1. Introduction
The use of data in recent years has increased dramatically and is continuing to do so with rising technologies such as artificial intelligence (AI), machine learning, blockchain, cloud gaming and augmented/virtual reality being deployed at a vast rate and scale. As well as these new emerging technologies existing technologies such as mobile data, video streaming and crypto mining are all still rising in data consumption. These activities are all facilitated by data centres located across the globe and collectively as an industry the energy consumption from data centres is increasing at a rate and scale never seen before. The global data centre capacity demand is estimated to be around 50 GW as of 2023 and is predicted to rise to 200 GW by 2030, with AI accounting for 15% of this demand in 2023 and predicted to account for 40% by 2030 [
1]. The energy consumption associated with this global data centre capacity demand is estimated by the International Energy Agency (IEA) to rise to 1000 TWh by 2026 from 460 TWh estimated in 2022 and is expected to rise from 1% to 6% of the UK energy demand by 2030 [
2,
3]. Some studies even suggest the global energy demand for data centres could increase to 8000 TWh by 2030 [
4].
The challenge faced is how to sustainably deliver energy to an industry growing at such a rapid pace. The long-term strategy has been to increase the renewable energy capacity globally, for which the data centre industry leaders such as Google, Microsoft and Amazon all are massively contributing with power purchase agreements (PPAs) covering 100% of their annual energy consumption, respectively [
5,
6,
7]. Power purchase agreements are contractual arrangements between a generator and a consumer in which an agreed amount of energy is procured at an agreed cost for an agreed period of time in advance of the energy generation. However, insufficient investment in grid infrastructure globally is leading to increasingly constrained grids which is starting to slow the progress of the deployment of renewable energy, with over 3000 GW of renewable generation projects awaiting grid connections as of 2023 [
8]. These constraints are also inhibiting the construction and operation of new data centres globally [
9].
This paper presents a solution to this challenge by deploying hybrid renewable energy systems for providing off-grid energy to data centres. This approach uses wind, solar, gas and batteries to provide reliable and sustainable energy to data centres that cannot obtain a connection to local grids. Each technology is integrated into a microgrid electrical network, which will have an energy management scheme facilitating the automatic operation of the system. The batteries will be used in a grid-forming control mode stabilising the system voltage and frequency. The batteries will also be sized such that the system’s maximum and minimum load steps can be accepted by the system without causing the voltage and frequency of the system to fall outside of the data centre’s operational limits. The wind generation will operate in synchronisation with the battery system at maximum output unless curtailed by the energy management system. Solar generation will be connected to the microgrid via multiple inverters which will be in synchronisation with the battery system inverters but will not control the microgrid voltage and frequency. Gas generation is used as peak lopping and will be issued real and reactive power set points from the energy management system based on the deficit power between the renewable generation and the data centre demand. The gas generation will have sufficient capacity to serve the entire demand of the data centre, eliminating the need for backup diesel generators as required for a data centre to achieve its uptime tier rating.
The economic feasibility of an energy generation scheme is typically quantified by the levelized cost of energy (LCOE), which is the cost per unit of energy generated considering the capital expenditure (CAPEX) and operating expenditure (OPEX) of the system. Previous works presenting hybrid renewable energy systems for off-grid energy do not consider the economic impact of energy generation and consumption mismatch and the effect of curtailment. This paper presents a levelized cost of energy utilised (LCOEu) metric to quantify this impact, which considers the cost per unit of energy consumed.
This paper presents five case studies located in the five key data centre hotspots in Europe, locations which collectively account for over 85% of Europe’s data centre demand, often referred to as FLAPD consisting of Frankfurt, London, Amsterdam, Paris, and Dublin [
10].
This paper recognises that time is critical in bringing data centres online and that developing renewable energy generation plants can often take much longer than developing and constructing the data centre. This is often caused by prospective renewable generation plants facing challenges during the planning phase. For this reason, this paper considers the economic and sustainability impacts of operating a gas-only generation solution which can facilitate the operations of the data centre whilst the renewable energy generation schemes are in development and construction. Further to this, implementing a hybrid renewable energy system requires a large CAPEX unlike utilising a PPA for energy procurement, and it is typical that during the initial operating years, the data centre demand would be small in comparison to its capacity. Therefore, there is an economic incentive to shift the installation of renewable energy technologies further into the operation lifetime of the data centre and support the data centre load using gas generators and battery technologies during early operation in order to obtain maximum economic benefit.
