Australian tidal currents – assessment of a barotropic model with an unstructured grid

While the variations of tidal range are large and fairly well known across Australia (less than 1 m near Perth but more than 14 m in King Sound), the properties of the tidal currents are not. We describe a new regional model of Australian tides and assess it against a validation dataset comprising tidal height and velocity constituents at 615 tide gauge sites and 95 current meter sites. The model is a barotropic implementation of COMPAS, an unstructured-grid primitive-equation model that is forced at the open boundaries by TPXO9v1. The Mean Absolute value of the Error (MAE) of the modelled M2 height 10 amplitude is 9.3 cm, or 13 % of the 73 cm mean observed amplitude. The MAE of phase (11°), however, is significant, so the M2 Mean Magnitude of Vector Error (MMVE, 20 cm) is significantly greater. Results for 5 other major constituents are similar. We conclude that while the model has skill at height in all regions, there is definitely room for improvement (especially at some specific locations) before harmonic predictions based on observations are rendered obsolete. For the M2 major-axis velocity amplitude, the MAE across the 95 current meter sites, where the observed amplitude ranges from 0.1 cm s to 144 15 cm s, is 6.5 cm s, or 20 % of the 31.7 cm s observed mean. This nationwide average result is not much greater than the equivalent for height, but it conceals a larger regional variation. Relative errors on the narrow shelves of NSW and Western Australia exceed 100 %, but tidal currents are weak and negligible there compared to non-tidal currents. We show that the model has predictive value for much of the 79 % of Australia’s shelf seas where tides are a major component of the total velocity variability. In descending order this includes the Bass Strait, Kimberley to Arnhem Land and Southern Great Barrier 20 Reef regions. There is limited evidence the model is also valuable for currents in other regions across northern Australia. We plan to commence publishing ‘unofficial’ tidal current predictions for chosen regions in the near future, based on both the limited number of observations, and the COMPAS model.


Introduction
Tidal currents are a major component of the velocity variability for most of the Australian continental shelf, yet tidal current 25 predictions are only listed in the Australian National Tide Tables for 7 sites, 5 of which are in Torres Strait. As part of a project to map Australia's tidal energy resource, and as a step towards an operational, model-based tidal current forecasting ability, we have compiled a tidal currents harmonic constituents validation dataset at 95 sites based on observations acquired by a number of agencies. This is a significant number of sites, but it is still small compared to the 683 sites for which the Bureau of Meteorology Tidal Unit has estimates of tidal height harmonic constituents. We use these validation datasets for currents and Reef (Beaman, 2010) and northern Australia (https://ecat.ga.gov.au/geonetwork/srv/eng/catalog.search#/metadata/121620).
Onsite depth measurements at the locations (Fig. 2) of the tidal currents validation data discussed below were not used for estimating the model topography. The minimum depth (at zero tide) in the model is 4 m for most of the grid, but 8 m in the NW, NE and in Gulf St Vincent. Depth was median filtered to remove sharp gradients. A deeper channel of 12 m was also included in King Sound (in the NW). These bathymetric changes had significant effect on the tidal response, and it is 75 anticipated that further model improvement will require targeted bathymetry modification at the local scale. The tide is introduced through eight tidal constituents (M2 S2 N2 K2 K1 O1 P1 Q1) from the TPXO9v1 1/6° global model (Egbert and Erofeeva, 2002; http://volkov.oce.orst.edu/tides/otps.html) and applied at the open boundary using the condition 80 described by . The  scheme includes a normal and tangential velocity Dirichlet condition with provision for a local flux adjustment on normal velocity to maintain domain-wide volume continuity. Thus, the surface height is not directly constrained at the boundary but is instead computed via volume flux divergence as it is in the model interior. For the present application, we found that flux adjustments to constrain the sea surface height were not required; prescribing the transports at the boundary was sufficient to achieve the target height. This situation is quite unusual. One 85 necessary step to achieve this was to use the TPXO components of transport on their native (Arakawa C) grid and use the depths in COMPAS to convert the transports to depth-averaged, cell-edge normal velocity, thus accounting for bathymetry differences between our model and the TPXO model. The model was run in 2D mode only, using a time-step of 1 s, achieving a runtime of ~5:1 on twelve processors. A spatially constant bottom drag coefficient of 0.003 was used to compute bottom stress. Tidal potential forcing is optionally applied in the model but we found that it made very little difference compared with 90 other parameters such as friction, so we have omitted it for the configuration described here.
For many test runs of the model, it was started from rest and run for either 7 or 30 days from 24 Feb 2017 including a 1-day ramp period. Results were compared with harmonically synthesised time series at all sites for which tidal constituents are available (see below) using T-Tide (v1.3b, Pawlowicz et al., 2002) and then parameters were manually adjusted, or model code de-bugged or improved, in search of closer agreement with the observations. Apart from this 'tuning', no data assimilation 95 was used with these model runs.
For the model configuration described here, it was run for 222 days from 24 Feb to 5 Oct 2017, and then tidally analysed so that 1) its performance can be described for individual constituents, and 2) predictions can be made for any time or place within the domain without having to run the model. The COMPAS model code, the output time series and tidal constituents at all points of the mesh are freely available, as described in Sections 9 and 10. 100

