This document is written as a summary of the discussions in the kick-off meeting, but includes a more detailed discussion about the strategy we agreed upon. The document is therefore intended to be explanatory rather than compact. So this document should rather be seen as a more detailed project plan than the contract. Our discussions also worked as a catalyst for a few new ideas already included below.
Another reason for writing this document fairly detailed is that the PDF file with the presentations from the meeting will most likely be subject of quite some questions without further explanation. The presentation can be downloaded at http://www.ucc.ie/serg/honeymoon/downloads/HONEYMOON_kickoff.pdf
The HONEYMOON system will be designed as a partly central and partly de-central system. The aim was originally to go for a fully integrated de-central project and system. The integration issue was two-fold, maintenance and political, but what we have lost in maintenance is gained by having the system run central. The political consequences of a central solution are hard to evaluate. Therefore we will keep the design in such a way that the system can be installed partially or fully decentral. Please note, that the numerical weather prediction model used in the HONEYMOON system has also been changed to ensure that a decentral solution can be offered in the future. In this respect it is important to note that only by changing the NWP model, the original strategy of the HONEYMOON project can be followed. The NWP model is one of four modules in the system and hence the original description of the project to be ``designed in a Plug-and Play mode for highest flexibility of the work packages'' is not affected by this change. The details of this change are described in the NWP section of this document.
HEPS and SEPT will be run in INM in the real-time demonstration. SEPT will be developed by UCC and INM's is focusing on HEPS.
The purpose of these two components is identified by the predictability of the weather pattern. SEPT will produce typically 3 hourly values of the predictability of wind speed weighted with a standard power curve. The horizontal resolution of the predictability doesn't need to be in higher resolution than around 1.0 deg. The background for only providing the predictability is that the steep power curves blow up the uncertainty such that the absolute values are not very useful. The HEPS system will also be very different in comparison to the deterministic model. Hence, it will not be simple to translate wind power between the detailed deterministic forecast and the ensemble forecasts. A multi model ensemble system will for sure have different Weibull distributions, which would be required for every site. These Weibull distributions might even have an annual variability. It would therefore not be within the scope of the project to base the power prediction directly on the ensemble.
The goal and assumption therefore is that the deterministic forecast gives the best prediction measured over long time. The prediction error however varies with time and must be predicted from HEPS with SEPT. In practical terms this means that we expect that backup power to be required to compensate for the expected deterministic forecast errors.
We aim for running SEPT every 12 hours based on the assumption that the predictability of the atmospheric states predictability does not vary strongly with time. The background for this assumption is that the predictability is lying in the weather pattern and it is on a large scale. So if a 12UTC+48H and a 18UTC+42H valid the same time differ significantly, the ensemble spread should be large at 12UTC+48 and hopefully also 6 hour before. This assumption holds mostly for the ECMWF ensemble and a short range ensemble system should have this as a target as well. There is in addition a technical reason for using only 00UTC and 12UTC, which is that this strategy also allow for using the ECMWF ensemble in case HEPS is running into a delay. We must take this precaution because HEPS depends on several factors that INM is not fully responsible for, such as the computer delivery and access to boundary data from other Met Centres. Strictly speaking it would be more optimal to run HEPS on 06UTC and 18UTC for a wind energy purpose, because the observations over the Atlantic from AIR-craft is likely to give a better spread the prediction for the following day in a 3Dvar or OI analysis based forecasting system. However the ECMWF 4Dvar system is in any case using as many AIR-craft observations over the Atlantic as feasible, so in that case it does not matter that only 00UTC and 12UTC ensemble forecasts are available. The output frequency of the predictability generated by SEPT must be at least 3 hourly although it is poor resolution of the reduced predictability that follows fronts. However the ECMWF EPS is at present only available in 3 hourly time resolution.
Historic power data from the end-users in the project will be fed into PCAT, which is seeking for 1-6 hour forecasts valid at the same time and at the same point. PCAT will take wind speed and direction into account. PCAT also must make sure that the forecast quality is OK to include the observation in the power curve estimate. Because of the confidentiality of power data and the fact that the HONEYMOON system should not only be useful for transmission system operators, PCAT is extended to also be able to produce power curves with grid averaged wind speeds and power data. In the case of a supplier that doesn't have observations himself, the information of the installed capacity and the type of turbines in the relevant grid boxes together with the manufacturer's power curves should be enough to produce power curves with PCAT.
