GB2525573A

GB2525573A - Control of array of energy producing devices

Info

Publication number: GB2525573A
Application number: GB1320797.2A
Authority: GB
Inventors: Thomas Henry Clark
Original assignee: OCEAN ARRAY SYSTEMS Ltd
Current assignee: OCEAN ARRAY SYSTEMS Ltd
Priority date: 2013-11-26
Filing date: 2013-11-26
Publication date: 2015-11-04
Also published as: GB201320797D0

Abstract

A control unit 300 for an array of power generating devices, such as wind or tidal turbines, or wave power generators, uses real-time data from the devices to determine a state of the array, and dynamically control operation of devices in the array. The control system may use a number of algorithms, and use a reconciliation module to select the best control algorithm for the conditions, or combine the algorithms to improve control of the array. The algorithms may include Array-As-A-Sensor which observes the state of the devices and sensors in the array to predict impending events; real-time feature recognition, which recognises signatures of events such as blade break or impact; and a wave propagation algorithm.

Description

Control of array of energy producing devices The invention relates to systems and methods to control an array of devices adapted to use kinetic energy of a stream of fluid to produce electrical energy.

Onshore or offshore power plants comprising farms of devices, comprising e.g., wind turbines, tidal stream turbines and/or wave energy converters, produce energy from a fluid in movement, such as wind or sea currents.

It is known that the issues of turbulence, wave propagation, meteorological effects and many other (typically unsteady) fluid flow and environmental effects impact the generation of power from the devices, and these issues are usually all considered at the design and analysis stages of building power plants.

Attempts have also been made to take into account changes in the fluid flow during operation of the farms -Figure 1 shows schematically an example of a typical known farm 1 comprising energy producing devices 3, such as wind turbines, tidal stream turbines and wave energy converters. The farm 1 is usually equipped with a Supervisory, Command And Data Acquisition (SCADA) system 2. Such SCADA systems are commonplace for use in control of multiple devices. The SCADA system 2 typically has data input for local area sensor data (such as coming from a flow determination sensor, usually located upstream at least one of the devices 3); and for non-real-time processed data (such as coming from a sensor or for run-time data) from respective individual control modules 30 of the individual devices 3 of the farm, as shown in Figure 1.

The farm 1 is thus supervised by the SCADA system 2 in that the SCADA system 2 can send non-real-time operational commands (such as an on/off command or an operation set point) to the individual control modules 30 of the devices 3.

However, as explained above, in conventional farms, power production is not optimal, as the farms only take into account simple measurements about e.g., the changes in the fluid flow, and are not configured to take into account any physical understanding of the fluid dynamic (unsteady) effects involved, such as environmental and turbine4urbine interference effects.

Aspects of the invention address or at least ameliorate at least one of the above issues.

In one aspect, the invention provides a control unit configured to control an operation of an array comprising at least two devices adapted to use kinetic energy of a stream of fluid to produce electrical energy, the unit being configured to: determine a state of the array using real-time data relating to an operation of at least one of the devices of the array; modify the operation of at least one device of the array, as a function of the determination.

The real-time data may refer to at least one of the following parameters: torque, thrust, rotations per minute, shaft strain and/or shaft stress. The unit may further be adapted to use external data, such as environmental data and/or the economic, operational and logistic data. The unit may further be configured to: output a real-time control command for the at least one device, as a function of real-time data relating to an operation of at least two devices in the array.

In another aspect, the invention provides a control unit configured to control an operation of an array of at least one device adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least two algorithm modules, each module being configured to generate at least one real-time control response for at least one of the devices, by implementing, a respective control algorithm; and a reconciliation module configured to reconcile the real-time control responses in order to output the real-time control command.

The unit according may be configured to: generate, as a function of real-time data relating to the operation of all the devices in operation in the array, at least one real-time control response for all the devices in operation in the array; and output a real-time control command for all the devices in operation in the array.

The algorithm module may be configured to implement at least one of the following control algorithms: Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response, Wave propagation, Thrust/mass flux distribution response, Unsteady inflow vs. yield response, Unsteady flow vs. peak load response, Unsteady flow vs. dynamic range response, Fluid feature and debris advection, Awareness model. The awareness model may include at least one of the following: Set point response, Peak allowable load awareness model, Lifetime awareness model, Economic awareness model, Operational and/or logistical awareness model, and/or Environmental data response.

In another aspect, the invention provides a control unit configured to control an operation of an array of at least one device adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least algorithm module configured to implement at least one of the following control algorithms: Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response.

The algorithm module may be configured to generate at least one real-time control response for at least one of the devices, by implementing the at least one control algorithm. The algorithm module may further be configured to implement at least one of the following control algorithms: Wave propagation, Thrust/mass flux distribution response, Unsteady inflow vs. yield response, Unsteady flow vs. peak load response, Unsteady flow vs. dynamic range response, Fluid feature and debris advection, Awareness model. The awareness model may include at least one of the following: Set point response, Peak allowable load awareness model, Lifetime awareness model, Economic awareness model, Operational and/or logistical awareness model, and/or Environmental data response. The unit may further comprise at least one optimisation module configured to implement Machine Learning and/or Artificial Intelligence to the algorithm module. The unit may further comprise an optimisation module configured to implement Machine Learning and/or Artificial Intelligence to the reconciliation module.

The unit may be implemented at least partially as software or firmware, and/or implemented at least partially in a physical casing, separate from the devices.

In another aspect, the invention provides a system comprising: an array of at least two devices adapted to convert kinetic energy of a fluid into electrical energy; and at least one control unit according to any one of aspects of the invention.

The system may comprise at least two redundant control units. The device may comprise a wind turbine, a tidal stream turbine or a wave energy converter.

In another aspect, the invention provides a method for controlling an operation of an array of at least two devices adapted to use kinetic energy of a stream of fluid to produce electrical energy, comprising a control unit: determining a state of the array using real-time data relating to an operation of at least one of the devices of the array; modifying the operation of at least one device of the array, as a function of the determination.

The real-time data may refer to at least one of the following parameters: torque, thrust, rotations per minute, shaft strain and/or shaft stress. The unit may use external data, such as environmental data and/or the economic, operational and logistic data. The unit may output a real-time control command for the at least one device, as a function of real-time data relating to an operation of at least two devices in the array.

In another aspect, the invention provides a method for controlling an operation of an array of devices adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least two algorithm modules generating at least one real-time control response for at least one of the devices, by implementing, a respective control algorithm; and a reconciliation module reconciling the real-time control responses in order to output the real-time control command.