Current meter observations
Acoustic Doppler current profilers (ADCPs) of various types have been deployed as part of Australia's Integrated Marine Observing System (IMOS) at 55 sites over the continental shelf around Australia, some for several years, since 2007. The ADCPs are almost all moored within a few metres of the sea bed, and sense the water velocity over the lower 80-85% of the water column. We have taken the depth-average of these observations, concatenated all records from individual instrument 105 deployments at the same nominal position, and determined the tidal constituents using the UTide software of Codiga (2011).
Apart from the deployments off the NW of the continent, these sites tend to be at locations where tidal currents are not particularly strong. As a means of quantifying the relative magnitude of tidal and sub-tidal depth-average velocity, we determined the principal axis of the subtidal variability (using singular value decomposition) and computed the root mean square (RMS) of the major and minor axis components. Details of the IMOS ADCP deployments are at 110 https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License. http://oceancurrent.imos.org.au/timeseries/ along with regional graphics comparing the tidal and sub-tidal ellipse parameters (as well as the mean velocity for each deployment). Penesis et al. (2020) give details of ADCP deployments that deliberately sought to observe tidal currents for two of Australia's most prospective tidal energy development regions. These include seven locations in the Clarence Channel near Darwin and seven locations in Banks Strait at the NE tip of Tasmania. We determined tidal velocity constituents, the mean and sub-tidal 115 ellipse parameters from these data as above.
We have included data from 10 of the sites where Middleton et al. (1984) and Griffin et al. (1987) deployed current meters on the Southern Great Barrier Reef (SGBR, see Fig. 2) in order to study both the anomalous tides and the sub-tidal variability.
These observations were made by single, mechanical RCM4 Aanderaa current meters with several drawbacks compared to ADCPs. Due to limited storage capacity, the flow direction was only sampled instantaneously once an hour, so short-period 120 changes of direction were not averaged. To minimise noise due to waves, the instruments were moored fairly low in the water column (typically 7 m off the seabed), thereby probably underestimating the depth-average velocity. Some had to be deployed close to islands, with the result that they recorded effects (such as asymmetric ebb and flood directions) that the model is unlikely to be able to reproduce at specific locations due to its imperfect representation of topography. Nevertheless, we have included these records in our validation dataset, processed as above, despite the quality questions because 1) the tides in this 125 region are important for navigation (e.g. through Hydrographers Passage), and 2) in the hope that future models with finer meshes and better topography may be able to better distinguish observation error from model error.
Lastly, we also extracted 13 current meter records from the CSIRO archives (https://www.cmar.csiro.au/data/trawler/), choosing sites in Bass Strait, the NW shelf and the Gulf of Carpentaria where tidal currents are significant. These were mostly point measurements, either by acoustic or mechanical (Aanderaa) current meters. Where two instruments were deployed on a 130 mooring, we simply averaged the data for the period when both were operating.
In support of this paper and future studies of the tides of Australia, we have published this validation data set as a netCDF file containing up to 13 tidal constituents, and the subtidal statistics, for each of the 95 locations discussed above (see Section 10). https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.

