Power generation equipment power prediction and state monitoring system based on real-time data fusion

Abstract

The invention relates to the technical field of photovoltaic power generation, and discloses a power generation equipment power prediction and state monitoring system based on real-time data fusion, and the system comprises a differentiable physical digital twin model which is used for predicting available power according to real-time meteorological data and internal health parameters; the internal state deviation decoupling and credit distribution module is used for calculating the deviation between a predicted value and an actual value and quantizing and positioning key health parameters causing the deviation; and the twinborn self-consistent correction and parameter updating module is used for carrying out closed-loop online updating on health parameters of the twinborn model by utilizing differentiable characteristics and credit distribution results of the model. According to the system, a physical mechanism and data driving are deeply fused, the problems that a traditional model is static and cannot be evolved are solved through a closed-loop self-consistent correction mechanism, the running state of the photovoltaic power station is accurately depicted and predicted, and the intelligent level of operation and maintenance decision making is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, specifically to a power prediction and condition monitoring system for power generation equipment based on real-time data fusion. Background Technology

[0002] With the transformation of the global energy structure, new energy power generation, represented by photovoltaics, has become an important component of the power system. However, photovoltaic power generation is characterized by significant intermittency and fluctuations, and its power output is affected by both weather conditions and the health status of the equipment itself. Therefore, accurate power forecasting for photovoltaic power plants is a crucial prerequisite for ensuring the safe and stable operation of the power grid and optimizing power dispatch decisions. Simultaneously, effective monitoring and diagnosis of the operating status of thousands of photovoltaic modules and related equipment within the power plant is of great significance for improving power generation efficiency, reducing operation and maintenance costs, and preventing major failures.

[0003] Currently, the technical approaches for photovoltaic power prediction and condition monitoring mainly fall into two categories. One category is based on physical mechanism models, which establish detailed physical and thermodynamic equations for photovoltaic modules, inverters, and other equipment, and solve them using real-time meteorological data. These models have good interpretability and can clearly reflect the impact of different physical parameters on power generation. However, their core drawback lies in the static nature of the model. The health parameters (such as equivalent thermal resistance, material aging coefficient, etc.) set at the beginning of the model's establishment are fixed values, which cannot reflect the actual performance degradation caused by aging, wear, dust accumulation, and other factors during long-term operation. This causes the predictive accuracy of the physical model to gradually decrease over time, resulting in a "twin deviation" between the model and the physical entity.

[0004] Another type is data-driven models based on artificial intelligence, such as deep learning networks. These models learn from massive amounts of historical operation and maintenance data, enabling them to dynamically capture patterns in power generation changes and often exhibiting superior prediction accuracy. However, their inherent "black box" nature makes them lack physical interpretability. When the model predicts an abnormal drop in power, it cannot inform maintenance personnel whether this drop is caused by abnormal component temperatures, inverter efficiency degradation, or other specific physical reasons. This lack of interpretability significantly limits their application value in refined operation and maintenance and fault diagnosis, preventing maintenance personnel from taking targeted maintenance measures based on the model's results.

[0005] Therefore, existing technologies generally face the contradiction between the "static inaccuracy" of physical models and the "black box blindness" of data models. How to deeply integrate the clear interpretability of physical mechanisms with the dynamic adaptive capabilities of data-driven approaches to build an intelligent model that can accurately predict, trace the source of diagnosis, and co-evolve with physical entities is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the aforementioned technical issues, this disclosure provides a power prediction and condition monitoring system for power generation equipment based on real-time data fusion, which solves the fundamental problems in existing photovoltaic models where physical models are inaccurate due to static parameters and data models are unable to make diagnoses due to a lack of interpretability.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a power prediction and condition monitoring system for power generation equipment based on real-time data fusion. This system can couple power prediction with condition diagnosis, and perform online correction of the health status parameters inside the model by analyzing prediction deviations, thereby simultaneously improving the accuracy of power prediction and the precision of condition diagnosis.

[0008] In one embodiment, the system includes a processor and a memory, the memory storing instructions executable by the processor, which, when executed, cause the processor to perform the following functions: A differentiable physical digital twin model is constructed. This model is a computational model that incorporates the key physical operating mechanisms of the power generation equipment. Internally, it defines a structured health parameter space for quantitatively characterizing the health status of the power generation equipment. The structured health parameter space It is a multidimensional vector, and its mathematical expression is: ; Each element ( Each of these is a health parameter with a definite physical meaning. In some embodiments, the health parameter may be the equivalent thermal resistance. Capacitor attenuation coefficient Inductance drift coefficient or maximum power point tracking efficiency offset At least one of them.

