A system and method for monitoring and optimizing the performance of renewable energy production facilities

Description

2 SCADA (Supervisory Control and Data Acquisition) systems already exist on the market that enable monitoring, alarms, and basic analyses. A limitation of such systems is that they generally do not perform coordinated analysis across multiple areas or climate zones, 5 as a result of which knowledge and optimization insights remain fragmented. The invention that is the subject of this patent application aims to provide a solution to at least one of the aforementioned disadvantages by means of 10 an overarching analysis framework for renewable energy installations. The invention relates to a system for overseeing and analyzing the performance of facilities for producing renewable energy, in which the 15 system uses virtual models, data analysis, and coordination mechanisms to generate performance insights across multiple areas. Although the invention can work with self-cleaning solar panel systems, it is not limited to such 20 systems and can be applied to variousconfigurations of renewable energy production facilities where local controls use adjustable parameters. 25 There is a need for targeted coordination at fleet level of the cleaning, inspection and maintenance of solar and wind farms. In addition, there is also a need for digital systems that improve the efficiency of solar panel and wind farms by accurately predicting efficiencies and wear. BE2025 / 7131 3 The invention relates to a system for monitoring and optimizing the performance of renewable energy production facilities, where the system comprises: 5 -a virtual model of each of the one or more renewable energy production facilities consisting of sensor data from the local controls; -one machine learning module comprising physical models10 or a machine learning module for each virtual model comprising physical models; -an advice / parameter output module configured to forward actuation parameters to the local 15 controllers;where the machine-learning-module is configured to perform simulations based on the virtual models in order to determine the actuation-parameters, where the advice-20 / parameteroutput-module is configured to forward the actuation-parameters to the local controls; where the system is further equipped with a feedback loop module configured to analyze the difference between the simulations of the machine learning module and measured performance of the local controls and to send a request for recalibration to the machine learning module, where the system is equipped with a fleet analysis module configured to compare performance and environmental data from multiple renewable energy production facilities, derive global performance indicators and generate standard parameters for use in model fittings and comparison of different virtual models, without determining or executing operational cleaning actions. An advantage of this is that a system according to theThe invention can determine how the cleaning of a facility for producing renewable energy can be optimally performed based on physical boundary conditions such as acoustic limits, coating wear, and energy consumption. An additional benefit is that less wear will occur because the local control is informed of the safe operational cleaning margins. An advantage of the system according to the invention is that it can supervise multiple facilities for producing renewable energy simultaneously. It is important to note here that the machine learning module learns and the system uses the learned result. The machine learning module executes the learning process, adjusts model parameters and / or model weights, learns an optimal target measure, and does this exclusively using a model. The system, as a whole, comprises the machine learning module,30 uses the KPI learned by the machine learning module and BE2025 / 7131 5 deploys these for supervision, comparisons and recalibration; but the system will otherwise not control or make operational decisions.In a typical implementation form of the invention, the 5 actuation parameters shall contain one or more of the following optimization parameters: the duration (t) of the vibrations, the frequency (f) of the vibrations and amplitude (A) of the vibrations, and where optionally these optimization parameters are determined by the system model when configuration proposals are passed on to local controls, without the system itself directly determining or activating operational cleaning actions. 15 In one implementation form of the invention, the sensor data is processed by the local control via an edge module and the sensor data contains information on the degree of dirt, irradiation, humidity, wind speed and temperature. The virtual model can also include data on the measured performance of one or more renewable energy production facilities. In typical implementation forms, the virtual model can be updated at intervals of, for example, 1 to 5 minutes. In the preferred implementation form, the time between measurements is theforwarding of the data to the virtual model preferably shorter than approximately 10 seconds. In a common implementation form, the spatial resolution of the virtual model for a 30 solar panel installation, for example, can be smaller than 1 m² BE2025 / 7131 6 per model unit. In an implementation form of the invention, the devices for producing renewable energy shall comprise photovoltaic panels and / or wind turbines.5 An advantage of a system according to the invention is that the energy yield is estimated at +3-6% for solar panels and +1-2% for wind turbines due to better actuation parameter