Model-data comparison method
All model-data comparisons presented in this paper are based on the six major tidal constituents (M2, S2, N2, K1, O1 and P1) 140 determined from the model and observational time-series (rather than the time series approach used during model tuning) for all the usual reasons. The length of the analysed model run (222days) is long enough to resolve P1 from K1 but we do not discuss results for these lesser constituents (important as they are for predictions) in detail because the results for M2 are broadly representative of all constituents. Availability of the full set of model-data comparisons for 6 constituents, 18 regions and 5 variables is covered in Section 10. 145

Tide gauges
When comparing the model with tide gauges, we select the closest model grid point if one exists within 11 km. We calculate the model error (model minus observation) for amplitude and phase individually as well as the vector error (taking both phase and amplitude into account) for each tidal constituent. Summing over a number of sites within a certain geographic region, we then compute the Mean of the Absolute value of the amplitude Error (MAE), the Mean Magnitude of Vector Error 150 (MMVE), the mean of the amplitude error and the mean of the observed amplitude (for expressing the MAE or MMVE as a relative error or RE). We use MAE and MMVE in preference to root-mean-squared errors because the MAE and MMVE are less affected by outliers. Outliers are a significant issue, as we will discuss below with reference to Table 1, which lists the sites we have chosen to exclude from the tidal heights dataset. We combine analyses for the six constituents by computing the Root Sum of Squared (RSS) MAEs and MMVEs. In order to estimate the total regional-mean tidal relative error, we also 155 compute the RSS of the area-mean observed amplitudes. These statistics are computed for a number of regions (bounding boxes are shown in Fig. 1) around Australia as well as for the entire country and listed in Table 2. We have not attempted to account for the uneven distribution of the data points around Australia, other than to compute regional means as well as the nationwide means. Nor have we attempted to estimate errors of the observational tidal constituents based on factors such as record length or instrument type. 160

Current meters
When comparing with current meters, we select the grid point for which a penalty function J=D/(5C)+|H_m-H_o|/H_o is minimised, where D are the distances to the model grid point, C are the sizes of the cells, H_m are the model depths and H_o is the onsite depth at the observation point. This is an attempt to mitigate the effect of the model's imperfect topography, by finding the nearest depth-matching (if possible) model counterpart of the observation. We then proceed as for tide gauges, but 165 with the amplitude and phase of the major axis velocity taking the place of height. Errors of the major axis inclination and minor axis amplitude are shown graphically and are listed in Table 3 but are not otherwise included. Three sorts of site-specific relative error are listed in Table 3: 1) the M2 major axis velocity amplitude error relative to the observed amplitude 2 = (| | − | |)/| | , 2) the M2 major axis velocity vector error relative to the observed amplitude 2 = https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License. | − |/| | , and 3) reLF, which has the observed sub-tidal ('low frequency') RMS major axis velocity sub_o 170 included in both numerator and denominator. The first two measures characterise the model's ability to do what it is designed for, which is just to simulate tides. The first of these is for users who need to know tidal range but not at any particular time.
The second is for applications where timing is also important. The third acknowledges that tides are not the dominant component of velocity variability everywhere. Using a tidal model alone (i.e. without a model of other processes) to predict the total current (characterised by maj_o+sub_o) will result in an error determined by sub_o if the tidal error is zero. Where 175 tidal and sub-tidal variability are equal, the upper limit of reLF is 50%. Table 3 lists sites by ascending reLF, and includes averages of the sites with lowest, middle and greatest reLF, for most columns. For the 'm-o' column the average is mathematically an MAE, but with a non-geographic sample of sites. Table 4 is like Table 2, with major axis velocity amplitude and phase taking the place of height amplitude and phase, for the same six constituents. 180