The unit may: generate, as a function of real-time data relating to the operation of all the devices in operation in the array, at least one real-time control response for all the devices in operation in the array; and output a real-time control command for all the devices in operation in the array.

The algorithm may be at least one of the following control algorithms: Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response, Wave propagation, Thrust/mass flux distribution response, Unsteady inflow vs. yield response, Unsteady flow vs. peak load response, Unsteady flow vs. dynamic range response, Fluid feature and debris advection, Awareness model. The awareness model may include at least one of the following: Set point response, Peak allowable load awareness model, Lifetime awareness model, Economic awareness model, Operational and/or logistical awareness model, and/or Environmental data response.

In another aspect, the invention provides a method for controlling an operation of an array of devices adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least algorithm module implementing at least one of the following control algorithms: Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response.

The algorithm module may generate at least one real-time control response for at least one of the devices, by implementing the at least one control algorithm. The algorithm module may implement at least one of the following control algorithms: Wave propagation, Thrust/mass flux distribution response, Unsteady inflow vs. yield response, Unsteady flow vs. peak load response, Unsteady flow vs. dynamic range response, Fluid feature and debris advection, Awareness model. The awareness model may include at least one of the following: Set point response, Peak allowable load awareness model, Lifetime awareness model, Economic awareness model, Operational and/or logistical awareness model, and/or Environmental data response. At least one optimisation module may implement Machine Learning and/or Artificial Intelligence to the algorithm module. An optimisation module may implement Machine Learning and/orArtificial Intelligence to the reconciliation module.

Aspects of the invention extend to computer program products such as computer readable storage media having instructions stored thereon which are operable to program a programmable processor to carry out a method as described in the aspects and possibilities set out above or recited in the claims and/or to program a suitably adapted computer to provide the system recited in any of the claims.

The invention may have numerous advantages over the prior art.

The invention may enable taking into account physical understanding of the fluid dynamic (unsteady) effects involved, such as environmental and turbine-turbine interference effects. Thus the invention may enable the farm to handle the unsteady fluctuations in load and power generation, as well as spatial variations between different devices in the farm, and this during operation of the farm.

The invention may allow integration of all the data within a control approach in a farm, or some of the data may be used in tandem with the existing measurements, in order to further improve performance and robustness of a device and/or an array of devices.

The invention may provide a real-time (or nearly real-time) centralized control of the devices.

Where multiple rotors are used, e.g., in an array comprising either several single-rotor devices, or at least one device comprising multiple rotors, the invention may enable a wider array-level control approach, in order to address the effects of the fluid dynamic and environmental and turbine-turbine interference, thus improving the performance of the individual devices as well as the net performance and operability of the array.

A unit according to the disclosure may have:

the ability to make informed selection of control algorithm used (either pre-set or in real time); and/or the ability to incorporate the power of Machine Learning and/or Artificial Intelligence at more than one level, whilst retaining awareness of the physics at play; and/or the ability to handle high-dimensional problems, multiple solution problems and multiple objective problems.

Aspects of the invention will now be described, by way of example, with reference to the accompanying drawings in which: Figure 1, already discussed, schematically illustrates a typical layout of a known farm having a SCADA system; Figure 2 schematically illustrates a farm having a Control, Command And Data Acquisition (CCADA) system; Figure 3 schematically illustrates an exemplary implementation of external data feeds into the CCADA system of a farm comprising redundancy; Figure 4 schematically illustrates an exemplary implementation of redundancy in an array of a farm comprising redundancy; Figure 5 schematically illustrates an exemplary implementation of a CCADA system using multi-layer control system; Figure 6 schematically illustrates an exemplary implementation of a CCADA system using multi-layer control system, optionally but advantageously implementing Machine LearninglArtificial Intelligence (ML/Al); Figure 7A schematically illustrates the principle of an implementation of device-device interaction in an array of devices, or thrust vs. mass flux; Figure 7B schematically illustrates the principle of an implementation of an array thrust control in a case of an outage coverage, in an array of devices; Figure 8 schematically illustrates the principle of an implementation of response control for spatial variation in the environment in an array of devices; Figure 9 schematically illustrates an exemplary implementation of an external data feed in a CCADA system using multi-layer control system; and Figures 1OA and lOB schematically illustrate curves representing the behaviour of a device without real-time feature recognition (dotted lines) and the control of the device with real-time feature recognition (solid line).

In all of the Figures, similar parts are referred to by like numerical references.

Overview Figure 2 schematically illustrates a farm 1 comprising at least an array comprising at least one device 3 adapted to produce energy from the movement of a fluid, i.e. using kinetic energy of a stream or flow of fluid to produce electrical energy. The fluid may be a gas (such as air) or a liquid (such as sea water), and the device 3 may thus comprise at least one wind turbine, tidal stream turbine or wave energy converter.

The array may be multiple-rotor, i.e. the array may be formed of at least two rotors of one of the devices 3, or the array comprises at least two single-rotor devices 3.

The farm 1 comprises a Farm Control Unit (FCU) 300 configured to form a Control, Command And Data Acquisition (CCADA) system. The CCADA is configured to form a controller for the array of at least one device 3. The unit 300 is thus configured to: determine a state of the array using real-time data relating to the operation of at least one of the devices 3 of the array; and modify the operation of at least one device 3 of the array, as a function of the determination.

The state of the array refers to the operation status of all the devices 3 linked to the unit 300. The unit 300 is preferably linked to more than two devices 3. The status of one of the devices may, e.g., be "on" operation (under different possible regimes of operation, as it will be apparent from the present specification) or "off" operation (deliberate shut down or breakage).

As non-limiting examples, real-time data may refer to at least one of the following parameters of the device 3: torque, thrust, rotations per minute, shaft strain and/or shaft stress.

It will be appreciated that, in the specification, the term "real-time" encompasses "near real-time", i.e. the only time delay introduced between: the occurrence of an event during the operation of at least one of the devices 3 of the array and the transmission of the data to and/or the receipt of the data by the unit 300 is introduced by the basic processing of raw data by the device 3 and/or the data transmission time from the device 3 to the unit 300. There are thus no significant delays.

It is appreciated that in the farm of Figure 1, the SCADA system 2 does not process real-time data or near real-time data, as the data is processed first at least by the controls 30. The controls 30 then transmit delayed feedback of processed data to the SCADA system 2.