selection.10 In an implementation form of the invention, the system is located on a cloud infrastructure and / or on an edge-based system or implemented with a hybrid architecture.15 An advantage of this is that the system can be built modularly and is compatible with existing infrastructures. In addition, it can utilize the computing power of a cloud infrastructure, allowing multiple resites to be supervised simultaneously. An additional benefit is that the system is capable ofto be able to trace or retrieve every actuation parameter that is forwarded to a local control. This is important for safety, compliance and quality control. In a preferred form of implementation of the invention, all data, including sensor data, actuation parameters and both measured and predicted performance, that are exchanged within the system and between the system and the local control, are securely encrypted. Standard open communication protocols such as MQTT / OPC-UA / HTTPS may be used for this purpose. 5 In the most preferred form of implementation, the system according to the invention shall be equipped with a fleet analysis module configured to compare the sensor data and measured performance of the various renewable energy production facilities and to detect deviations and / or patterns and / or determine averages, whereby the fleet analysis module is configured to determine standardized actuation parameters based on the data from the local controls and whereby the fleet-The analysis module forwards the standardized actuation parameters to the machine learning modules and / or advice / parameter output module. The actuation parameters form only one possible form of expression of the optimization insights calculated by the system. An advantage is that, according to the invention, the system is now capable of applying fleet learning. In the most preferred form of implementation, the system according to the invention shall be equipped with a fleet analysis module configured to compare the sensor data and measured performance of the various renewable energy production facilities and to detect deviations and / or patterns and / or determine averages, whereby the fleet analysis module is configured to derive, on the basis of the local control data, standardized reference values ​​or normalization values ​​BE2025 / 7131 8 relating to the model-described effect of optimized parameters, whereby these standardized reference values ​​orNormalization values ​​are forwarded to the machine learning modules and / or advice / parameter output module, for use in simulation and model tuning of actuation-related parameters, without the fleet analysis module determining or initiating operational cleaning actions. This results in scalable standardization and harmonization being possible across all renewable energy production facilities. An additional benefit is that the optimization of the machine learning module for predicting actuation parameters will occur faster thanks to data from the fleet analysis module. In a preferred implementation form, the machine learning module is configured to take into account at least one of the following elements when determining the actuation parameters: noise limit, vibration frequency limit, mechanical resonance of the renewable energy production facility, or a grid curtailment window. 25 By sound limit is meant a maximumpermitted noise level that can be caused by the actuation / vibration generation of the mechanical components that are coupled to or near the energy asset, or the installation for the production of renewable energy30 (e.g. photovoltaic structure, support frame, BE2025 / 7131 9 nacelle / wind turbine tower). It has to do with environmental standards or, in another example, occupational safety standards. For example, at a photovoltaic site located close to a residential zone, a noise limit of, for example, 50 decibels at night may be required by legislation. In that case, the system will select actuation values ​​that will not exceed the noise limit. 10 Vibration frequency limit here refers to the vibration frequency limit of the generated vibration input (deexcitation) applied to the asset. The vibration frequency limit is an acceptable frequency range for generated mechanical vibrations, in the asset or those coupled to the asset, which are limited by structural integritys / or interference with sensors / electronics. 20For example, for a photovoltaic module, a maximum limit of, for instance, 90 Hertz can be set because higher frequency micro-resonances can amplify singles or load connectors. The system will then limit the parameter “f” to a safe range.25 In this way, it is ensured that the machine learning module takes into account not only the performance of the renewable energy production facilities, but also physical or other30 external constraints and parameters. BE2025 / 7131 10 Preferably, the system comprises a coating model that detects a change in surface tension or wetness of the renewable energy production facilities, whereby the machine learning module derives a model-based adjusted value for an actuation parameter, in particular a vibration frequency (f), without the system performing a vibration activation or determining a cleaning time 10 The detection and interpretation of changes in surface tension / wetness is performed by the machine-feedback-loop-module, based on indirectly measured performance indicators and / or sensor data. The adjustment of the actuation parameter frequency “f” occurs as an optimization proposal, not as direct activation. Preferably, the feedback-loop-module is configured to send the request for recalibration based on deviation trends