The five case study hybrid renewable energy systems presented within this paper are simulated using EMD International WindPRO 4.0 software. WindPRO is a leading software for the planning of wind farm projects and has modules for optimising wind farm layouts, solar farm layouts and hybrid systems. For each system, the levelized cost of energy (LCOE) and emissions are compared against the costs of an equivalent local grid-connected system, onsite gas-only generation, and renewable power purchase agreements.
The design criteria for a hybrid renewable energy scheme are typically based on two definitions: loss of power supply probability (LPSP) and levelized cost of energy (
LCOE) both for a given load profile [
11]. LPSP is the statistical possibility of the power generation system being unable to support the load due to a lack of generation caused by intermittent generation with insufficient storage and/or undersized dispatchable generation. The LPSP for data centres is zero, in order to maintain the data centre industry standard uptime merit of 99.999% and the
LCOE is to be as low as possible. Iverson et al. and Lian et al. both present optimised hybrid renewable energy systems for data centres using these metrics [
12,
13].
The capacity of each technology within the hybrid renewable energy system is critical to the reliability of the system and the commercial viability of the project. Therefore, each technology must be optimally sized to ensure that the data centre energy demand can be met at the minimum total cost.
2. Design and Methodology
This section details the design and methodology for the various hybrid renewable energy system solutions for providing energy to a hypothetical hyperscale data centre in each of the FLAPD markets. Hyperscale data centres vary in capacity anywhere from 20 MW to over 300 MW. This paper considers a hyperscale data centre with a capacity of 50 MW.
The load profile of a hyperscale data centre can vary significantly depending on the end user application. The demand could be that of internet service providers with daily fluctuations or, if the end-user has a constant workload for applications such as AI training, the demand could be flat. This paper considers a flat load profile. Fluctuations due to cooling are neglected as the total cooling demand in hyperscale data centres typically represents less than 10% of the total energy demand [
5,
14]. Further to this, changes in external temperature would have only minimal effect on hyperscale data centre cooling demand as they utilise more advanced and efficient cooling then traditional data centres.
2.1. Levelised Cost of Energy
The levelized cost of energy is the average cost per unit (MWh) and is the primary economic metric for analysing the cost of energy. This is calculated by taking the total lifetime costs and dividing by the amount of energy produced over that lifetime.
LCOE is the levelized cost of energy (GBP/MWh)
A is the annual total cost (GBP)
t is the year of operation
n is the lifetime in years
is the annual energy produced (MWh)
i is the interest rate (%)
The interest rate is set at 3% within this paper, as advised within Ofgem cost–benefit analysis guidance [
15].
The standard formula for
LCOE does not allow for a fair comparison of the proposed hybrid renewable energy systems with gas-only generation solutions and grid-connected equivalents. This is because the hybrid renewable energy system can produce more energy than is required by the data centre which if following Equation (1) will lead to a reduced
LCOE. This additional energy can be either stored or curtailed by the hybrid renewable energy system. To give a fair comparison the levelized cost of energy utilised (
LCOEu) is defined. This is calculated by taking the total lifetime costs and dividing them by the amount of energy consumed by the data centre over that lifetime. The value of
LCOEu will be higher than
LCOE.
To allow comparison of the costs of hybrid renewable energy system and equivalent power purchase agreements (PPAs) the
LCOEs,w is defined, which considers the
LCOE of only the solar and wind component of the hybrid renewable energy system. This is because the PPAs that data centres utilise comprise only solar and/or wind generation and do not consider the battery and gas generation as deployed within each proposed hybrid renewable energy system. This is a limitation of PPAs as they do not consider the cost of energy when the renewable resource is not available.
LCOEs,w is calculated by taking the total lifetime costs of the solar and wind components and dividing them by the amount of energy generated by the solar and wind generation over that lifetime.
