine foundation design and construction, which faces many challenges related to seafloor topography mapping, shallow geohazard detection, structure interpretation and modeling, soil type analysis and property estimation, among other considerations. While a set of existing workflows/algorithms are available to deal with these challenges, here we revisit these from the perspective of pattern recognition and present an integrated workflow in which deep learning appears suitable for assisting in eight essential tasks for better wind farm site characterization (figure 1), including:
1) Cone penetration testing quality control: Given the fact that CPT measurements are often incomplete and noisy, data QC – particularly outlier detection and missing segment reconstruction – is necessary for improving the CPT quality and benefiting the following components, such as soil-type classification and property estimation. By treating the CPT logs as 1D curves, it is feasible to automate the process of CPT outlier detection and log reconstruction through an encoder-decoder and more similar DL architectures.
2) Soil-type classification: Differentiating the soil units is important for understanding the lithology in shallow subsurface and identifying zones for optimal turbine placement. Both simple machine learning algorithms such as random forest and deep learning algorithms such as long-short-term memory are capable of analyzing the CPT patterns and classifying the soil types into clay, sand, silty clay/sand and so on.
3) Denoising: While the seismic data acquired for the wind farm industry are of ultra-high resolution compared to the oil and gas industry, it is also contaminated with more noise and/or processing artifacts, particularly seabed multiples, due to stricter budget constraints on seismic processing. Therefore, denoising is expected to improve the quality of UHR seismic and assist with tasks such as horizon interpretation. Treating UHR seismic data as images, the applicable DL algorithms include autoencoder and generative adversarial network, as well as physics-guided convolutional neural network (CNN), which can be applied to improve the quality and interpretability of the seismic data.
4) Geohazard detection: Geologic complexities in near subsurface, such as the presence of boulders, are identified as geohazards that may cause risk for wind turbine placement. Therefore, detecting shallow geohazards is crucial to the success of a wind farm site development. While the presence of such geohazards is traceable in UHR seismic images, many DL algorithms – particularly U-net and its derivatives – are capable of identifying geohazard-related patterns as image anomalies and locating the existing geohazards.
5) Seismic-well tie: As the UHR seismic data are collected in time and the CPT testing is in depth, the process of seismic-well tie is required to calibrate both measurements before integrating data from both sources and building reliable subsurface models. The objective of matching UHR seismic patterns with CPT logs can be achieved by physics-guided CNN, flownet and more deep neural networks.
6) Seafloor mapping: The bathymetry map illustrates the seafloor topography and helps to assess the suitable type of wind turbine, such as Monopile and Tripod, for a given location. With the seafloor well imaged in UHR seismic, its mapping can be automated by implementing U-net and its derivatives.
7) Horizon picking: Horizon interpretation plays a crucial role in site characterization by not only serving as the input to velocity modeling but also providing spatial guidance to extend the soil model from 2-D to 3-D. While the limited quality of UHR seismic challenges the existing horizon-picking tools, this task can be accelerated and improved by implementing a classification and/or regression CNN.
8) Soil-property estimation: Estimating soil properties is crucial to understand characteristics such as stiffness of near subsurface and identification of optimal zones for placing wind turbines. A typical workflow usually starts with deriving a set of attributes and/or elastic properties from seismic (primarily acoustic impedance), then correlates the soil properties from CPT data with the derived seismic attributes and properties by either empirical methods or simple machine-learning schemes, such as artificial neural networks, and finally propagates the correlation in 3-D to reveal the variation of soil properties. With the recent advance in deep learning, the workflow can be further automated by implementing a convolutional neural network for directly linking UHR seismic and CPT measurements.
Application
Figure 2. An example of deep learning assistingthree tasks for the HKZ wind site characterization.(a) The layout of available UHR seismic and CPTlocations, (b) the seismic amplitude along lineNo. 45. (c) The picking of three horizons by thetwo-branch CNN. (d) The mapping of seafloor bythe unsupervised learning method, and (e) theestimation of FRR (left) and RES (right) properties onseismic line No. 45 by the semi-supervised method.
We use an example in figure 2 to demonstrate the results of tasks 6-8 on the public data from wind farm Hollandse Kust Zuid (HKZ), which lies 18 kilometers off the coast in the zone between the Hague and Zandvoort and is of an area of 235.8 square kilometers. The available data include a total of 125 UHR seismic lines and 50 CPT locations. For the DL algorithms, a two-branch CNN is used to pick three horizons, an unsupervised learning method for extracting the seafloor, and a semi-supervised learning method to estimate the soil properties, including cone-tip resistance (RES), sleeve friction (FRES) and friction ratio (FRR), which are essential to evaluate the stability and strength of the near subsurface and further the determination of the foundation to be built for wind turbines.
As demonstrated in the results, first, the ML prediction is at correct positions for all three horizons, with minimal mis-picks. Second, the seafloor topography clearly reveals the presence of sand dunes in this area, which represent the “no-go” zones for wind turbine placement. Last but not the least, several potential clay layers are indicated by the estimated low RES and high FRR. All of them verify the capability of DL in wind farm data processing and interpretation and moreover indicate its potential in assisting other tasks and further automating the site characterization workflow (figure 1).
Acknowledgements: We thank the Netherlands Enterprise Agency (RVO) for providing the geophysical and geotechnical data under the creative commons license 4.0 and SLB for granting its permission to publish the work.