over a predetermined time span or across multiple renewable energy production facilities. This has the advantage that the recalibration will take into account historical data, measurements and / or data, in order to perform a more accurate recalibration. In a practical implementation form, the feedback loop module is configured to detect a need for recalibration independently of each virtual model,30 and to that end generate a recalibration request BE2025 / 7131 11, whereby the machine learning module is configured to recalibrate the virtual models independently of each other based on this recalibration request, and whereby the fleet analysis module compares global performance indicators to5generate standard parameters for use by various virtual models. The global performance indicators are aggregated, asset-transcending numbers that describe the performance behavior of multiple energy assets together, independent of individual cleaning actions or activation actions. They are used by the fleet analysis module to detect trends across multiple installations, to derive normalization and reference values, and to generate standard parameters for model recalibration. They decide nothing and they activate nothing. They serve only for model tuning and comparison. Example of a global performance indicator: the average rate of performance recovery after rainfall, measured across multiple installations within the same region. Preferably, the system is configured to calculate a net energy gain model-based, where the net energy gain (NE) can be calculated model-based as follows: NE = ∆Erecovered – Eequiv – Pcoating, where "∆Erecovered" represents a model-based estimated additional energy yield that has the potential tocan be recovered by reducing pollution or performance loss, where “Eequiv” represents a model-determined energy equivalent associated with a hypothetical deactivation event, and where “Pcoating” encompasses a penalty term representing the long-term impact of wear on an existing surface coating, and where a machine learning module (6) learns as a model parameter based on the calculated net energy gain and re-actuation target measure or frequency, without the system (1) determining an activation instruction or a cleaning time. Preferably, the system is configured to calculate a net energy gain model, where the net energy gain is equal to the extra amount of electrical energy (kWh) that can potentially be recovered, reduced by an estimated energy equivalent associated with an assumed deactivation event, and reduced by a penalty term representing coating wear. In an implementation form, for a given asset, or thefacilities for the production of renewable energy, and for a given evaluation period a net energy gain calculated by, on the one hand, modeling the additional25 energy yield that can be recovered by reducing pollution or performance loss, and on the other hand subtracting from this a model-estimated energy consumption associated with a potential activation, as well as a wear term that reflects the impact on an existing30 surface coating. BE2025 / 7131 13 This wear term is preferably expressed as a coating wear coefficient that may depend on environmental conditions and on the number or intensity of assumed activation events, so that the net energy gain not only reflects the immediate yield5 but also model-enables a long-term effect on the behavior of the asset. The re-actuation frequency or target measure derived from a machine learning module is used by the system exclusively as a model-based KPI for supervision and calibration, without the systemdetermines the execution of the activation instruction and a cleaning time. 15 Formula: NE=∆Erecovered–Eequiv–Pcoating; where: -NE=net energy gain;20 -∆Erecovered=model-estimated extra energy yield (kWh) that can potentially be recovered by reducing contamination or performance loss over an evaluation period; -Eequiv=model-estimated energy equivalent (kWh)25 associated with a presumed activation event (not executed, not controlled); -Pcoating=penalty term (kWh-equivalent) representing the long-term impact on an existing surface coating30, derived from a coating wear coefficient. All terms are model-based and the system does not activate anything and does not determine a cleaning time. BE2025 / 7131 14 5 A concrete example here is for a photovoltaic site in Europe, where, for an evaluation period of 30 days, ∆E recovered = 120 kWh (model predicts recovery after pollution reduction), Eequiv = 20 kWh (assumed light actuation) and Pcoating = 5 kWh10 (low wear under mild conditions).Using the above formula, we get: NE = 120 - 20 - 5 = 95 kWh.15 Interpretation is then model-based favorable, KPI points to high net energy gain without operational decision. It should be noted here that ∆E recovered is not a measured gain and is not a result of a cleaning performed, but the difference between two predictions, namely: ∆E recovered = E predicted, reference - E predicted, current;25 where: -Reference: ideal / expected behavior under comparable conditions -Current = modeled behavior with established performance loss.30 BE2025 / 7131 15 Sources for the estimate can be: historical performance curves, climate data, fleet comparison with comparable assets, coating model (effect of wetness / aging). 