[0009] Calculate the internal state deviation signal This signal is used to characterize the difference between the output of the differentiable physical digital twin model and the actual operating state of the physical device. In one embodiment, the calculation process is as follows: First, real-time collected meteorological data... and the current structured health parameter space The input is fed into the differentiable physical digital twin model to calculate the predicted value of available power. Then, the predicted available power value Subtract the actual output power of the power generation equipment collected synchronously. The internal state deviation signal is obtained. The calculation formula is as follows: ; According to the internal state deviation signal The signal characteristics are the structured health parameter space. Health parameter allocation, credit allocation weight In one embodiment, the allocation process is as follows: First, the internal state deviation signal within the time window is... The sequence is processed to extract a signal feature vector. Subsequently, the signal feature vector Input to a pre-defined credit allocation network The output of the network, after being processed by the softmax function, yields a result consistent with the structured health parameter space. The same-dimensional, normalized credit allocation weights The calculation formula is as follows: ; Based on the credit allocation weight For the structured health parameter space Perform online updates. In one embodiment, the update process is as follows: First, the internal state deviation signal is... A function, such as its square, as the loss function. Secondly, calculate the loss function. For the structured health parameter space gradient Finally, the credit allocation weights are used. The gradient is weighted by a Hadamard product to update the structured health parameter space. Its update formula is: ; in, For learning rate, This is the Hadamard product operator.

[0010] In other embodiments, the system may also perform the following functions: inputting future weather forecast data and the updated structured health parameter space into the differentiable physical digital twin model to generate power generation prediction results; and outputting the time variation trajectory of the structured health parameter space as the condition diagnosis result of the power generation equipment.

[0011] A second aspect of the present invention provides a method for power prediction and condition monitoring of power generation equipment based on real-time data fusion, wherein the method is executed by the system described in any of the foregoing embodiments.

[0012] In one embodiment, the method includes the following steps: Step 1: Construct a differentiable physical digital twin model, wherein the differentiable physical digital twin model defines a structured health parameter space for quantifying the health status of the power generation equipment; Step 2: Calculate an internal state deviation signal, which is the deviation between the predicted available power output by the differentiable physical digital twin model and the actual output power of the power generation equipment. Step 3: Based on the signal characteristics of the internal state deviation signal, assign credit allocation weights to the health parameters in the structured health parameter space; Step 4: Update the structured health parameter space online according to the credit allocation weight.

[0013] In some embodiments, step three is specifically implemented as follows: performing time-domain, frequency-domain, or time-frequency-domain analysis on the internal state deviation signal sequence within the time window to extract the signal features; and inputting the signal features into a preset credit allocation network to output the normalized credit allocation weights with the same dimension as the structured health parameter space.

[0014] In some embodiments, step four is specifically implemented as follows: using the function of the internal state deviation signal as the loss function; calculating the gradient of the loss function with respect to the structured health parameter space; and weighting the gradient using the credit allocation weights to update the structured health parameter space.

[0015] In other embodiments, the method further includes steps five and six: Step five, inputting future weather forecast data and the updated structured health parameter space into the differentiable physical digital twin model to generate power generation prediction results; and Step six, outputting the time variation trajectory of the structured health parameter space as the condition diagnosis result of the power generation equipment.

[0016] In the method and system of the present invention described above, by constructing a differentiable physical digital twin model, the deviation between predicted and actual power is used to correct physically meaningful health parameters within the model in a closed-loop, online manner, enabling the model to age and evolve synchronously with the physical entity. This not only fundamentally solves the static inaccuracy problem of traditional models and improves long-term prediction accuracy, but also enables early warning of equipment failures and accurate prediction of remaining service life by real-time monitoring of the changes in these health parameters, providing strong decision support for predictive maintenance and asset health management of power plants. Furthermore, the method and system of the present invention are not only applicable to single devices, but can also be extended to power generation systems composed of multiple devices, improving the overall system's operating efficiency and stability through collaborative optimization.

[0017] The technical solution provided in this disclosure has the following advantages compared with the prior art: 1. This invention calculates the internal state deviation signal between the output of a differentiable physical digital twin model and the actual output power of the device, and uses this signal to update the structured health parameter space within the model online. This enables the prediction model to continuously self-correct, quantifying and reflecting the performance degradation caused by aging or wear in the model in real time, thereby ensuring that power prediction is always based on the current true health state of the device, significantly improving the accuracy and reliability of the prediction results throughout the entire life cycle of the device.

