The proposed schemes CEGA and BLEA were applied to Horns Rev and Princess Amalia wind farms, by considering the cases with V80 and NREL5 turbine models. As mentioned, although the Gaussian wake model
was assumed as the reference in this work, the algorithms were also applied with the Jensen wake model
for comparison, due to its extended usage in the literature. Results showed that the current methodology (through either CEGA or BLEA algorithms) outperformed the corresponding baseline layout output power (
) for all cases. However, the performance showed a high variability depending on the wake model used, as well as the turbine model and wind farm considered. All obtained layouts were evaluated with the
× 1 m/s resolution wind roses (evalWR, see
Figure 2).
5.1. Results for the Crossover-Elitist Genetic Algorithm (CEGA)
CEGA results revealed a systematic higher performance through
, showing power output increases against the baseline more than two times those obtained through
. The fact that these differences are wake model-dependent was clear by crossing the wake model of the evolved layouts: the Jensen-evolved solutions abruptly decreased their performance when evaluated with
(J-G cases), whereas Gaussian-evolved solutions sharply increased their performance when evaluated with the Jensen model (G-J cases). This can be explained due to a general underestimation of the output power under the Jensen model, shown in Niayifar and Porté-Agel [
8] for the baseline layout of the HR-V80 case with respect to the Gaussian model and LES results. There, bigger wake effects are observed under the
model for almost all wind directions and multiple and diverse levels of wake interaction. Naturally, if the baseline layout is being underestimated under
, and a hypothetical perfect optimization (with no wake interaction) would provide the same performance under both schemes, any midway reduction of wake interactions will yield a bigger relative performance under
compared to
.
Figure 4 summarizes the power performances obtained with respect to the baseline layouts. Results for both wake models show bigger improvements for the PA wind farm and the NREL5 turbine, with
Po increases of 0.89% and 1.84% by respectively using
and
. On the other side, V80 turbines in the HR wind farm show the smallest rate of improvement, with power increases of 0.24% (
) and 0.72% (
). The other two cases provide midway results, with quite similar improvements (centred around 0.5% with
, and ∼1.15% with
). The fact that a bigger improvement is obtained through a bigger wind turbine as NREL5 (126 m-5 MW for NREL5 vs. 80 m-2 MW for V80) can be explained since bigger turbines imply higher power densities, which produce in turn stronger wake effects, as a consequence establishing favourable conditions for a bigger margin of improvement in an optimized layout.
The obtained results also highlight the high sensitivity of the obtained
Po improvement to the resolution of the wind rose used for evaluation, as shown in Feng and Shen [
53]. For instance, for the PA-NREL5 case (Jensen), the improvement rises to +2.03% and +3.54% when respectively using a
1 m/s evaluation wind rose or the evolWR (
resolution and constant speed bins). These values are comparable with results provided in recent works by other methods such as Sequential Quadratic Programming, which yielded more modest power improvements (+1.5% [
6] and +2.3% [
57] respectively). In a further comparison with the existing literature, the obtained improvements against the baseline layout for HR-V80 (again within
for a proper comparison) are shown to be nearly double (+0.72%) improvements provided using Random Search (+0.37% [
55] or ∼0.38% [
53]).
For each case study, the best performing individual at each generation was evaluated with evalWR to obtain an impression of the real performance during the whole process. Results on this
evaluated evolution for both
and
models are shown in
Figure 5 in terms of their relative performance against a population of random layouts (1st generation). This relative representation of the
improvement is useful to provide evidence of the real optimization capacity of the evolved layouts, independently of the baseline layouts. In addition, it also works as a metric for the optimization level that the baseline layouts per se have within their corresponding wind farms. Results show a higher level of optimization for the baseline layout in HR compared to PA (represented by blue lines in
Figure 5), for each turbine and wake model, this explaining why a bigger margin of optimization can be systematically attained in PA compared to that in HR.