Tidal height
Since we have no reliable, objective (model independent) way of knowing which tide gauge observations (or more precisely, the analysed tidal constituents) are more accurate than others, we have cautiously employed a largely model-based quality control procedure. This procedure excludes sites if: 185 • The absolute value of M2 error exceeds 20 cm and an observed M2 amplitude within 10 km is less by more than 20 cm (excludes four sites) • The observed amplitude is less than 4 cm (two sites) • The observed amplitude exceeds 10 cm and is less than half, or more than twice the model amplitude (14 sites) • The observed and modelled phase differ by more than 90° (six sites). 190 195 With the 20 sites listed in Table 1 excluded, the M2 MAE across 615 sites is 9.3 cm, or 13 % of the mean observed amplitude, which is 72.5 cm. The resulting scatter plot (Fig. 3, note the log-log axes) of model vs observed height amplitude still has points that could be considered outliers; at 5 % of sites the negative errors are ~3 to 10 times the MAE. But we have not excluded these along with the other 20, for lack of clear evidence that they are due to observation error rather than model error.
The bias is relatively small but not insignificant (-2.3 cm, see Table 2). It is negative because most of the biggest errors are 200 cases of the observation exceeding the modelled amplitude. The region with the biggest negative M2 bias (-11 cm) is clearly (see Table 2) the Southern Great Barrier Reef, where the model underpredicts the large tides within about 100 km of the head of Broad Sound The region with the biggest M2 amplitude MAE is the one we abbreviate here as 'Arnhem' (rather than Joseph Bonaparte Gulf and Arnhem Land) but across this region there is a mix of under and over-prediction. The modelled M2 height amplitude is 205 too small in Van Diemen Gulf and the head of Joseph Bonaparte Gulf but too great at many of the offshore sites where the observed amplitude is small.
There are large M2 phase errors (Fig. 4) at many sites. While some are possibly due to observation error, the predominance of positive phase errors, especially in regions of strong tides, points to a problem in the model. The region with the biggest M2 phase MAE is the Kimberley (20°) ( Table 2), nearly twice the all-site average of 11.5°. The significant phase errors are why 210 the Australia-wide M2 MMVE (20 cm) is so much greater than the M2 MAE (9.3 cm).
The next most energetic constituent after M2 (72.5 cm averaged across all sites) is S2 (35.8 cm). S2 has the next-greatest MMVE (12.9 cm, because of large phase errors in the Kimberley), then K1 (with 29.6 cm observed amplitude) has a MMVE of 7.7cm.
Summing over six constituents, and taking both phase and amplitude errors into account, the RSS MMVE across all sites is 215 are Central Great Barrier Reef, New South Wales and the South West, while the three regions with the highest (38.9, 36.4 and 35.6%) are the wide shallow seas in the tropics: Torres Strait, Joseph Bonaparte Gulf and Arnhem Land and Gulf of Carpentaria. Thus, the greatest regional-average relative errors of modelled height are about twice the size of the least. Both are small enough to conclude that the model has skill, but large enough to conclude that there is still room for improvement. 220