LCOEs,w is the levelized cost of solar and wind energy (GBP/MWh)
is the annual energy produced by solar and wind (MWh)
2.2. Technology Choice
Renewable energy alone cannot provide reliable power for data centres as the sources are intermittent. However, a hybrid scheme with a combination of multiple renewable energy sources accompanied by an energy storage scheme can improve the reliability of the renewable energy scheme and reduce the reliance on onsite generation from oil and gas or utility. In the instance where the hybrid renewable energy system operates without connection to the local electrical grid, this system can be described as a microgrid. The choice of renewable energy technologies is an important consideration for a hybrid renewable energy system, with solar, wind, hydropower and biomass power being the most common. Within this paper, only solar and wind are selected as viable technologies. Hydropower is disregarded as it cannot be adopted near any of the data centre hotspot locations due to a lack of resources. Biomass is also disregarded due to challenges in the reliable availability of resources [
16]. Wind and solar are chosen as they are well-developed technologies with low
LCOE across Europe [
8]. Having a combination of technologies allows for each technology to potentially compensate for the other in times of low generation reducing the reliance on dispatchable generation and lowering energy storage requirements.
The choice of energy storage technologies is also an important consideration for a hybrid renewable energy system, with batteries and hydrogen being the most suitable. Iverson et al., Haddad et al., and Clúa et al. all present hydrogen storage in their hybrid renewable energy systems with Haddad et al. having batteries accompanying the hydrogen [
12,
17,
18]. The hybrid renewable energy system presented by Clúa et al. is supported by the grid so that the electrolyser operates at constant power to maintain high purity hydrogen. If the grid is not as ‘green’ as the hybrid renewable energy system then the environmental benefit of using hydrogen for energy storage will be reduced. Hydrogen is more suitable for long-term storage compared to batteries with Microsoft promoting a ‘first of its kind’ 3 MW polymer electrolyte membrane (PEM) fuel cell backup power generator for data centre applications [
19]. Within this paper, batteries are chosen as the energy storage technology as they are suitable for utility scale-up, unlike hydrogen. Battery energy storage is also a more mature technology than hydrogen and has lower costs and maintenance requirements.
2.2.1. Wind
The selection of a suitable wind turbine to provide reliable, resilient, economically viable, and environmentally sustainable power is site-specific and is dependent upon the wind resource at a given site. For the selected 50 MW demand, no single turbine is sufficient and so a wind farm is required.
There are a variety of considerations when developing a wind farm, including site resources, site infrastructure, legal obligations, and turbine layout. The site resource is dependent on location, with factors such as surface roughness, height above sea level and interference with other structures affecting the wind speed and the available power. In this paper locations are fixed near the data centre hotspots and the resource is assessed using meteorological data available from nearby measurement sites. Site infrastructure such as access for vehicles for turbine transport is not considered within this paper. Legal obligations such as shadow flicker, noise and effect on bird migration are also not considered, although it is appreciated that these concerns could be a barrier to allowing wind farm development for the data centre.
Wind farm layout is an important consideration as the wake of a wind turbine can affect the available power for nearby turbines. Traditional layouts are such that they are uniform with distance ellipses between each turbine, typically with a radius of seven times the rotor diameter on the major axis and four times the rotor diameter on the minor axis, with the major axis of the ellipse pointing in the direction of the mean wind direction as shown on a wind rose [
20]. Meyers and Meneveau challenge this approach by presenting a mathematical method of determining the optimum non-elliptical wind turbine spacing considering the annual energy yield and cost of both the land and turbine [
21]. Their conclusions state that the average optimal spacing between wind turbines is 15 times the rotor diameter, more than double the factor of seven that traditional layouts suggest.
WindPRO software has been used for optimising the wind farm layout and results are compared with the methods presented by Meyers and Meneveau. WindPRO has been selected as it includes a vast database of meteorological data globally and available wind turbine models. WindPRO calculates the annual energy production (AEP) for a single turbine using wind speed statistics from meteorological data and the chosen wind turbine power curve. The AEP for the wind farm is then calculated using an improved N.O. Jensen wake model. WindPRO then optimises the wind farm layout using a proprietary ‘smart’ iterative method.
2.2.2. Solar
Solar power is an intermittent sustainable energy source using photovoltaics, which generates direct current from semiconductors being illuminated by photons from the sun’s light. The energy that can be produced by a solar farm depends on the solar irradiation as well as the solar farm layout considering tilt angle, azimuth angle, row spacing, shading and albedo of the ground.