The better optimized baseline in HR can be explained partly due to the outline of the wind farm area, which was defined according to the minimum-sized convex polygon containing the turbines. For instance, in the case of PA, this criterion produces the result that some regions of the wind farm perimeter, especially over the northeast (see
Figure 1b), remain with a relatively low density of turbines, thus generating favourable conditions for a power output improvement in a WFLO experiment. This contrasts with the baseline layout in HR, where the whole perimeter is shown equally filled up with turbines, so that the whole wind farm area is better harnessed. Moreover, the higher margin of improvement in PA compared to that in HR could also be related to the wind rose morphology (see
Figure 2). In this way, higher efficiencies are observed when optimizing with a unidirectional wind rose compared to a multiple-direction wind rose (e.g., [
45,
46,
55]). In this work, this idea can be promoting that a wind rose with some clear prevailing wind directions as PA allows a further power optimization compared to a site with a more homogeneous wind rose (as HR), with turbines tending to align spanwise from the prevailing wind direction.
Within the eight optimizations carried out through CEGA, different evolutionary behaviours were observed depending on the type of condition met to generate the transition point
Pt between exploration and exploitation. The two HR-V80 cases (
and
) and PA-NREL5 (
) met the
Ud diversity metric to switch mode, while the rest of cases did so due to improvement stagnation (improvement <0.02% in 1000 generations). Because evolWR and evalWR produce slightly different results, the
evaluated evolution can show small decreases, and the highest attained performances (represented by green dots in
Figure 5) might not be placed in the last generation.
Table 4 summarizes the type of
and other technical data about the evolution, such as the number of generations evolved or the processed computational time. The algorithms were written on
(Mathworks, Natick, Massachusetts, USA) and run under the ’parpool’ mode on a 12-core computer cluster, each core running with 64 GB DDR3 RAM and 2.6 GHz.
The obtained layouts for the best performing solutions through CEGA are shown in
Figure 6,
Figure 7,
Figure 8 and
Figure 9. The figures show the power output at each of the turbines, for both the obtained optimizations as well as for the baseline layouts. In addition, the difference in power output density is depicted (plots c, f in
Figure 6,
Figure 7,
Figure 8 and
Figure 9). Power densities and the difference between the obtained solutions and the baseline layouts were computed through Inverse Weight Distance Interpolation ([
91], with defined values power = 2, radius = Inf and Euclidean norm = 2). This interpolation method was used in order to properly be able to represent the power density due to the effect of each wind turbine, with a high value around the turbine but progressively decreasing with the distance from the turbine. As it can be observed, for most cases, the obtained solution increases the number of turbines in the perimeter of the wind farm with respect to the baseline. This in turn allows the turbines at the central part of the domains to increase their power output compared to the baseline layouts (see red areas in the power density difference (plots c, f in
Figure 6,
Figure 7,
Figure 8 and
Figure 9). Although the turbines on the perimeter suffer a power decrease compared to the baseline, the overall performance trade-off is positive due to a higher performance of the inner turbines. The tendency of the turbines to concentrate over the western side of the perimeter is also noteworthy, in line with the direction of the prevalent flow, followed by a rather wide void space just downstream. This also promotes a minimization of the losses, in line with results showing the second line turbines being largely the most penalized in relation with the immediately upstream turbines [
8,
92]. It is also remarkable that
systematically produces solutions that place more turbines on the perimeter of the wind farm area compared to those obtained with
. This can be explained by the fact that the Jensen model is related to a general underestimation of
and thus associated with greater wake effects, so that less turbines are susceptible to remain in the interior area of the wind farm (and thus remain exposed to wake effects from all directions).