It will be appreciated that the modification of the operation of the at least one device 3 of the array may also preferably be real-time (and thus near real-time), i.e. the only time delay introduced between: the determination of a state of the array and the modification of the operation of at least one device 3 of the array is introduced by the basic processing of raw data by the unit 300 and/or the command transmission time from the unit 300 to the device 3. There are thus no significant delays.

It is appreciated that in the farm of Figure 1, the SCADA system 2 does not modify the operation of the device in real-time, as the command is processed first at least by the controls 30.

The unit 300 may further be adapted to use data external to the devices 3, such as: environmental data, e.g. local area sensor data 6; and/or economic, operational and logistic data, e.g. external and/or Wide Area Network data 7 (the feed for the economic and operational data may be indirect, as explained in greater detail below).

In order to control the array of devices 3, the control unit 300 is also thus configured to: generate, as a function of real-time data relating to an operation of at least two devices 3 in the array, at least one real-time control response for at least one of the devices 3, by implementing, when generating each control response, a respective control algorithm; and output a real-time control command for the at least one device 3, as a function of the generated real-time control responses.

Preferably, the unit 300 is configured to generate at least one real-time control response for all the devices 3 in operation in the array, as a function of real-time data relating to the operation of all the devices 3 in operation in the array. The unit 300 is then preferably configured to output a real-time control command for all the devices 3 in operation in the array.

As shown in Figure 2, control lines 31, or loops, interconnect the control unit 300 and at least two, preferably each, of the devices 3.

The farm 1 may also comprise a conventional Supervisory, Command And Data Acquisition (SCADA) system 2.

The SCADA system 2 may typically have data input 21 for non-real-time processed data from the unit 300 (such as coming from a sensor of a device 3 or for run-time data), as shown in Figure 2. The SCADA system 2 is also preferably configured to send, via an output 22, non-real-time operational commands (such as an on/off command or an operation set point) to the unit 300. The SCADA system 2 may also have a data output 23 to, e.g., a shoreside sewer facility 4.

The facility 4 is linked, via command and/or data links 41 and 42, to an online interface and/or dashboard 5 operated by a human and/or automated operator of the farm 1. The operator may then feed the server facility 4 with commands and/or data via the interface and/or dashboard 5.

Therefore, the control lines 31 are preferably able to transmit both commands and data, such that the system 2 may be operated as a conventional farm comprising only a SCADA system 2 as shown in Figure 1, e.g., in the event of a failure or a maintenance operation of the unit 300. Furthermore, command and supervisory functions may appear identical to a typical SCADA from the perspective of an operator, which may also facilitate retro-fitting on existing farms.

The server facility 4 may also have an input 43 for an external and/or Wide Area Network (WAN) data 7 feed, feeding data such as environmental data and/or economic and operational data as explained in greater detail below.

Preferably, communication down control lines 31, and to shore, may be made via fibre optic cables, as the latter have a high bandwidth, but other forms of communication, such as TCP/IP, are also possible.

It will be appreciated that Figure 2 is only schematic, and that the unit 300 may be physically located at a shore station, in a subsea hub or similar subsea container, on a service platform (at or above sea level, floating or fixed) or within the devices (e.g. inside nacelles or platforms) in the array. It will also be appreciated that the unit 300 may be implemented at least partly as software or firmware, i.e. relying on parallel or distributed computing capability (e.g. distributed to a cloud server, or to local processors in other control systems within the array).

Similarly, the SCADA system 2 may be implemented at least partly as software or rmware, and be a part of the unit 300 and/or be a part of the sewer facility 4, and/or be located in a physical casing, separate from the unit 300 and/orthe sewer facility 4.

Optionally, each device 3 may have an onboard backup controller 30. However, in an advantageous normal way of operation, control functions are undertaken centrally by the unit 300.

As known to those skilled in the art, registration signals may be periodically sent via the same communication infrastructure to and from the FCU 300 (e.g. by multiplexing a periodic registration signal down the same lines as the control and data signals) allowing the unit 300 to identify and safely manage communication interruptions (e.g. broken cable or similar fault states or planned interruptions and outages).

As shown in Figures 3 and 4, in order to mitigate the risk of failure of an FCU 300 or of communication lines to the unit 300, multiple physical FCUs 300 may be incorporated as redundant units into an array. In this case, the FCUs 300 may be incorporated into some device nacelles, making these devices (referred to as 3') key devices, and other devices (referred to as 3) non-key devices. Figure 3 shows the diagrammatic layout of a turbine farm, excluding the cabling, hubs and shoreside features. The non-key turbines 3 may house onboard safety layer with basic backup control system (shown as 30 in Figure 2) which can be activated under a variety of conditions e.g. fault state of the remote controller. The Farm Control Units 300 may reside in subsea hubs, platforms, non-locally or in the key turbine 3'.

It will be appreciated that array maintenance operations preferably take into account the requirement for minimum redundancy when removing key devices 3' or e.g., subsea hubs containing FCUs 300.

Redundancy (multiple FCUs, one or more active at a given time) allows key turbines 3' or FCUs 300 to be disconnected for maintenance. The degree of redundancy (ratio of key to non-key turbines in the illustrated case) preferably allows for both the likelihood of fault states and the likelihood of key turbine 3' removal (maintenance, etc.). In an extremely redundant scenario, an FCU 300 may be incorporated into each device 3. Figure 4 shows that which FCU is active at any given time may be controlled via operational input or via fault state logic. Whilst the system is foreseen to be run with a sole FCU active at a time in order to prevent conflicting behaviour, modes in which multiple FCUs run at once may be envisaged.

The currently operating FCU 300 in a farm with redundancy can be switched manually by an operator (locally or remotely), and/or automatically on occurrence of a fault state or communication interruption.

Farm Central Unit As already mentioned, the unit 300 is configured to generate at least one real-time control response for at least one of the devices 3 by implementing at least one respective control algorithm. The unit 300 thus enables the use of a plurality of models and control algorithms, for real-time (and near-real-time) control of the devices 3 of the array. The control is said to be real4ime (or near-real-time as explained above) because it may use real-time data, e.g., relating to an operation of at least one device (preferably at least two devices, and very preferably all the devices, in operation in the array), i.e. knowledge of the current operating state of the devices, the said device being taken either from the point of view of a device or from the point of view of a sensor as explained in greater detail below.

The unit may use real-time data external to the devices, such as environmental data and/or economic and operational data.