Tidal currents
Perhaps the most striking difference between maps of the M2 major axis amplitude (Fig. 5) and the M2 height amplitude (Fig.   3) is that the currents have more small-scale variability, clearly associated with the local topography, as well as the regional variability that broadly reflects the regional variations of tidal range. Characterising and analysing the distribution of the errors https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.
as well as the signal is not straightforward, but is what we will attempt to do, after looking at some of the site-specific results 260 listed in Table 3.
The first line of Table 3 is for a Gulf of Carpentaria site sampled in 1987 by CSIRO. It is the first line because it has the lowest reLF, which in turn is because the errors of the M2 major axis velocity phase and amplitude are both small (0° and 1 cm s -1 ), while the amplitude of the observed M2 tidal currents is large (21 cm s -1 ) compared to the rms sub-tidal velocity (3 cm s -1 ).
Site CW3 (line 2) sampled by Penesis et al. (2020) in Banks Strait is a much more energetic site but the errors of the major 265 axis velocity phase and amplitude are both relatively small (8° and 1 cm s -1 ) nevertheless. It is also a tidally dominated site, (98 cm s -1 for M2 compared to the sub-tidal velocity of just 7 cm s -1 ). As it happens, the error of the minor axis is also very small (both are essentially zero) here, but the error of the inclination is not (-28°T observed but -52°T modelled). Site CW1 (line 3) is about 3 km away (just one grid cell) and has a greater amplitude error (13 cm s -1 ) but less inclination error (3°).
Looking down the table we see that 9 of the 19 lowest-error sites are in Banks Strait. This is clearly a region where the model 270 in its present form is capable of producing current velocity predictions with low relative error (reM2, re 2 and reLF 17, 26 and 30 % at most), so is the first to be discussed in the next section.
At the other extreme (at the bottom of Table 3) is NRSNIN, an IMOS ADCP at the Ningaloo Reef National Reference Site in Western Australia, where the observed M2 major axis amplitude is just 7 cm s -1 . The model estimate is listed as being more than twice this, at 19 cm s -1 . From the prediction point of view, these errors are compounded by there being a fairly large 275 (18 cm s -1 ) sub-tidal variance at this site. Similarly, the next-highest relative error is at GBRLSL, a site off the Great Barrier Reef in 330 m of water where the observed M2 major axis velocity is essentially zero, but the model estimate is 7 cm s -1 . One thing these two sites have in common is that they are over steep topography where sharp gradients are common, so part of the poor agreement is bound to be due to representation error (that error that occurs when you compare a point measurement with an area-average). But even so, these are probably not sites where tidal predictions will be of much practical use. 280 Table 3 includes statistics that characterise model error averaged over sites grouped according to whether reLF is in the lowest, middle and highest third. The MAE over this first third is 7 cm s -1 (an 11 % average relative error), while the MMVE is 13 cm s -1 , a 22 % average relative error or 29 % if sub-tidal currents are taken into account as well. For the locations that these sites are representative of, you could argue that the tidal model is not only useful, but is enough by itself, i.e. a short-term forecast of sub-tidal current velocity would not often make a significant contribution (since its rms value is around 6 cm s -1 , just 10% 285 of the average M2 amplitude). For the middle group the average M2 tidal current amplitude (30 cm s -1 ) alone still exceeds the sub-tidal variability (9 cm s -1 ), but the dominance is less than for the first third and the errors of the tidal model are not insignificant. The average reLF for this group is 57 %, which could be argued as being acceptable, but with there being much room for reduction, either by improvements to the tidal model or addition in near-real time of a skilful forecast of sub-tidal variability. For the final third, the observed tidal currents are mostly insignificant (3 cm s -1 compared to 22 cm s -1 ), so it doesn't 290 really matter what the predicted tidal velocity is, as long as it is weak. This last group includes all 11 sites in New South Wales and south-east Queensland regions, five of the deeper (~100 m or more) sites in South Australia, and all eight of the sites in https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.
south-west Western Australia. We will now look more closely at the regions where tidal currents are a large fraction of the variability.      https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.

Bass Strait (including Banks Strait)
The tide comes into Bass Strait from both the east and west, with the strongest flows (Fig. 6) either side of the central basin (see Fig. 2) where the tidal range (Fig. 3) Table 3). This is also the biggest error in Bass Strait, but it is still quite a small (17%) relative error. Taking the phase error also into account takes this to 26%. Table 4 lists the M2 MAE across the 18 validation sites in 335 Bass Strait as 8.5 cm s -1 . The RSS across 6 constituents is 9.9 cm s -1 , or 14.3 % of the 69 cm s -1 mean observed RSS of amplitudesa much better than average (22% across Australia) relative error. Figure 6 and Table 3 show that, across Bass Strait, the modelled M2 current ellipse eccentricities and orientations are mostly in good agreement with observations. The phase errors range from -10° to 12°. Summing over six constituents, and taking the phase errors into account (Table 4)