The use of wind speeds in PCAT also needs some explanation. We cannot realistically use higher than hourly forecast values. Technically it is possible, but we can not justify higher resolution from an accuracy point of view. However, power data is often available in higher time resolution such as 5, 10 or 15 minutes. The model time step will be less than 30 seconds. During one hour we therefore have at least 120 wind speeds. This is then averaged with a standard power curve instead of a linear weight function. This method takes account for the non-linear effects of varying wind speed. With a mean wind speed of 11 m/s and an amplitude of 5m/s for the departure from the mean, a considerable drop in effective power is observed compared to a constant wind of 11m/s. The same arguments apply with a mean wind speed of 7 m/s, which leads to an increase in power compared to a constant wind of 7m/s. The model therefore produces a wind speed that is constructed for evaluation in power space and not be compared with wind speeds. The direction is included in this construction. However, it is best written out as normal wind components, internally averaged in power space and including air density. We refer to this as the power-linear wind speed (plws). The difference between plws and linear wind speed (ws) is mostly small. The infrequent systematic errors in ws are however considered large enough to to test and compute plws to find out its real impact. It is also different compared to what others are doing. The power curve estimate should be more accurate with plws.
The power curve estimate process is iterative, so that the estimation process excludes the periods where the forecasts and observations disagree too much. The first iteration can run with a high error tolerance and then gradually reduce the error tolerance in the next iterations. It is in most cases best to estimate the power curve direction-dependent , then compute a direction-independent part with even weight on all directions. Finally a direction-dependent correction is applied to the mean in bigger wind speed bins. This is mainly because some wind speeds occur very seldom from certain directions. The estimate becomes unreliable, if the wind speed bin is too small. The direction-independent part of the power curve is suggested to be in 0.5m/s bins, while the direction-dependent part should be a correction added to the direction independent power. It is enough to compute around 5 values for each direction bin, if these are all functions of the relative power given by the direction-independent curve. It sounds rather complicated, but it allows for packing detailed power curves in a very comprehensive and compact form. It is more clear by looking at a formula:
Pideal(plws,direction) = Peak * ( P(plws) + C(P(plws),direction) )
where Peak is the maximum power, P is interpolated in a table with plws in 0.5m/s bins, C interpolates in 5 bins of relative power and 8-12 direction bins. The tables used by P and C are generated from a 2-dimensional table with plws and direction in 0.5m/s and 8-12 direction bins. These two matrices can include all turbines in the grid box or a representative average. This is to deal with the fact that we might not have access to all turbines output, but be able to estimate a representative average. An end-user might also want to see the production from other farms in the area than his/her own, so it might be an idea to use a set of power curves for the different groups of turbines.
PCAT does not need to be running automatic in the project since we assume that once a power curve is made it will not change. We assume that no other buildings or turbines will be put up close to the farm. Another reason for PCAT not running automatically is that in a lot of cases the power data end-users are based on windows databases that are not accessible automatically from LINUX. It is worth noting, that this also had impact on the PCAT programming strategy, because it excludes the use of a Kalman filter strategy to be applied in PCAT, because there will be no automatic feed in of new data.
The power curve data will be send to the central installation system installed in UCC. We must expect that very few customers want PCAT running centrally because of data confidentially reasons. The power curves generated by PCAT don't reveal much about the turbines, because they include the model bias statistics at that particular site. The output power curves will therefore not be of value for anybody. We can go one step further in making the PCAT power curves anonymous. A horizontal sum of all turbines in the model grid box will completely eliminate the possibility of abusing the power curves. This holds of course only for grid boxes with more than one turbine or farm. The horizontal summation means that all turbines inside the grid box will feel In the project we need however to use the area summed data and transfer this to the central installation to prevent that PCAT should be rerun again and again at the end-user until the final PCAT version is ready.
The efficiency based power parametrisation is closely related to PCAT. However, in the estimate of the power curve, PCAT will have to ignore observations that differ too much from the power curve. This is in fact what is most interesting in this task. We do not know whether it is the model's inaccuracy or whether a phenomena occurred that changed the power output in a way so that the power didn't fit the average power curve in this particular case. A typical example could be a 45-90 deg. relative sudden shift in the wind direction, where it will take a while before all turbines have adapted to the new direction. Other examples are days with strong convective precipitation, wind gusts and shorter calm periods between the rain. Also very stable conditions with a strong wind shear could be a challenge.
The most handy formulation of the parametrisation seems to be
Peff = Pideal(plws,direction) * Coef(
wind shear,
direction changes,
static stability,
strong rain,
ice up
)
where Coef is normally in the interval between 0.7 and 1.0 with an average value of 0.95. All the terms in Coef are model grid box variables. The local effects of the wind shear are included in Pideal. So the assumption is that Coef should only contain the effects that generally have impact on the power production.