At least one sensor in the device 3 is configured to measure the real-time state of the device (such as data relating to torque, thrust, rotations per minute, shaft strain and/or shaft stress) thus determining the state of the array. The unit 300 is in turn configured to receive the data (e.g., the measurements -10-of the sensor) via the control lines 31 forming a feedback loop for each of the devices 3. The unit 300 is furthermore configured to apply at least one control algorithm to correct for any deviation between the desired and actual states, i.e. modify the operation at least one device 3 of the array or output a real-time control command for the at least one device 3.

The desired state can be set in advance (it is the case e.g., for a basic set operating point) or dynamically updated based on user commands or automated processes, such as machine learning to improve performance over time as described in more detail below. It is appreciated that a single desired state (i.e. one value for each of the independent variables) must be chosen (either by the operator or through the automated process). For example, a standalone turbine with fixed pitch blades may have as little as one independent variable (e.g. terminal voltage varied in order to control rotational speed and shaft torque for optimal power generation). However, even for a single dependent variable (as in the example above: maximising power generation) it is appreciated that there may be multiple methods for determining what state will maximise the desired property. For the simple example above, lookup tables of known performance (embodying engineer's experience), a machine learning and/or artificial intelligence (ML/Al) algorithm, or an optimisation algorithm could be used, and different methods would give different responses to the input state.

Figures 5 and 6 show examples of a unit 300 comprising at least one algorithm module 303 configured to implement at least one control algorithm. The unit 300 thus forms a Multi-Layered Control System (MLCS). The unit 300 may thus take advantage of the fact that, in some circumstances, one algorithm may yield a more desirable response than another. In other cases, several algorithms may each have an associated inaccuracy; so blending of several algorithms can improve the overall accuracy in prediction of the most favourable response. In still other cases, some models might capture important aspects of physics at play, offering a better insight into the likely dynamics of the response than those which do not.

The algorithm module 303 may be configured to implement at least one of the following control algorithms and/or models, described in greater detail below: Array-As-A-Sensor; Array-As-A-Device; Operational response; and/or Real-time feature recognition response; and/or Wave propagation; Thrust/mass flux distribution response; Unsteady inflow vs. yield response; Unsteady flow vs. peak load response; Unsteady flow vs. dynamic range response; Fluid feature and debris advection; and/or Awareness models, including at least one of the following: -11 -Set point response; Peak allowable load awareness model; Lifetime awareness model; Economic awareness model; Operational and/or logistical awareness model; and/or Environmental data response.

It will be appreciated that in a device array, there are many independent control variables, many sensors and highly complex interactions occurring between devices and the environment. Many different additional methods for selection of the best operating point may also be used.

It will be also appreciated that each different response model provides a different, possibly conflicting, recommendation for the command to the devices. The unit thus comprises a reconciliation module 301 configured to reconcile the real-time control responses for the devices 3 from the modules 303, in order to output the real-time control command to at least one, preferably all, of the devices 3. The MLCS thus offers ability to make informed selection of the control algorithm used, either statically (e.g. by defining preset conditions under which each method is used, with blending or weighting between them) or dynamically (e.g. using Machine Learning (ML) and/or Artificial Intelligence (Al), or similar, to determine under which conditions each method performs best, then select and blend accordingly).

As known by those skilled in the art, each artificial intelligence and/or machine learning module can be viewed as a black box' which receives raw sensor data and outputs the best response, having received some initial training (the latter based usually on enginee(s insight). However the integrity and rate of learning decrease with the number of degrees of freedom, and modules implementing ML/AI are not physics aware' which means that their learning is based on minimising the difference between expected and actual behaviour based on correlation between key variables. Therefore the capacity of artificial intelligence and/or machine learning modules to identify and respond to unusual events is limited, as is awareness of why and when key variables are correlated. Figure 6 shows an example using ML/AI in a more powerful way by compartmentalising the tasks required.

To advantage, the unit 300 thus further may comprise an optimisation module 302 configured to implement Machine Learning and/or Artificial Intelligence to the reconciliation module 301.

Similarly, the unit 300 may further comprise at least one optimisation module 304 configured to implement Machine Learning and/or Artificial Intelligence to at least one of the algorithm module 303.

It is thus appreciated that by implementing the optimisation module 302 directly and only to the reconciliation module 301 and/or by implementing an optimisation module 304 directly and only to one of the algorithm modules 303 (thus compartmentalising the tasks required), the integrity and rate of learning stays high, because the number of degrees of freedom is locally limited, and some algorithm modules 303 may be physics aware', with better capacity to identify and respond to unusual -12 -events and with awareness of why and when key variables are correlated. For example, in lower stages (such as in feature recognition module 303), the ML/Al module 304 provides fine tuning to individual models, improving accuracy of prediction. In higher stages (such as in array-as-a-sensor), the ML/Al module 304 can be used in a physics aware' mode, accepting inputs relating to physical metrics or model outputs (rather than just raw sensor signals) to assist in the reconciliation process in the reconciliation module 301.

The unit may thus enable the operation of systems with high-degree of freedom having multiple solutions for possible commands, because the reconciliation process in the module 301 has the inherent capacity to handle multi-objective problems. An example would be in an array of for example four turbines, attempting to generate a maximum power. With any two turbines operating off-design (i.e. at a lower power), local acceleration of flow through the other two turbines can result in a local maximum in the hyperspace that consists of the independent control variables and the power.

However, elsewhere in the hyperspace, a different solution with greater or equal power (such as all turbines operating uniformly) can exist. By making the reconciliation module 301 aware of different aspects of the physics and/or economics in play (through use of multiple models in the algorithm modules 303), the MLCS has the ability to identify and suppress the effect of multiple solution problems, since a different version of the hyperspace is used in each model, and the control algorithms do not get caught in local minima or oscillate between different solutions.

Control algorithms and models As shown in Figure 6, the unit 300 may use different control and/or management models, including at least one of the following models or algorithms: Array-As-A-Sensor, e.g., implemented in the module 3031; Array-As-A-Device, e.g., implemented in the module 3032; Operational response, e.g., implemented in the module 3033.

Array-As-A-Sensor (AAAS) model In the AAAS model, the module 3031 is configured to observe the state of other devices (and wider sensor data 7) to predict impending events and/or input state changes elsewhere in the array. The AAAS module 3031 may preferably incorporate the principle of the device as an upstream sensor.

The module 3031 allows the ultimate device and array response to be updated in real or near-real time as a response to events happening elsewhere within the network.

Array-As-A-Device (AAAD model In the AAAD model, the module 3032 is configured to manage the state of some, or all, of the devices in the array, in order to improve performance of under-performing individual devices within the array, as well as the net array performance. For example, thrust distribution throughout an array can be -13-managed to optimise net array yield, as explained in greater detail below. The AAAD module 3032 may advantageously incorporate a module 3034 implementing a device-device-environment interaction model.