Optimising the solar farm layout depends on the site location with respect to the position of the sun. WindPRO has been selected for the solar farm optimisation as it allows for all technologies within the hybrid renewable energy system to be modelled using only one software package eliminating the need to integrate the results from multiple simulations. It is recognised that other software packages such as PVSyst version 8 could be more appropriate for obtaining more detailed and comprehensive results.
The WindPRO analysis assumes that the tilt angle of the panels can be optimised continuously. Tilt control is often not implemented due to maintenance requirements but, as this paper is considering a hybrid renewable energy systems solution for a mission-critical application, there will be operators onsite 24/7. Therefore, regular maintenance checks are less of an inconvenience in comparison to remote solar farms which are often left unattended. Including tilt control has a relatively minor impact on the AEP of each solar farm with the a difference in the WindPRO of circa 5%. The solar azimuth angle and row spacing were optimised using the WindPRO solar optimiser. This can provide optimisation for either maximum annual energy production (AEP), minimum levelized cost of energy (LCOE) or maximum net present value (NPV).
2.3. Locations
Data centres are often geographically located in clusters. This is for many reasons such as available power, connectivity, workforce, politics and policy among others. However, the main drivers are power and connectivity.
The cost of land for each if the FLAPD locations is detailed in
Table 1. It has been assumed that the land required is purchased rather than leased, thereby inflating the
LCOE at each site.
2.4. Technology Costs
In order to calculate the
LCOE for each technology and to allow WindPRO to optimise the hybrid renewable energy system configuration, the cost per unit of each technology is required and is shown in
Table 2. The cost per unit is given as capital expenditure (CAPEX) and operating expenditure (OPEX). The
LCOE includes both the CAPEX and OPEX.
2.5. Wind Turbines
Within WindPRO there are modelling parameters available for a vast number of wind turbine designs. The same turbine design has been selected for all site locations for ease of comparability, although it is appreciated that the optimum wind turbine choice may differ across sites.
Within WindPRO five wind turbines of different power ratings and hub heights were selected for comparison to find the optimum wind turbine that gave the lowest
LCOE over 20 years. The five turbines selected are detailed in
Table 3 and the optimum wind turbine was found to be the Vestas V172-7.2.
The results shown in
Table 3 are for the Southwest England site. It is appreciated that the AEP and
LCOE will be different at each site; however, it is assumed that the optimum turbine will prevail across all sites and as such this process is not repeated for each site.
It is recognised that the
LCOE results for both the Vestas V172-7.2 and GE WIND ENERGY GE 2.5-120 are much lower than the 3.3 MW utility-scale example shown by the NREL [
25] and the UK government prediction for 2025 for onshore wind [
31] which show 39 USD/MWh and 38 GBP/MWh, respectively. This difference is due to these wind turbines having higher hub heights than most typical onshore wind turbines. This could potentially lead to planning permission issues and result in requiring a wind turbine of lower hub height. It is expected that a lower hub height would result in the wind farm having a lower capacity factor thereby increasing the
LCOE.
2.6. Wind Farm Configuration
The wind farm layout for each location is optimised using the WindPRO wind farm layout optimisation function. Using this function the
LCOE is improved compared to the traditional spacing method of 7D × 4D and the 15D spacing method suggested by Meyers and Meneveau [
21]. The results for a site in Slough, England are shown in
Table 4.
For each of the proposed locations, a square area of circa 100 ha is drawn in WindPRO which is angled so that the mean wind direction at the turbine hub height shown from the site wind rose is perpendicular to the area.
Figure 1 shows how this is seen visually in WindPRO. It is appreciated that the availability of land is a primary consideration for each hybrid renewable energy system and that this ideal land geometry will most likely not be available. However, using the optimum geometry at each site allows for a fair comparison between sites.
Once the area is defined, using the wind farm optimiser in WindPRO a variety of configurations with varying numbers of wind turbines is simulated giving the optimum layout for annual energy production (AEP). From the AEP of each layout, the optimum
LCOE can be found.
Figure 2 shows the WindPRO wind farm optimiser output layout.