The major question is, if we can see the effect we expect. The next problem then is, if we can predict the phenomena accurate enough to use it in prediction mode. Suppose that we can parametrise the efficiency well in the 1-6 hour forecasts, but the accuracy of the same parametrisation based on a 24-48 hour forecast is low, then it is still worth to include it. Such effects of the parametrisation will mostly give a positive signal in an area integral, if the turbines are spread over a reasonably large area, but a negative or neutral signal on a single site. If a strong frontal zone hits the target area with a correct arrival time, but the turbines only adjust to the new wind direction after 30 minutes, the forecasted power will be overestimated for half an hour although the wind speed is correct in the model.
So the purpose of the efficiency based power parametrisation is to correct the computation of the power when the circumstances are not optimal for the turbines. The correction is required, because it is occasionally seen that a forecast of the wind is near perfect, but power is not. So as the prediction of wind becomes better, the more important this module becomes.
DMI will in 3 or 4 month toward the end of the project deliver analysis and boundary data for the coupled NWP forecast system. The data will be delivered in 6 hourly frequency.
DMI is conducting a gust factor analysis that mainly focuses on the cutoff wind speed. The question is also how accurate is this forecasted wind speed. The outcome of this analysis might be used to improve the vertical diffusion. The generation of direction dependent roughness climate files will be done by DMI. The impact of the direction will be of most interest for the efficiency base power parameterisation.
In the meeting we have learnt that the electricity grids over Europe are working rather different. It became only clear toward the end of the meeting that a central installation of the system is not the most clever solution for the UK-grid for example. The UK market is so dependent on the very short-range that traditional weather forecasting has very difficult conditions and more focus on nowacasting is required. The other extreme is Germany, where the planing is more long term and where a feed-in-law is active. The actual system implementation must therefore be very different and also focus on different forecast lengths. The UK system for example is more dependent on fresh observations than the German system will be.
The different focus on forecast lengths has impact on the verification of the system. It seems to be necessary to make a market survey and find out how we verify the system in such a way that there will be a large group of potential end-users.
If we assume that the future lies in trading of certificates and it is the black suppliers responsibility to buy enough of these, then the end-user might be a such a supplier, that doesn't know any details of the turbines and their location when the certificate is purchased. This has impact on PCAT, which is originally thought as being based on statistics only.
If we assume that the future customer is a TSO, then we can rely on that at least the input and output of the electrical grid is known. Therefore PCAT will rather be used to generate power curves for areas than individual turbines. These two examples show the importance of the extension in the design of PCAT to meet the requirements also in these two cases.
From this discussion we can see that the DSTAT module plays an important role in the marketing of the Honeymoon system.
A deterministic down-scaling model for wind energy purposes must include certain features that traditional NWP models from Met Centres do not have. This has become clear over the past year, where many circumstances forced us to reconsider the strategy described in the original proposal. Together with the fact that it became clear in the contract preparation phase that the HIRLAM system cannot be used outside a met centre of the HIRLAM Group in real-time, it can be stated now that a system as it is needed for the purpose of wind energy forecasting is not readily available anywhere. It requires therefore a programming from scratch of the dynamic core and interfacing to the physics. The disadvantage of a total programming from scratch is that only the sum of all changes relative to a more traditional forecast can be diagnosed although it will be interesting to implement and test them individually. The project's time frame does however not allow for such a careful strategy. The scientific arguments for the approach are so strong on the other hand, that it would be too much a compromise not to follow up on the approach. Given the progress of this programming we estimate 1 year for the task in total and we are already 7 months ahead in this process.
The change of the deterministic NWP model for Honeymoon was not discussed in detail. As explained, the reason was that these are meteorological and would have required one more afternoon allocated to the meeting. In the following a detailed description of the new model is given in the light of the HONEYMOON system and also from a meteorological point of view.
As we have discussed in the meeting, the deterministic forecast with the new model will be run centrally in the demonstration phase. This model is especially developed for the purpose of giving accurate boundary layer winds in the most efficient way. It is designed to also be run de-central at the end-user after the project.
The model system is based on a isentropic hybrid scheme using a time split scheme for the gravity waves, but a Semi-Lagrangian approach for the advection terms. The model has spectral boundary relaxation in the free atmosphere, no boundary relaxation on the outflow zone and a second order time scheme in the boundary layer inflow zone, which is local. No horizontal diffusion is required apart from the inherent damping in the advection term.