Operational response model In the Operational response model, the module 3033 is configured to combine engineering metrics, operational and economic cost functions and/or constraints to provide optimal trade-off between maximisation of revenue and operational costs, including device lifetime management and assisting in decision making between lost revenue and unplanned or more frequent maintenance or shorter lifetimes. The model may incorporate Economic and Logistic metrics, as described in more detail below.

The above-mentioned models or algorithms may use at least one of the following sub-models.

Real-time feature recognition (RTFR) response; Wave propagation; Thrustmass flux distribution response; Unsteady inflow vs. yield response; Unsteady flow vs. peak load response; Unsteady flow vs. dynamic range response; Fluid feature and debris advection; and/or Awareness models, including at least one of the following: Set point response; Peak allowable load awareness model; Lifetime awareness model; Economic awareness model Operational and/or logistical awareness model, and/or Environmental data response.

Real-time feature recognition response The real-time feature recognition response, described in more detail below, allows disaster mitigation, peak load relief and unsteady yield optimisation in wind, wave and tidal industries, as it allows a signal or group of signals to be monitored in real time (or near-real time), and features within the signal identified (e.g. blade break, impact, impingement of a particular type of turbulent flow structure). The real-time feature recognition response may take into account at least one signature data of upstream events.

As shown in Figure 6, at least one of the RTFR modules 303 may utilise an optimisation module 304 implementing a leaming pack. The learning pack comprises instructions configured to teach the -14-module 303 to recognise events. For example, impact of marine debris on a blade can be simulated in a laboratory or virtual environment and the results used to create a learning pack for identification of such events. Learning packs may include but are not limited to: marine mammal proximity, marine mammal strike, bird proximity, bird strike, fish proximity, fish strike, impact of debris, impingement of turbulent flow structures of different types, sizes or strengths, blade breakage, gearbox breakage, short circuit or other electrical power system events, vessel proximity, UAV/ROV proximity, plane or helicopter proximity.

Wave pmpagation Using knowledge of wave parameters (e.g. height) and direction at one position, a predictive model can be used in at least one of the modules 303 to ascertain the passage of waves in real time from that position to another in the same locale. For example, a wave passing over a wave height sensor (this is an example of sensors integrated at a farm level rather than at a device level) with known direction and known bathymetry will propagate some distance DX in time DT. For example, as waves pass over turbines 3 in an array, fluctuations in the inflow velocity affect the turbine. Using a wave propagation model, measurement of waves at the edge of an array allows prediction of the time at which waves arrive at turbine locations within the array, as well as predicting the wave characteristics (e.g. amplitude, wavelength) at the turbine locations. The wave propagation model thus allows fluctuations in load and power throughout the array caused by a wave to be predicted and responded to if necessary.

The measurement of waves at the edge of an array can be done directly, e.g. using sensors measuring wave height and direction (or similar parameters), or can be done indirectly, e.g. by recognition of the signature of a wave affecting a turbine at the edge of an array.

As explained above the wave propagation model can be refined (or established entirely, in the lack of an analytical, computational or experimental model as a basis) using machine learning or artificial intelligence.

Thrust vs. mass flux distribution response For a given device installation, the thrust applied to the flow by the device, e.g., a turbine, is related strongly (and nonlinearly) to the mass flow through the device. It is this relationship that gives rise to various models (such as the Betz Limit) for optimal power output of a turbine, where Thrust and Mass flux are optimally balanced for greatest power capture. Even for a single turbine, increasing the thrust coefficient decreases mass flow through the device, and as known by those skilled in the art, there exists an optimum thrust (which in some cases can be analytically derived, e.g., by the Betz' Theorem) at which power is maximised.

An important aspect, as shown in Figures 7A and 7B, is that altering thrust at one turbine affects thrust, mass flow and power output at all other turbines in an array (to varying degrees depending on -15-flow conditions and relative locations and taking into account aspects such as viscous performance, bathymetry, surface effects and interactions between multiple turbines). Figures 7A shows that for arrays, turbines interfere with one another. In Figure 7A, changing thrust coefficient at turbine alters mass flux not only through turbine but also through turbines 2 and 33. There is an optimal thrust distribution (balance of thrust between turbines) as well as net thrust. Depending on the spatial arrangement of the array layout, there may be multiple optimal distributions and local optima close to arbitrary thrust distributions. Figure 7A shows four (of many) possible configurations in which device-device interaction can take place.

The net thrust has a primary effect on net yield. However it is known that there is a secondary effect on net yield and that the distribution of thrust affects individual turbine yield, loading and wear rates.

Thus loading, wear rates and yield can be traded off between individual units in the array for best economic benefit.

In Figure 7A(1), low thrust at turbine leads to higher mass flux through turbine 2 in its wake. In Figure 7A(2), high thrust at turbine leads to low mass flux through turbine 2 in its wake. In Figure 7A(3), high thrust at turbine relative to turbines 2 and 33 leads to higher mass flux through turbines 32 and 33, outside but adjacent to the turbine 3 stream tube: net thrust affects the net mass flux through and yield of the array. Figure 7A(4) shows that decreasing thrust at turbine relative to turbines 2 and 33 leads to reduced mass flux through turbines 2 and 33, outside but adjacent to the turbine 1 stream tube: again, net thrust affects the net mass flux through and yield of the array.

More generally, numerical or experimental models of device (e.g., turbine) arrays (or even careful variation within full scale arrays) may be used to ascertain the relationship between thrust (at each device -i.e. the thrust distribution in the array) and the mass flow and power output of each device (and therefore the entire array). This relationship is the thrust/mass flux response model.

Thrust exerted by an individual turbine can be controlled using various means; designing blades to have a particular flexibility or aerodynamic characteristic, regulating voltage or current at the generator using the electrical power system, feathering blades (in a controllable pitch system), altering gearing in the drivetrain or actively altering aero/hydrodynamic characteristics or shape.

Thus, with a thrust/mass flux response model, together with the ability to control turbine thrust, the individual turbine loads and outputs across the array as well as the net array thrust and output can be varied in order to achieve a desirable state.