5 Another example in the Middle East (sand zone / desert zone): Reference: 800 kWh / 14 days; Current: 600 kWh / 14 days; So ∆E recovered = 200 kWh.10 It should also be noted here that the term Eequiv is the model-estimated energy equivalent (kWh) that represents the energy cost of a supposedactivation event. The system does not activate anything, is 15 no hardware control, and is an abstract cost item for optimization comparison. It is determined on the basis of the type assumed actuation (e.g. light versus intensive vibration), duration and 20 intensity (as hypothetical parameters), historical energy profiles of comparable systems, and safety and limit conditions. For example, a light actuation of short duration can have an energy equivalent of, for example, 17 kWh, an intensive actuation 25 (in a sand region) can have a higher intensity actuation with an energy equivalent of 34 kWh. It is also noted here that the term Pcoating is not actual energy consumption, but a normalized cost 30 that model-wise translates lifespan reduction, reduced hydrophobicities BE2025 / 7131 16 increased maintenance requirements into an energy measure. Pcoating=kxg(environment, intensity, frequency) where: 5 - K=coating-wear coefficient(dependent on material and dependent on age);-g(environment, intensity, frequency) = function of environmental conditions and assumed activation load.10 For example: Europe has a mild climate, with low wear and so Pcoating = 8 kWh equivalent. In contrast, there is sand in the Middle East15, with the consequence of higher mechanical and thermal loads, and so Pcoating = 25 kWh equivalent. Preferably, the system is configured to work together with existing facilities for the production of renewable energy via non-invasive sensors and standard communication protocols. This has the advantage that the system according to the invention can easily be implemented on existing facilities. Preferably, the virtual model is integrated with a fleet analysis module configured to compare model-based performance forecasts of multiple facilities for the production of renewable energy and parameter standardization at the model level. In a practical implementation form, the feedback loop module is configured toto initiate a recalibration process in the machine learning module, whereby the machine learning module performs dynamic recalibration by increasing the weight of recent data and reducing older data. By assigning a higher factor10 to more recent data, data, or measurements than older data, data, or measurements, it is ensured that the most recent data have a greater influence on the recalibration. Preferably, the machine learning module is configured15 to recalibrate one or more virtual models in response to a recalibration request generated by the feedback loop module, by adjusting internal model parameters, such that the simulated performance of the virtual models corresponds better with measured20 performance of the corresponding renewable energy production facilities, whereby the machine learning module performs the recalibration without determining or initiating operational cleaning actions or activation actions,25 whereby the result of the recalibration consists ofupdated model settings used for simulation, parameter advice comparisons. Preferably compare fleet analysis module performance 30 and environmental data from multiple energy assets BE2025 / 7131 18 in different climate zones to determine climate-specific normalization factors, factors used as reference values ​​to recalibrate at least one prediction model of another asset via the machine learning module.5 Preferably, the normalization factor takes into account at least variations in background pollution, temperature profiles, humidity patterns, or wind speeds characteristic of a specific geographical region,10 where these variations are used exclusively for model-based parameter adjustment comparisons. Preferably, deviations between measured and predicted power outside a climate-specific tolerance range15 trigger the generation of a recalibration request, whereby model parameters are adjusted by the machine learning moduleare adjusted within predefined constraints. Preferably, the fleet analysis module detects correlations20 between environmental events (rain, wind gusts, sandstorms) at a first site and performance improvements or performance losses at a second site, and where these correlations are used as statistical reference information to model-adjust parameters of the prediction model25 for assets with similar environmental characteristics. Preferably, the output of this system consists of optimized model parameters, calibration values ​​or30 threshold settings for simulation purposes and BE2025 / 7131 19 supervision purposes, and not instructions for activating any cleaning mechanism. Preferably, the system only transmits model-based optimization parameters to local control module5 information or configuration advice and where all decisions regarding activation or timing of cleaning mechanisms are performed exclusively by a separate local control system. 10 The invention also concerns a method for overseeingand optimizing the performance of facilities for producing renewable energy using a system according to the invention characterized by the method consisting of:15 -retrieving the measured performance of the aforementioned facilities by the feedback loop module; -predicting the predicted performance of the aforementioned facilities using the Machine Learning module based on the virtual model and forwarding this to the feedback loop module; -calculating the difference between the measured and predicted performance using the feedback loop module; -sending a request to the Machine Learning module for recalibration if the said difference is greater than a predetermined threshold value. In certain implementation forms of the invention, the feedback loop module will take into account the measured performance over a predetermined time span.30 In one implementation form of the invention, the method