The Semi-Lagrangian scheme is used in a simplified way. The maximum time step permitted will be similar to the Eulerian formulation. One of the reasons why we refer to this scheme as a simplified scheme is that the surface pressure tendency is fully explicit and Eulerian. We must preserve accuracy with relative steep orography, which is not possible with a traditional Semi-Lagrangian scheme. Using long semi-implicit time steps, apply tricks in the continuity equation and perhaps too short waves in the model orography can lead to unrealistic flows around mountains in traditional Semi-Lagrangian implementations. It is therefore safer to handle the continuity equation with a time split scheme. The worst inherent damping characteristics of the Semi-Lagrangian schemes are due to the combination of long time steps and interpolation. The damping in the new efficient short time step scheme therefore mainly works on the very short numeric waves. This simplified scheme allows for a more efficient computer algorithm, which limits the number of times each advected variable are copied from 2nd level cache to primary cache. The cost of the scheme is therefore only slightly more expensive than an Eulerian time-split scheme. The use of an explicit or also called time-split scheme allows for more complicated boundary relaxation schemes than the semi-implicit time scheme. In addition, the more accurate phase speeds of short gravity waves should be a slight benefit. The frontal phase speed is of major concern in wind energy and the contributes most to forecast errors after 18 hours integration. It is also a fact that explicit schemes are in theory more accurate than implicit schemes. So alone this issue is enough argument for using a time split scheme. The phase speed is traditionally accepted to be better in the Semi-Lagrangian than in the Eulerian schemes. A major reason why the Semilagranian method is chosen was the importance of phase errors in wind energy.
Another benefit in the Semi-Lagrangian scheme is that the strong shortwave damping results in a more accurate advection near the surface. It is worth noting that the short waves are undesirable. Although they might sometimes look like eddies, they are not. And the attempts from the model dynamics to generate such waves or eddies are due to imbalances mainly in the hydrostatic pressure gradient term, which is very different from the real nature of the boundary layer eddies. The most efficient way to handle this is to have a relative strong divergence damping in the momentum equation in the boundary layer. In this way the spurious tendencies are dampened, which are often observed near the surface in high resolution. These have a potential to either trigger convection or to lead to random output values, unless a proper time filter is applied to the output data. The real effect of boundary layer eddies however needs to be handled by the vertical diffusion alone.
The isentropic hybrid coordinate is not easy to combine with Semi-Lagrangian vertical advection. This is however not required, because the surfaces are closer to material surfaces and the vertical velocity near the lower boundary is also not very high in the terrain following coordinate. Therefore, the vertical advection is expected to be sufficiently accurate and stable to be handled Eulerian with central differences. The need for horizontal diffusion in the Eulerian advection schemes however tends to become problematic in increased resolution. The many local orographic extremes are catalysators of vertical transport in height coordinates. The traditional means of reducing this vertical transport problem in the Eulerian case, such as pressure level diffusion or reduced diffusion in the boundary layer, are not very convenient in high resolution. By using an explicit Semi-Lagrangian scheme, we avoid the horizontal diffusion, we limit this systematic transport without introducing the drawbacks of the semi-implicit Semi-Lagrangian schemes that cannot handle the complexity of the flow. The Semi-Lagrangian explicit scheme will however also generate systematic vertical transport, but it is possible to include Semi-Lagrangian vertical advection in the non-isentropic-hybrid levels near the surface to compensate for the systematic vertical transport from the horizontal advection and thereby eliminate the problem.
The physical parametrisation schemes will be selected of the schemes available in MM5. The actual choice will only be taken after a significant amount of experiments throughout the project. The philosophy is however that a simple scheme is chosen rather than a complicated scheme. The physics is also handled with a time-split scheme.
As mentioned before, the system is deterministic. Therefore, it is planned to use well tested schemes to keep the same schemes over time. The corrections should rather come from the statistical coefficients generated by PCAT. These are dependent on very long-term statistics, such that it will not be beneficial to change the statistical properties of the model system once the system is running in real time (see section on PCAT). Therefore, the model system is programmed to a level that should be competitive for many years and applicable in a variety of resolutions.
