Identifying & Costing Optimal Wind Turbine Construction Scenarios

The placement of wind turbines, and their connecting paths, is a challenging, multi-objective problem. GeoSynergy has been investigating whether this problem can be modeled using neural networks, and whether the resulting wind turbine construction "scenarios" can be costed. Such a workflow would allow for the rapid screening of new, candidate wind turbine sites, using remote sensed data.

In the first instance, we developed a training dataset of ~50 Australian wind turbine sites. For each site, we generated layers including wind speed, capacity factor, DEM, and ground truth WT/path positions (Fig.1). Note that layers such as DEM and, where applicable visible channel satellite products, were limited to time periods before construction had commenced.

Fig.1 Example input layers and ground truth for the WT placement model (~50 sites used).

We then trained custom U-Nets/FCNs to predict turbine position, and pathing, heatmaps from the stacked input layers, using custom loss functions that promoted placement quality (Fig.2).

Fig.2. Example U-Net model architecture, and training data for a wind turbine placement model.

The following heatmaps show some example wind turbine placement predictions, versus selected input layers, and the ground truth turbine placements (Fig.3).

Fig.3 Example wind turbine placement predictions for two Australian sites. Selected input layers, and ground truth wind turbine positions, are also shown.

Figures 4 and 5 show aggregated predictions over two, larger regions. Compare the prediction heatmaps with the ground truth turbine positions.

Fig.4 Left: aggregated turbine predictions for this site (multiple tiles). Middle: IEC1 Capacity Factor layer for reference. Right: ground truth wind turbine positions over base map.

Fig.5 Left: aggregated turbine predictions for this site (multiple tiles). Middle: IEC1 Capacity Factor layer for reference. Right: ground truth wind turbine positions over base map.

We also compared some of our regional predictions with the results of a published GIS approach by the 100% Renewable Energy Group at ANU [1] (Fig.6). Notably, this ANU approach incorporates additional data layers, including protected land status, and high voltage grid access.

Fig.6. Comparison of our turbine placement model predictions (top row), to some selected predictions from the published ANU GIS modality (bottom row, [1]). Note the strong agreement in the top left prediction, which is quite striking, considering the small size of our training dataset.

Our initial results from modelling wind turbine placement suggest that 1) this problem is tractable and suited to neural network modelling; 2) modelling results will benefit from expanding the training dataset to a large number of sites (likely necessitating the use of international sites); and 3) modelling results will benefit from expanding the training data layers to include variables such as land conservation status, high voltage grid access, restricted air space, highway proximity, and population density.

Once we have generated a heatmap of priority wind turbine placement sites, the next step in developing a scenario is to find optimal wind turbine positions within that heatmap (Fig.7). Note that we are not considering inter-turbine wake turbulence, this is a simplification for the purposes of rapid scenario generation and costing. Instead, we expand the wind turbine point to a "disk", by applying a fixed buffer. This problem then resembles a "maximal disk cover problem", which we can optimize using simulated annealing. Because the theoretical maximum coverage is known (it is the sum of n ordered heatmap pixels, where n is the total disk area pixels), we can use it to benchmark our coverage results.

Fig.7 Example maximal coverage optimization, for establishing optimal wind turbine positions. Note that we are using IEC1 capacity factor in this case, for demonstration.

Once wind turbine positions have been established, we need to establish the pathing that connects these turbines. We have applied the multi-objective NN approach to pathing prediction (Fig.8).

Fig.8 Example wind turbine pathing prediction. Ground truth post-construction paths are shown in yellow (OSM). Ground truth wind turbine locations are shown in white.

For model training, we incorporated a (pre-construction) visible satellite product layer, ground truth wind turbine locations, along with other pertinent layers, and extracted ground truth paths from post-contruction OSM. Early results look promising. We suggest that this problem is less tractable, and likely more dependent on scaling the training dataset significantly, versus turbine placement. Scaling may require the training of an intermediate path segmentation model, from visible satellite products, as OSM data is often incomplete for these sites.

In light of the early stage of development of our pathing prediction, we have substituted Dijkstra's for the following discussion on scenario costing, where we connect the wind turbine nodes, whilst minimizing the of total slope traversed (Fig.9).

Fig.9 Top: Example Dijkstra's algorithm pathing for IEC1 capacity factor. Bottom: Example output of the predicted path to Shapefile, and overlay onto various layers.

The last step in scenario costing is to calculate construction costs, and turbine revenue, over time. For demonstration, we use historical wind energy pricing here. To cost a scenario, we set a construction completion date for each wind turbine. We then calculate it's revenue over time, using it's IEC capacity factor, and estimate time to profit (Fig.10). Various scenarios can be explored, for any site, by varying, for example, the number of turbines, or the dates of construction. We can incorporate Power Purchase Agreements (PPAs), Renewable Energy Credits (RECs), and other costing modifiers. We can also modify turbine outputs globally for the site, or individually per turbine, to evaluate a variety of outcomes.

Fig.10 This costing scenario relates to the previous optimization shown above, with n turbines 10, IEC1 capacity factor, and construction dates / energy prices from 01.01.2008.

In conclusion, this proof of concept work shows the potential for developing neural network models for rapid prototyping of costed, wind turbine construction scenarios. Such tools, together with conventional GIS modalities, may enhance the ability of stakeholders make informed decisions by efficiently evaluating multiple site configurations, optimizing costs, and balancing energy yield with environmental and logistical constraints. By integrating neural network models with traditional assessment methods, this approach has the potential to streamline planning processes, reduce uncertainty, and support more sustainable and cost-effective wind energy development.

References

[1] Cheng, C., Silalahi, D.F., Roberts, L., Nadolny, A., Weber, T., Blakers, A. and Catchpole, K., 2025. Heatmaps to Guide Siting of Solar and Wind Farms. Energies, 18(4), p.891.