Figure 7B shows an example of control of the thrust distribution (per Device-Device Interaction') in case of a failure of a turbine, in which array yield (and therefore lost revenue) may be de-sensitised to prolonged outage or downtime, by taking into account aspects of operations and maintenance costs, such as including vessel availability, weather windows and lost revenue. In the example of Figure 7B, marine operations in case of a failure of a turbine can be conducted with increased flexibility and less weather constraints (thus in safer and more cost effective conditions). Figure 7B(1) shows normal operation of an array forming a tidal fence. Figure 7B(2) shows that a failure (represented by a -16-missing turbine) causes change in thrust distribution. Mass flux thus increases disproportionally through the failed turbine, and the remainder of the array suffers decreased yield in addition to the lost yield form the failed device. Figure 7B(3) shows that in a thrust control scenario, devices are operated at a point associated with increased thrust, and distribution of thrust within the array in this way is used to alter (e.g. increase) mass flow through the rest of the array, allowing the net yield to be less sensitive to the outage.

As explained above, the model can be refined (or established entirely, in the lack of an analytical, computational or experimental model as a basis) using machine learning or artificial intelligence.

Unsteady inflow vs. yield response Whether wave impingement, environmental turbulence from the marine boundary layer and bathymetry, atmospheric turbulence in the wind, or turbulence created by upstream structures or devices, unsteady flow features impinging on a device (as well as in the region of, but not directly impinging on a device) has an effect on energy yield of devices. For example, aero/hydrodynamic performance of blades is affected by the change in stall characteristics of 2D sections resulting from the presence of viscous scale turbulence -while blade sections are also affected by larger scale turbulent motions, affecting the angles of inflow to blade sections in an unsteady way. Thus presence of turbulence and more general unsteady flow effects alter both mean and instantaneous power output from a device. Intensity and scale length spectra of inlet turbulence can be inferred or measured either a priori (i.e. before turbine installation) or during operation. Computational modelling, engineering assumption, correlations from lab and full scale data and machine learning and/or artificial intelligence can all be used to ascertain the effect of turbulence on yield both in the mean and in real-time or near-real time. Similar procedures can be carried our for other unsteady flow effects such as waves, internal waves and advection of thermoclines, etc. The classification of the resultant change in energy yield due to these effects are taken into the unsteady flow vs. yield response model, and includes both the effect of a particular turbulent spectrum on the mean yield, and the effect of individual (or groups of) turbulent structures on the power output of a device as they advect through it.

Having established this response, it can be used to inform siting of turbines as well as improve future energy yield estimates before turbine installation and during operation. Applied in real time or near-real time, the model informs the control unit of the effect of impending unsteady flow features on energy yield, allowing the device response to be dynamically tuned for optimal energy yield through unsteady events.

As explained above, the model can be refined (or established entirely, in the lack of an analytical, computational or experimental model as a basis) using machine learning or artificial intelligence. -17-

Unsteady flow vs. peak load response Similar to the unsteady flow vs. yield response model, the unsteady flow vs. peak load response model allows peak loading on a device to be predicted as a function of an unsteady flow characterisation, either a priori or in real time.

For example, the peak load caused by an identified unsteady flow feature (such as a wave or turbulent structure) about to impinge on a device can be computed from the response model in real or near-real time. The effect of changing turbine control parameters on the peak load caused by that event can be ascertained. Thus using the unsteady flow vs. peak load response model in real time, the control unit can alter control parameters to reduce the magnitude of a load condition about to occur or in progress.

The model can be refined (or established entirely, in the lack of an analytical, computational or experimental model as a basis) using machine learning or artificial intelligence.

Unsteady flow vs. dynamic range response Similar to the unsteady flow vs. yield response model, the unsteady flow vs. dynamic range response model allows the likely dynamic range (which can be expressed either as an absolute range or as a normalised value relative to the mean) of loading (or other parameters) on a device to be predicted as a function of an unsteady flow characterisation, either a priori or in real time.

For example, typical dynamic ranges caused by a particular sea state likely to occur in the region of a device (e.g. an impending storm event) can be computed from the response model in real or near-realtime. The effect of changing turbine control parameters on the dynamic range during the event is also calculable from the response model. Thus using the unsteady flow vs. dynamic range response model in realtime, the control system can alter control parameters during the storm event to improve the utility factor and/or lifetime of the device (both of which are dependent on dynamic range, see Dynamic range vs. Lifetime Awareness model').

Figure 8 shows an example where different turbines 3 (or turbine groups) within the same array are subject to different mean flow speeds and different turbulent flow characteristics. In this figure, a group of three turbines experiences a highly turbulent environmental flow while a neighbouring pair of turbines experiences less intense turbulence. More severe turbulence (higher intensity at the important scales) can increase the dynamic range of the loadings experienced by devices. To compensate, a more conservative control strategy is put in place to relieve cyclic/unsteady loadings and extend component lifetimes. Less turbulent inflow on this side leads to lower fatigue rates and peak loading on devices. A less conservative control strategy can be applied to these devices, -18-favouring yield over fatigue life. This maximises yield of the array and keeps lifetimes and wear rates consistent between all devices.

Fluid feature and debris advection model As explained above, the wave propagation model ascertains, for a given wave event at the boundary of an array or some distance from a device, the time taken for the wave to propagate to the device.

In a very similar way, the fluid feature and debris advection model relates a fluid event at one turbine or sensor (the passage of a patch of turbulence or item of debris) and uses an advection model to ascertain the later time at which the fluid event (or an evolved version of it) advects through another device location.

The advection model can be based on computational simulations (e.g. from software such as of the marks Telemac or Mike2l) by observing flow through the field then computing Lagrangian trajectories originating at all device or sensor locations. Timescales can be derived from the Lagrangian trajectories between the originating points and the closest passing point to relevant locations downstream.

One possible method of determining relevant locations is to take a Lagrangian trajectory and use a cone whose angle is govemed by the turbulent viscosity of the fluid (i.e. the rate at which turbulent diffusion occurs at the appropriate Reynolds number). Any devices residing in the downstream cone may be considered to be at risk of debris impact, or affected by turbulent features, where debris or turbulent features pass through the originating location of that trajectory. The cone may have a non-singular radius at the originating location govemed by a characteristic scale length of the debris or turbulent feature. In an example implementation, an item of debris is identified (or impacts) at a device location on the upstream boundary of an array. The debris continues to advect through the array with the flow. Using the fluid feature and debris advection model, the devices downstream which may be affected can be identified, and the time at which they are likely to be affected estimated, in real or near-realtime. Control parameters for these devices can be adjusted to respond appropriately to the impending event.