Kimberley
The Kimberley region of Australia includes King Sound, where the greatest tidal range in Australia occurs. The entrance to 345 King Sound has such strong tidal currents that tourists go out to see them in RIBs, helicopters and other vessels. There are not, however, any available instrumental records of the flows in the most energetic regions, so the percentiles of the model (across ~30,000 cells, see Fig. 8) are very different to the percentiles of the observations. Figure 8 shows that the model agrees quite well with the seven available records, including the change from nearly circular M2 ellipses at KIM050 to the shore-normal https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License. rectilinear flows at CAM050 and CAM100, and then the shore-parallel ellipses at TIMORS88. The M2 amplitude errors at 350 KIM200 and KIM100 are just 3 and 0 % of the observed amplitude. It is only with the phase taken into account that the M2 relative errors are significant (5 and 12%). The RSS MMVE is 11.4 cm s -1 , or 27 % of the observed RSS amplitude, just 5 points higher than the Bass Strait figure.  Figure 9 shows that M2 velocity errors are relatively low at six of the eight sites in the Darwin-Clarence Channel region. Table   3 (lines 46 and 47) identifies the two noticeable exceptions as being the Darwin-CW1 and C1 sites, where the M2 major axis 360 amplitude errors are 56 and 76 cm s -1 . At C1 the problem is clearly the topography; model depth is only 8 m but the in situ depth is 52 m. It is less clear why the error at CW1 is large but we will not be surprised if rebuilding the mesh using recentlyacquired topography data does not reduce these errors. At present however, the velocity major axis RSS MMVE remains listed as 35.8 cm s -1 , or 44.7 % of the observed RSS amplitude. The modelled tidal height amplitude in Van Diemen Gulf (Fig. 10, see Christine Reef for example) is significantly weaker than the observations, for reasons that we are yet to determine. 365 https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.

Southern Great Barrier Reef
The Barrier Reef is dense off Broad Sound, causing tides to enter the reef lagoon from both the NW and SE. These waves meet in the lagoon outside Broad Sound then further amplification of the wave entering the Sound occurs due to the geometry of 370 the Sound (Middleton, Buchwald and Huthnance, 1984). Our model simulates the first process satisfactorily in a qualitative https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.

South Australia
A distinctive feature of the tides of South Australia is that the amplitude of S2 exceeds that of M2 (barely), leading to a very strong spring-neap cycle. The vanishing semidiurnal tide on days when M2 and S2 are out of phase is locally known as the Dodge tide. Table 2 lists the SA-average observed M2 and S2 height and major axis amplitudes as 25.6 and 26.7 cm, and 4.3 385 and 4.6 cm s -1 . The model M2 and S2 height and major axis amplitudes (not listed) are also nearly equal, at 23 and 28 cm, and 5.3 and 6.2 cm s -1 so Dodge tides will also occur (imperfectly) in model-generated predictions. The maximum modelled M2 https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License. major axis amplitude is 40 cm s -1 in the South Australian region (Fig. 12), but we have no observations to validate the model at that location. The maximum observed M2 major axis amplitude is 9.3 cm s -1 at both SAM6IS and SAM8SG (rows 41 and 63 of Table 3) where the model is in very close and moderate agreement, respectively. The RSS MMVE is 4.2 cm s -1 , or 45.5 % 390 of the observed amplitude. Table 3 lists results for just five sites in the Pilbara region (one being the Ningaloo site mentioned earlier as having the greatest error). Unfortunately, these are all we have in our validation dataset despite the economic importance of marine traffic in this region. Results for the three IMOS ADCPs near 20° S (PIL050, 100 and 200) include M2 vector errors of 7 to 25 % of the 395 observed amplitude. But this region is well known for strong internal tides (Book et al., 2016), to which our analysis method is essentially blind, and thus underestimates the errors. Internal tides aside, the RSS MMVE for this region is 10.7 cm s -1 , or 49.8 % of the observed amplitude.

Gulf of Carpentaria, Torres Strait, Central Great Barrier Reef
The GOC and CGBR regions have intermediate (40.3 and 52.7 %) relative errors of the RSS MMVE, but being based on just 400 3 and 5 sites, these statistics are uncertain. Nevertheless, we see value in publishing tidal current predictions for these two regions, with appropriate warnings, partly because the sub-tidal currents are weak in these two regions. As mentioned earlier, Torres Strait is one of the few places where official tidal current predictions are already published. We have not yet compared those predictions, or observation-based constituents with our model.