Finally, the electricity cable length needed to interconnect the turbines was computed for every case, retrospectively to the evolution of the optimized solutions. The cable length was computed according to the estimation of the minimum spanning tree. From simple geometry, it can be derived that regular grids, as those observed in the baseline layouts, hold maximum spanning trees. Thus, a modification of the layout as done here, within a constant area constraint, is highly likely to see reduced the amount of cable needed, and thus allow additional improvement from the point of view of the efficiency of material usage. This has also been proven in previous works such as Fleming et al. [
6], where reductions of cable length of 13.1% are obtained for the conditions equivalent to the current PA-NREL case. In the case of CEGA solutions (see
Figure 6,
Figure 7,
Figure 8 and
Figure 9), the cable length was reduced between 17% and 28% with respect to the baseline, depending on the cases. Solutions using Jensen provided larger cable length reductions. The largest reductions are obtained for the PA-V80 case, with 28.6% and 24.3% reduction, respectively, for
and
. In turn, the smallest reductions are found in the PA-NREL case, with 20.1% and 17.3% reductions, respectively.
5.2. Results for the Baseline Layout Evolutionary Algorithm (BLEA)
All power output performances of the optimized layouts using BLEA showed some improvement compared to the baseline layouts for the four cases considered. However, these improvements were mostly relatively small compared to CEGA, this proving that most of the baseline layouts considered for this work represent a relative optimum regionally, this proving that, in general terms, it does not result in it being easy to escape from their surrounding search space. In turn, BLEA only provided higher improvements than CEGA solutions in the PA-NREL case (for both
and
). As CEGA explores all the search space, and not only the surroundings of the baseline, the better results in CEGA show that, in addition to representing a regional optimum, in most cases (at least 3 of 4), the baseline layouts belong to search spaces that actually represent sub-optimal (below optimum) solution regions of the overall problem. Relative power output improvements of the obtained layout solutions with BLEA are shown in
Table 5.
As with CEGA, improvements with the Jensen wake model yield bigger values than with the Gaussian one. Results highlight that the Horns Rev layout is very close to its regional optimum, especially when considering the V80 turbines. On the other side, Princess Amalia reflects the existence of a bigger margin of improvement around its baseline solution. In the case of the NREL turbine, this margin is big enough to even outperform the overall search of CEGA, yielding power performance improvements up to 1.91% and 0.95% under and , respectively.
In the following, we focus just on the results attained in the PA-NREL case, as it represents the only case showing a bigger
improvement than the overall search with CEGA.
Figure 10 shows the evolution of the performance for both Jensen and Gaussian wake models. Note the decreases in the evolution, due to the fact that the selected parent may not be the highest performing one, in order to preserve diversity. In addition to this, because layout evaluations with evolWR and evalWR can provide slightly different values, the maximum improvement during the evaluation can be found more than 1000 generations before the end of the evolution. Again, computation costs were remarkably higher under
.
Figure 11 shows the optimized layouts obtained in the PA-NREL case, for both
and
models, and the power density difference with the baseline. In addition to showing a higher performance compared to the CEGA solutions, in the BLEA ones, a layout with a more regular grid can be appreciated, providing smaller differences with the baseline layout turbine positions (11% and 25% smaller distances for
and
, respectively). This shows that, in some cases, the baseline layout can still provide information to yield a higher performance than if it was not considered. This leads us to highlight that the usage of BLEA can result in an appropriate procedure in cases where the baseline has a large margin of improvement but its surrounding search space has big potential. Since it can be difficult to find out if a certain baseline layout fulfils these features beforehand, it can be concluded that running BLEA in parallel to CEGA can result in an appropriate procedure. As in CEGA, the optimal layout enhances the power output of the inner turbines, while keeping positive the trade-off resulting from penalizing some power in the peripheral ones (
Figure 11c,f). Again, the resulting layouts are consistent with the prevailing flow regimes from SW, showing a wider void area just after the first row of turbines in that direction (especially in
due to its bigger wake estimates).
Finally, the retrospective computation of the minimum needed electricity cable length in the PA-NREL case yielded remarkable reductions with respect to the baseline layout compared with the CEGA scheme, with decreases of 16.8% and 6.3% for and wake models, respectively. This smaller reduction is consistent with the fact that BLEA solutions keep a bigger similarity to the regular-gridded baseline layouts (which exhibit maximum spanning trees).