Awareness models The awareness models apply desirability or constraint functions to core engineering metrics, allowing the implications of a changing metric to be evaluated. These awareness models are often computed a priori to device deployment but may be updated through-life. They can be used in real time in conjunction with the above models to help inform the control system of the most desirable response to impending and current events. -19-

As shown in Figure 9, external data 7 may be fed into the CCADA control system 300, possibly via the facility 4. Environmental data (such as tidal chart data, or weather forecast) are typically provided via live feed, or a periodically updated database and/or lookup table. Models for costs and constraints of logistics, operations and power production can be uploaded from head office' or a designated authority, allowing the cost/energy ratio value to be optimised by the FCU 300. The device, e.g., turbine, model provides the expected performance responses per designer's analyses, maintenance requirements and expected lifetimes (such as Failure Mode Effects Analysis (FMEA) data).

The awareness models may include at least of the following.

Peak allowable load awareness model The peak allowable load awareness model is informed by engineering simulations and calculations of the structural and electrical properties of the device (and network). As named, it is used to define the allowable peak load for a given structure (beyond which failure occurs or insufficient safety factor is maintained).

Set point response The set point response may take into account models of steady state individual device performance and/or maps of safety or operational constraints.

Lifetime awareness model The lifetime awareness model is informed by engineering simulations and calculations of the structural and electrical properties of the device (and network). It is used to define the relationship between engineering metrics (such as time histories of loading, impulsive loading events, peak load values, mean load values, and dynamic ranges) and the expected product lifetime and maintenance intervals.

Economic awareness model The economic awareness model uses economic analysis (incorporating for example revenue from sale of power, subsidies, contractual penalties, and cost of capital) and considerations of operational constraints (such as vessel cost and availability, weather windows and port operations) in order to relate engineering performance metrics to costs and financial returns associated with operating an array.

Operational and/or logistical awareness model The operational and/or logistical awareness model uses operational costs and constraints (such as vessel cost and availability, weather windows, port operations, knowledge of scheduled maintenance operations) in order to ascertain cost and ability to undertake marine operations at given times.

Environmental data response As shown in Figures 2 and 9, the environmental data response may take into account wide area network data or external environmental data 7, such as tidal chart and/or weather forecast and/or sea state forecast.

-20 -As shown in Figures 2 and 9, the CCADA control system or unit 300 may also take into account data 6 from at least one local area sensor, such as coming from e.g., an Acoustic Doppler Current Profilers (ADCF) (far upstream) and/orwave height and direction sensors (far field).

It is known that Acoustic Doppler Current Proffers (ADCP) have some ability to measure turbulence, and they can be used in combination with a unit according to the disclosure in order to derive turbulent quantities from a mean profile provided by the ADCP, mean flow profiles converging more rapidly than turbulence metrics. The ADCP may thus enable taking into account the issue of turbulence, along with wave propagation, meteorological effects and many other (typically unsteady) fluid flow and environmental effects, in the generation of power from wind turbines, tidal stream turbines and wave energy converters. As a result, the disclosure may be applied in the fields of wind and tidal power engineering for assessment of available resource, energy yield and structural loading characteristics.

All awareness models can be simplistic (e.g. take a single parameter of maximum peak load allowed) or advanced in nature (e.g. adjust the peak load allowable depending on the time history of the loading and other metrics from elsewhere in the network, or a combination of variables).

In an example application of the awareness models, collection of engineering metrics from an operating device takes place in real (or near-real) time by the device's control unit. The awareness models are used (again in real or near-real time) to convert between these metrics and time-to-maintenance and/or time-to-failure estimates, which can in turn be used in calculation of cost metrics (which may include cash flow parameters, cost of energy, ROl metrics, etc.). The control unit compares these estimates to target values, and uses the response models to update control parameters (i.e. provide a control response) in order to optimise the output of the farm to meet the operator's requirements.

Figures 1 OA and 1 OB illustrate exemplary applications of the models for tidal array control.

The example of Figure iDA relates to a debris impact. In that example, an item of debris impacts a device 3 (e.g., "Turbine 1") atti, causing blade breakage (seethe interrupted line in Figure bA). The module 3031 implementing the Array-As-A-Sensor model uses real time feature recognition implemented by module 303 on Turbine 1, in order to identify the event. The module 303 implementing the fluid feature and debris advection model in the module 3031 identifies which other devices (e.g., Turbine 2") in the array may be affected by the debris advecting through the site, and identifies a danger period AT from the likely time (given by At) at which debris will reach each device location (i.e. Turbine 2). The module 3033 implementing the Operational model for example may clearly indicate that potential damage is to be avoided due to the cost of maintenance far outweighing lost revenue for the danger period (likely to be a constraint hard coded within the Operational model module 3033 and not requiring any financial calculation in this clear-cut case). The module 3032 -21 -implementing the Array-As-A-Device model is temporarily overridden by the module 301 implementing the reconciliation algorithm in the MLCS unit 300, since the Operational model applies a constraint. The unit 300 thus outputs a command in order to perform a managed shut down of affected devices (i.e. Turbine 2) for their danger periods before bringing them back to full power (see solid line in Figure bA). The dotted lines show that without the disclosure, Turbine 2 continues to operate and is at risk of similar damage from a convecting object (i.e. dotted line similar to interrupted line).

Once devices which remained intact are back online, a module 3034 implementing a thrust/mass flux distribution response model, along with allowable peak load and dynamic range awareness models, in the Array-As-A-Device module 3032 may utilise to redistribute thrust in the array in order to maximise yield without violating engineering constraints. The Operational module 3033 may interact with the Array-As-A-Device module 3032 to determine whether to maximise yield using thrust redistribution as shown in Figure 7B (penalising lifetimes of the remaining devices) until the next planned maintenance period, or schedule unplanned maintenance to recover lost revenue (and if so indicate the window in which it must be scheduled), or whether to sacrifice yield in order to meet desired lifetimes and maintenance intervals.

The example of Figure lOB relates to an energetic gust event. In that example, an energetic gust impinges on Turbine I at tI (see interrupted line). The load spike can be similar in magnitude to the impact of Figure bA, but the signature of the feature is different. The module 3031 implementing the Array-As-A-Sensor model uses real time feature recognition implemented by module 303 on Turbine 1, in order to identify the event. The module 303 implementing the fluid feature and debris advection model in the module 3031 identifies which other devices (e.g., "Turbine 2") in the array may be impinged by the gust through the site, and identifies an energetic period AT from the likely time (given by At) at which the gust will reach each device location (i.e. Turbine 2). The unit 300 may then output a command in order to dynamically tune Turbine 2 to the gust as it arrives (see solid line), capturing more energy from the gust and reducing dynamic load range. The dotted lines show that without the invention, Turbine 2 continues to operate has a similar response to Turbine 1.