The model is designed such that global data are allocated in slabs. This allows for an irregular physical grid, where the density of the grid points are higher over land than sea. This is a computational efficient way of gaining resolution, because the physical processes allow for longer time steps if the dynamic tendencies are not too spurious. The drawback of this approach is that thermally driven circulation such as sea breezes can not be predicted explicitly in the physical grid, but only in the dynamical grid. There is no simple solution to overcome this problem. However we should not worry too much about this problem because it is a special problem, which will only occur along complicated coast lines, but the strength of such sea breeze are strong enough to give full power production. If a major wind farm is build in an area, where such thermally driven forces exist below the dynamic resolution, the most computational effective way of solving the problem is to solve a balanced set of equations every hour with the tendencies and atmospheric state averaged over one hour. This is sufficiently accurate, because sea breezes last over a few hours. The alternative if is of course to spend the extra computer capacity for the job. It is rather easy to gain overview from a map with installed wind energy capacity if this is an important phenomena to simulate or not. If it turns out not to be of importance now, it can however change over time because of the exponential growth rate of the wind farms.
The electrical integration of wind power generated from sea breezes is another problem to overcome because it is a short term phenomena in the late afternoon.
The model is constructed for downscaling and focuses primarily on the winds in the boundary layer. It is expected that the model system performs well with around 20-24 levels. In the upper atmosphere, the model resembles however the boundary generating model. In this way, we can assure that the model will benefit from development of new parametrisation schemes in any boundary generating model, but will not be affected in the lower boundary.
The basic idea behind this model construction is to achieve a more smooth high resolution compatibility. This applies both on meteorological relevant scales and on the pure numerical noise level. The use of a Semi-Lagrangian scheme in the boundary layer applied on variables that are closely connected to the high frequent waves is an important way of damping noise. The disadvantages is that the convenient energy conservation feature in the energy conversion term is lost. The loss of formal conservation properties suggest also to use an Arakawa A grid. It is inconvenient not to have u,v in the same grid-point, but affordable to loose resolution in the divergence equation.
No real attempt is done to improve the upper levels. The numeric scheme, the isentropic hybrid surfaces and the boundary relaxation should be seen as a consequence of the aim to representing the large scale motion of the boundary generating model as much as possible. The main reason for using isentropic hybrid vertical coordinates is not to only get higher effective resolution. It is rather the accuracy of the scheme in connection with low-level fronts and inversions that makes it more suitable. The isentropic schemes give a more accurate description of the pressure gradient term through the better discretisation of the hydrostatic equation during these conditions. The changes in the wind power can be significant under these conditions. Even though some fraction of these case are predictable or regarded as such, it is felt important to incorporate features that improves this kind of weather regime.
It is also important that the system is compatible with other model systems. This gives in reality strong restrictions on the prognostic variables to be used. Cloud water and turbulent kinetic energy are examples of variables that might be present in the boundary files that are used, but they are not conservative enough to be used in boundary relaxation processes, especially because their vertical coupling is usually stronger than their horizontal coupling. For that reason the prognostic variables of the system must be limited to only u,v,T and q.
The design of the model system is therefore optimised for predictions of variables in the boundary layer in very high resolution, but still without any attempt to introduce non-hydrostatic effects. The correlation between predictable wind power and strong non-hydrostatic effects must be expected to be low. The non-hydrostatic effects that could be of interest for wind power are mainly extreme downdrafts in connection with thunder. Such down-drafts can cause turbines to switch off. The detailed air flow near steep orography might also not benefit from a non-hydrostatic approximation level, because the vertical velocity is a slave of the lower boundary condition and horizontal wind speed. The small changes due to a non-hydrostatic approximation level occur therefore in the unpredictable part of the weather spectra, and should therefore rather be diagnosed in an ensemble system. Another small complication is that the non-hydrostatic balance implies less radial extremes in average. This is then in agreement with the suggested strategy in the unpredictable cases, where it is recommended to smoothen out the prediction to minimise the error with ensembles. Therefore, a hydrostatic system is considered sufficient for the task of predicting wind energy even in very high resolution (e.g. 1.5km).
Only Linux on Pentium 4 or any other successors in the IA-32 processor family is supported. Parallelisation is only supported with MPI. The advantage of the dynamic scheme is that no global halo zone is required. There are only three processes that couple the system of equations. The advection of the prognostic equations requires one halo point and fetches it when required. The second process is the gravity wave equation. Only 3 prognostic multilevel variables need to be swapped to the neighbour-processors with one halo point and one single level field (ps) is swapped with two halo points. The last atmospheric process, which couples horizontally is the spectral boundary relaxation. The irregular physical grid requires also global communication before and after every physical time step. It uses not necessarily global communication, but from a programming point of view it is global. The construction of the message passing scheme makes it easier to handle all other model processes, because the basic model grid has no overlap with other grids. These are benefits that come with the use of the Arakawa A grid and explicit gravity wave handling. The system will also use many Fortran90-features and therefore requires a full F90 compiler.