BE2025 / 7131 20 includes an additional step. The additional step consists of generating new actuation parameters by therecalibrated Machine Learning module and the forwarding of the new actuation parameters by the advisory / parameter output module to the local controls.5 An advantage of this is that the machine learning module is continuously optimized and thus becomes more accurate in predicting the measured performance and generating actuation parameters.10 Another advantage of this is avoiding the need to manually recalibrate the machine learning module. With the insight to better demonstrate the features of the invention, the preferred form of implementation of a system according to the invention is described below as an example without any restrictive character, with reference to the accompanying drawings, in which: 20 Figure 1 schematically represents the system according to the invention with a general machine learning module; Figure 2 schematically represents the system according to the invention with different machine learning modules; 25 In Figure 1, a schematic system 1 is shown for monitoring and optimizing the performance of facilities 2 for the production of renewableenergy, which facilities may or may not be equipped with vibration-based cleaning devices,30 where such cleaning devices, if present, BE2025 / 7131 21 are equipped with one or more local controls, in which the system comprises the following: -a virtual model4 of each of the one or more facilities2 for producing renewable energy consisting of sensor data5 originating from the local5 controls3; The sensor data5 comprises physical information about the facilities2 for producing renewable energy recorded via local sensors. The information may, without being limited thereto, include data about the environment of the facility2 and the facility2 itself such as temperature, wind speed, irradiation, humidity and degree of dirt. 15 -one machine learning module6 which comprises physical models; The machine-learning-module6 generates actuation parameters 8 within certain boundary conditions. These boundary conditions20 relate to acoustic resonance and coating status without being limited thereto.The implementation of the invention shown in Figure 1 has one unique machine learning module6 that will generate the 25 parameters and predict the performance of various devices2 for producing renewable energy. A machine learning module6 can generate parameters and predict performance for a specific device2 based on a recalibration13 that occurs with sensor data5 of another device2. In this way, the machine learning module6 in this implementation of the invention can perform fleet learning. - an advice / parameter output module7 that is configured to transmit system-model-5 actuation parameters8 to local controls, 3 without the system1 itself performing or activating operational actions; Whereby the machine-learning-module6 is configured to perform 10 simulations based on the virtual models 4 in order to determine the actuation-parameters8, whereby the advice- / parameter-output-module7 is configured to forward these actuation-parameters8 to the local controls3.15Sensor data5 from the local controls3 serve as input variables for the machine learning module6. The machine learning module6 is configured to provide simulation as output variables, such as predicted20 performances9. Typically, the predicted performances9 correspond to the predicted power produced by a renewable energy production facility2. The machine learning module6 can also transmit boundary conditions25 to the advice / parameter output module7. The advice / parameter output module7 is configured to subsequently transmit these actuation parameters8 to the local controls3. 30 Whereby system1 is further provided with the feedback loop module BE2025 / 7131 23 which is configured to analyze the difference12 between the simulations of the machine learning module6 and measured performance11 of the local controls and3 and to send a request for recalibration13 to the machine learning module6.5 Furthermore, Figure 1 shows that system1 in this case, but not necessarily for the invention, is further providedis of a fleet analysis module14 that is configured to compare the sensor data5 and measured performance11 of the 10 different facilities2 for producing renewable energy and to detect deviations and / or patterns and / or determine averages, whereby the fleet analysis module14 is configured to determine standardized or global actuation parameters8 based on the data5 from the local controls3 whereby the fleet analysis module14 transmits these standardized actuation parameters8 to the machine learning module6 and / or advice / parameter output module7. 20 Typical of the fleet analysis module 14 is that it can engage in fleet learning based on the data 16 exchanged in the system 1. Typical input data 16 for the fleet analysis module 14 so that it can engage in fleet learning are performance, the measured performance, 25 environmental data, environmental data in cluster classification such as coastal zone or industrial zone, historical data across different facilities 2. Output data of the fleet analysis module 14 are, as examples, notrestrictive, insights and settings across the fleet30 such as standards, reference values, deviation detection BE2025 / 7131 24 between different facilities2, global parameter advice and parameter standardization. The fleet analysis module14 can, for example, demonstrate how the pollution level affects performance11 at different locations. For instance, the fleet analysis module145 can compare the impact of the pollution level on the coast with that inland. Additionally, the fleet analysis module14 can, by way of illustration, assess whether a specific facility2 systematically performs worse against a reference threshold determined across the entire fleet10. This can indicate defects. Furthermore, in this case, the system1 in the example of Figure 1 is located on a native cloud15. The fleet analysis module14 is able to use the computing power of the cloud15. Figure 2 shows a variant of a system 1 according to Figure 1, where in this case each virtual model 4 has its own machine learning module 6. 