South-east Queensland, New South Wales and South West 405
The relative error of the RSS MMVE for the SEQ, NSW and SW regions are 98, 114 and 115 %, respectively, suggesting that the model is not simulating the tidal currents in these regions very well, even though it is simulating the heights (recall that NSW is one of the regions with the lowest relative error of height). These narrow-shelf regions are also where the sub-tidal currents (Table 3) far exceed the tidal currents, so predictions of tidal currents would be of limited practical value even if they were accurate. For both these reasons, we will not be publishing tidal current predictions from the COMPAS model for these 410 regions.

Discussion
We have evaluated the tidal heights in our COMPAS model against a large number (615) of sites around Australia, giving a much more detailed picture than was given, for example, by Haigh et al. (2014) or Seifi et al. (2019), while being broadly consistent. But modelling tidal heights is not the principal motivation of this study. Our focus is on tidal currents (depth-415 averaged at this point), about which much less has been written (Stammer et al., 2014;Timko et al., 2013). Lyard et al. (2020) https://doi.org/10.5194/os-2020-107 Preprint. Discussion started: 11 December 2020 c Author(s) 2020. CC BY 4.0 License.
compare FES2014 with the same IMOS data we have used (just graphically). They conclude that for shelf currents, there is still a need for nested regional models (such as ours), with finer grids than global models have.
We have shown that our COMPAS model of the barotropic tide is in very good agreement with observed tidal currents at many, but certainly not all, of the 95 sites at which we have in situ validation data. A large number of the sites with high 420 relative errors are where the tides are very weak, so it could be argued that those errors are of little practical interest. Over the continental shelf, this is the case for the southern half of the continent from Ningaloo Reef in the west to Fraser Island in the east, excepting Bass Strait and the South Australian gulfs (i.e. the sections where the shelf is narrow). This leaves 79 % by area of Australia's shelf waters as being where tidal currents are both predictable and a significant proportion of the total variance. Bass Strait and the Kimberley region are where our model performs best, with the root sum of square (across 6 425 constituents), regional-average vector error of the major axis velocity being less than 27% of the observed signal. This measure of the relative error of the model's tidal predictions is between 40 and 50 % in the other regions where we think the predictions should be made available to the public.
We hope to expand our tidal currents validation dataset, especially at locations (mainly in the NW) where observations have been made by offshore industries, in order to guide development of the next version of our model. Incomplete as it is, we are 430 publishing it now along with the output of our model because we are sure it will have enduring value, for example, to developers of global models such as Lyard et al (2020) who used a preliminary version of the validation dataset.
It is well established (e.g. by Ray et al., 2011) that accurate topography is an essential component of a good tidal model and our results and those of Sahuc et al. (2020) bear this out. Some of the largest model errors are where there is a big discrepancy between the depth in the model and the depth that was recorded on site during mooring deployment. Improving the topography 435 in our model is certainly a priority for future model development. This will likely comprise a combination of inverse tuning where local bathymetry alterations are made to optimally correlate model predictions to observation, and capitalising on the results of the ausSeabed initiative (http://www.ausseabed.gov.au/about).
Boundary conditions are also, of course, an essential input for a regional tidal model. We have only tested our model using open boundary forcing from TPXO9v1, but hope to test it soon using TPXO9v2 and FES2020. 440

Conclusions
We have shown that for many regions around Australia's continental shelf, our model can predict depth-averaged tidal currents with enough accuracy to arguably be operationally useful for mariners and maritime industries. Regions where tidal currents are most predictable and in excess of non-tidal currents include Bass Strait, the Kimberley, Joseph Bonaparte Gulf to Arnhem Land and the southern Great Barrier Reef. Consequently, these are the regions for which we intend to commence publishing 445 'unofficial' predictions of tidal currents (both model-based and observation-based). They are also the regions of greatest interest to the renewable energy sector, for whom we have published maps based on the model discussed here. We intend also to publish tidal current predictions for the South Australian gulfs, the Pilbara, Gulf of Carpentaria, Torres Strait and the central