The WindPRO wind farm optimiser uses a ‘Global Atlas of Siting Parameter’ as the input for the wind speed data. Actual site conditions will deviate from this data, and WindPRO states that for “flat” terrain the mean error is 5.5% and the median error is 4.0%, recorded against 241 masts. This error is appreciated but not considered any further within this paper.
2.7. Solar Panels and Solar Farm Configuration
WindPRO has a variety of solar panel models within the software; however, unlike the wind turbine models which represent manufacturers and models available on the market, the solar panel models are generic. As with the wind turbine selection, the same solar panels will be used across all sites. The solar panel selected across all sites is the EMD-Generic 700 W due to having the highest power rating.
For this paper the tilt angle is tracked, therefore for optimisation purposes, the tilt angle is optimised within WindPRO and then fixed for the simulation for optimising the row distribution and azimuth angle. Row distribution is the spacing between adjacent rows of solar panels, sometimes referred to as row spacing. Row distribution affects how much one solar panel shades another at different tilt and solar angles, and therefore determines how much energy the shaded solar panel can produce.
Within WindPRO the solar panel azimuth can be altered such that it allows for a more consistent energy production profile with a lower maximum generation but increased at the extremes of sunrise and sunset. This arrangement is acknowledged with respect to its benefits for the flat data centre demand considered within this paper. However, for ease of comparability across each site a fixed optimum azimuth is determined for each site.
The solar irradiance data used for all calculations within WindPRO uses a SARAH-3 data set, this is a satellite-driven metrological measuring system that covers a spatial resolution of 5.5 km with a temporal resolution of 30 min. The mean error in data is 2.1% [
32]. This data set was selected as it provides high-resolution solar irradiation data across each site used within this paper, therefore the reliability of the solar irradiation data is uniform across all sites.
2.8. WindPRO Limitations and Uncertainties
This paper utilises the WindPRO micro-grid optimiser for HRES optimisation. The micro-grid optimiser in WindPRO can also include battery storage, curtailment, load shedding, grid connections and dispatchable generators (oil and gas) for a user-defined demand profile, covering the technology and topology of the systems proposed in this project. The WindPRO micro-grid optimiser, searches for the minimum lifetime cost of the system using an iterative method, adjusting the scaling of each technology capacity. The output gives predicted hourly generation data for the lifespan of the HRES. By including carbon pricing within the optimiser, emissions can be indirectly optimised.
The optimiser is limited in that it does not consider the non-linearity of technology costs therefore limits are implemented within the optimiser so that the system does not increase or decrease the magnitude of a given technology outside of the cost range for which the associated costs are reflective. The optimiser is also limited in that land requirements are not considered; for example, the lifetime cost of the system may be reduced by GBP 1 but require an additional hectare of land. Another limitation of the WindPRO hybrid optimiser is that it does not consider the degradation of the battery storage system which would be expected at a rate of more than 1.5% per year based on a single charge and discharge period per day [
33]. The battery storage system costs are also excluded from the
LCOE calculation within the optimiser therefore the
LCOE of each site is calculated using Excel.
4. Conclusions
This paper demonstrates that the challenges faced by data centre developers and renewable energy generators to connect to increasingly constrained grids can be solved by implementing off-grid hybrid renewable energy systems. The example systems presented within this paper show that hybrid renewable energy systems can significantly reduce data centres’ cost of energy and emissions compared to purchasing energy from their associated grids or equivalent gas generation-only schemes.
This paper also demonstrates the economic and environmental feasibility of deploying a gas-only solution to facilitate the initial operation of the data centre until renewable generation becomes available for integration into the system, or a grid connection becomes available. It was found in some instances that the gas-only solution is economically favourable in comparison to purchasing energy from local grids; however, the associated emissions would be higher.
The results of the hypothetical case studies presented within this paper give evidence to suggest that data centre developers should consider undertaking full feasibility studies for implementing off-grid hybrid renewable energy systems specific to their data centre requirements.
The paper also provides a new economic metric for evaluating off-grid hybrid renewable energy systems, the levelized cost of energy utilised (LCOEu). This metric gives a true representation of a data centre’s cost of energy when supplied from an off-grid hybrid renewable energy system.