20 In this case each virtual model will send the sensor data 5 tothe relevant machine-learning-model6 transmit. Otherwise, the operation is similar to the operation of system 1 in Figure 1.25 Figure 1 and Figure 2 also show that, according to the invention, system 1 forms a closed feedback loop with the devices 2 for producing renewable energy, in which the given current travels the following path30: BE2025 / 7131 25 -the local controls3 transmit measured performance and sensor data to system 1; -system 1 processes this data within one or more virtual models, in which a machine-learning module and a feedback-loop module are used to analyze deviations5 between measured and simulated performance and to initiate model-based recalibration; -Based on the recalibration, updated model parameters and / or optimization indications are generated, which are made available to the local controls via an advice / parameter output module, without the system itself determining or activating operational actions. Figure 1 and Figure 2 also show a working method for theoversee and optimize the performance of facilities2 for the production of renewable energy using a system1 in accordance with the invention, characterized in that the method consists of: - retrieving the measured performance11 of the aforementioned facilities20 by the feedback-loop-module10; - predicting the predicted performance9 of the aforementioned facilities2 using the Machine-Learning-module 6 or machine-learning-modules6 based on the virtual models4 and sending these to the feedback-loop-module10; - calculating the difference12 between the measured11 and the predicted9 performance using the feedback-loop-module10; - sending a request to the Machine-Learning-30 module for recalibration13 if the said difference BE2025 / 7131 26 12 is greater than a predetermined threshold value. In addition to recalibration, the machine-learning module 6 or machine-learning modules 6, following a request from the feedback-back loop module 10, can also be internal due to the excessive difference between the measured and predicted performance.retrain machine-learning model or adjust the model weights. With the insight to better demonstrate the characteristics of the method according to the invention, an example is described below, as an example without any limiting character, of a central element of a device2 for the production of renewable energy consisting of a solar panel park2.15 The Machine-Learning module6 predicts an initial performance 9 of the produced power of the solar panel park being at least 9.2 MW on the basis of sensor data5 derived from the virtual model4 of the relevant solar panel park regarding the degree of dirt, irradiation, temperature, wind speeds and humidity. In addition, a correction equal to -0.3 MW is applied derived from the physical models that described here the influence of cooling by wind and the temperature dependence of the 25 power. The final predicted performance of the Machine Learning module is equal to 8.9M and is sent to the feedback loop module 10. The feedback loop module 10 receives 30additional measured performance11 or measured power11 BE2025 / 7131 27 of the solar panel park2 originating from the local control equals 8.6 MW. The measured power is physically measured by a sensor on the solar panel park2. The feedback loop module10 will calculate the error between the measured power and the predicted power, being 8.6 MW - 8.9 MW5 = -0.3 MW. Because the error is greater than a predetermined threshold, the feedback loop module10 will send a request to the machine learning module6 to recalibrate it13. The machine learning module6 will recalibrate itself13. After this, the machine learning module will predict better performance9 of the solar panel park2 for the following days. In addition to predicting performance9, the machine learning module6 can also make predictions regarding wear and tear for the solar panel park. 15 Figure 1 and Figure 2 show system according to the invention with the additional element that all data16 generated or obtained by the system are sent to a secure web interface17. For the application ofUnder this patent, data16 are understood, without this being an exhaustive list: sensor data, actuation parameters, and both predicted and measured performance. Operators gain access via this web interface17 to critical performance indicators including net energy gain, coating conditions, vibration load, acoustic25 reports. Although the patent describes the implementation of the invention based on solar panel parks, the invention is also applicable to wind turbines and / or combinations thereof. In that case, the system analyzes performance, BE2025 / 7131 28 environmental and operational data of such installations in order to derive indicative optimization parameters or reference values ​​via virtual models and analysis methods. These parameters are intended to be used by external or local control systems, without the system itself interfering with the operational functioning of the installation. The current invention is by no means limited to the example described and