In another example (not shown in the Figures) a large wave propagates. In that example, a wave is recorded by a sensor buoy at the edge of an array. The Array-As-A-Sensor module 3031 applies Real Time Feature Recognition to the sensor buoy data, ascertaining wave amplitude, period and direction.

The wave propagation model (with knowledge of the bathymetry) ascertains the times at which that wave will arrive at each device in the array. The Array-As-A-Device module 3032 uses the unsteady flow response models to estimate the effect the impending wave will have on the operational state of each device in terms of peak load, energy yield, etc. The effect is further estimated for altered device control responses to determine how altering the device response affects peak load, energy yield, etc., given that the event is about to occur. The Operational model module 3033 indicates whether priority for each device is presently on lifetime preservation or power generation. If reaching the end of a 4- -22 -hour energy sale block contract with the energy quota not yet fulfilled, priority will be to maximise yield. If located in a particularly turbulent zone compared to other turbines, and needing to extend the lifetime to meet planned maintenance operations, priority will be to extend lifetime.

The reconciliation module 301 may take account of AAAD, Operational model and AAAS inputs to adjust the devices responses as recommended by the AAAD, noting the weighting preference between yield and lifetime indicated by the Operational model, over the time periods indicated by AAAS.

The above embodiments are to be understood as illustrative examples of the invention.

Further embodiments of the invention are envisaged. It is to be understood that any feature described in relation to any one embodiment may be used alone, or in combination with other features described, and may also be used in combination with one or more features of any other of the embodiments, or any combination of any other of the embodiments. Furthermore, equivalents and modifications not described above may also be employed without departing from the scope of the invention, which is defined in the accompanying claims.

Claims

-23 -CLAIMS1. A control unit (300) configured to control an operation of an array comprising at least two devices (3) adapted to use kinetic energy of a stream of fluid to produce electrical energy, the unit being configured to: determine a state of the array using real-time data relating to an operation of at least one of the devices of the array; modify the operation of at least one device (3) of the array, as a function of the determination.
2. The unit (300) of claim 1, wherein the real-time data refer to at least one of the following parameters: torque, thrust, rotations per minute, shaft strain and/or shaft stress.
3. The unit according to any one of claims 1 or 2, further adapted to use external data, such as environmental data and/or the economic, operational and logistic data.
4. The unit according to any one of claims 1 to 3, further configured to: output a real-time control command for the at least one device (3), as a function of real-time data relating to an operation of at least two devices (3) in the array.
5. A control unit (300) configured to control an operation of an array of at least one device (3) adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least two algorithm modules (303), each module (303) being configured to generate at least one real-time control response for at least one of the devices (3), by implementing, a respective control algorithm; and a reconciliation module (301) configured to reconcile the real-time control responses in order to output the real-time control command.
6. The unit according to claim 5, configured to: generate, as a function of real-time data relating to the operation of all the devices (3) in operation in the array, at least one real-time control response for all the devices (3) in operation in the array; and output a real-time control command for all the devices (3) in operation in the array.
7. The unit according to any one of claims 5 or 6, wherein the algorithm module (303) is configured to implement at least one of the following control algorithms: Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response, Wave propagation, Thrust/mass flux distribution response, Unsteady inflow vs. yield response, Unsteady flow vs. peak -24 -load response, Unsteady flow vs. dynamic range response, Fluid feature and debris advection, Awareness model.
8. The unit according to claim 7, wherein the awareness model includes at least one of the following: Set point response, Peak allowable load awareness model, Lifetime awareness model, Economic awareness model, Operational and/or logistical awareness model, and/or Environmental data response.
9. A control unit (300) configured to control an operation of an array of at least one device (3) adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least algorithm module (303) configured to implement at least one of the following control algorithms: Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response.
10. The unit according to claim 9, wherein the algorithm module (303) is configured to generate at least one real-time control response for at least one of the devices (3), by implementing the at least one control algorithm.
11. The unit according to any one of claims 9 or 10, wherein the algorithm module (303) is further configured to implement at least one of the following control algorithms: Wave propagation, Thrust1mass flux distribution response, Unsteady inflow vs. yield response, Unsteady flow vs. peak load response, Unsteady flow vs. dynamic range response, Fluid feature and debris advection, Awareness model.
12. The unit according to claim 11, wherein the awareness model includes at least one of the following: Set point response, Peak allowable load awareness model, Lifetime awareness model, Economic awareness model, Operational and/or logistical awareness model, and/or Environmental data response.
13. The unit according to any one of claims S to 12, further comprising at least one optimisation module (304) configured to implement Machine Learning and/or Artificial Intelligence to the algorithm module (303).

-25 -
14. The unit according to any one of claim 5 to 8, further comprising an optimisation module (302) configured to implement Machine Learning and/or Artificial Intelligence to the reconciliation module (301).
15. The unit according to any one of claims I to 14, implemented at least partially as software or firmware -
16. The unit according to any one of claims 1 to 15, implemented at least partially in a physical casing, separate from the devices.
17. A system (1) comprising: an array of at least two devices (3) adapted to convert kinetic energy of a fluid into electrical energy; and at least one control unit (300) according to any one of claims 1 to 16.
18. The system according to claim 17, comprising at least two redundant control units (300).
19. The system according to any one of claims 17 or 18, wherein the device comprises a wind turbine, a tidal stream turbine or a wave energy converter.
20. A method for controlling an operation of an array of at least two devices (3) adapted to use kinetic energy of a stream of fluid to produce electrical energy, comprising a control unit (300): determining a state of the array using real-time data relating to an operation of at least one of the devices of the array; modifying the operation of at least one device (3) of the array, as a function of the determination.
21. A method for controlling an operation of an array of devices (3) adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least two algorithm modules (303) generating at least one real-time control response for at least one of the devices (3), by implementing, a respective control algorithm; and a reconciliation module (301) reconciling the real-time control responses in orderto output the real-time control command.
22. A method for controlling an operation of an array of devices (3) adapted to convert kinetic energy of a fluid into electrical energy, comprising: at least algorithm module (303) implementing at least one of the following control algorithms: -26 -Array-As-A-Sensor, Array-As-A-Device, Operational response, and/or Real-time feature recognition response.
23. Computer readable storage media having instructions stored thereon which are operable to program a programmable processor to carry out a method according to any one of claims 20 to 22 and/or to program a suitably adapted computer to provide a unit according to any one of claims 1 to 16.