[0018] 2. This invention directly outputs the temporal variation trajectory of a structured health parameter space. Since each health parameter in this space (such as equivalent thermal resistance, capacitance attenuation coefficient, etc.) has a clear physical meaning, maintenance personnel can directly and quantitatively assess the performance degradation of critical components by observing the changing trends and amplitudes of these parameter values. This approach replaces the traditional "black box" diagnosis that relies on expert experience or complex signal analysis, providing an intuitive and interpretable diagnostic basis, which is beneficial for achieving accurate predictive maintenance.

[0019] 3. This invention extracts the signal features of internal state deviation signals and uses a credit allocation network to map specific features to credit allocation weights for different health parameters. This mechanism can intelligently determine the most likely physical source based on the "fingerprint" characteristics of the deviation signal (such as high-frequency noise or slow drift) and accurately apply update pressure to the corresponding health parameters. This avoids blindly or incorrectly adjusting parameters in multi-parameter models, achieving rapid and accurate attribution of the root cause of performance degradation. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0021] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the system functional modules of the present invention; Figure 2 This is a schematic diagram of the forward computation process inside the differentiable physical digital twin model of the present invention; Figure 3 This is a functional diagram of the internal state deviation decoupling and credit allocation module of the present invention; Figure 4 This is a schematic diagram of the twin self-consistent correction and parameter update process of the present invention; Figure 5 This is a schematic diagram of the combined output results of the present invention. Detailed Implementation

[0023] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0024] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0025] See attached document Figure 1 , Figure 1 This is a flowchart of a power prediction and condition monitoring method for power generation equipment based on real-time data fusion according to an embodiment of the present invention. In a specific embodiment, the method provided by the present invention may include the following steps: S100, Construct a differentiable physical digital twin model, wherein the differentiable physical digital twin model defines a structured health parameter space for quantifying the health status of the power generation equipment.

[0026] S200, Calculate the internal state deviation signal, which is the deviation between the predicted available power output by the differentiable physical digital twin model and the actual output power of the power generation equipment.

[0027] S300, based on the signal characteristics of the internal state deviation signal, assign credit allocation weights to the health parameters in the structured health parameter space.

[0028] S400, the structured health parameter space is updated online according to the credit allocation weight.

[0029] In one specific embodiment, refer to the appendix. Figure 1 The system may include: a data acquisition module, a differentiable physical digital twin model module, an internal state deviation decoupling and credit allocation module, a twin self-consistent correction and parameter update module, and a joint output module.

[0030] The data acquisition module is used to acquire data during the real-time operation of the power generation equipment. Specifically, the acquired data includes, but is not limited to, real-time meteorological data vectors. For example, solar irradiance, ambient temperature, etc.; and the actual output power of the power generation equipment. The data acquisition module transmits the acquired data to other modules of the system for processing.

[0031] During system operation, the differentiable physical digital twin model module executes step S100. This module contains a differentiable physical digital twin model, which is a computational model embedding the key physical operating mechanisms of the power generation equipment. The model internally defines a structured health parameter space. This space is a multidimensional vector used to quantify the overall health status of the device.

[0032] ; in, For the total number of health parameters, each element All of these are internal parameters with clear physical meaning that change as the equipment ages or wears out.

[0033] Subsequently, the system enters a continuous closed-loop correction and prediction process. At any given moment... The differentiable physical digital twin model module receives real-time meteorological data provided by the data acquisition module. It also incorporates the structured health parameter space for the current moment provided by the twin self-consistent correction and parameter update module. Perform forward calculations and output a theoretical predicted value of available power. The internal state deviation decoupling and credit allocation module executes steps S200 and S300. This module receives the available power prediction value output by the differentiable physical digital twin model module. and the actual output power provided by the data acquisition module. It first obtains the internal state deviation signal by calculating the difference between the two. : ; get Subsequently, the module extracts signal features from the deviation signal sequence within a time window and inputs the extracted signal features into a preset credit allocation network to calculate the correlation with the structured health parameter space. Credit allocation weight vector of the same dimension .

[0034] The twin self-consistent correction and parameter update module executes step S400. This module receives the credit allocation weight vector output by the internal state deviation decoupling and credit allocation module. and internal state deviation signal It is with A loss function is constructed based on this, and the loss function is calculated on the structured health parameter space based on the differentiability of the differentiable physical-digital twin model module. The gradient is then used to assign weights using credit. The updated health parameter space is calculated by weighting the gradient. 1).

[0035] Updated health parameter space It is then re-provided to the differentiable physical digital twin model module for calculating the available power prediction value at the next time step, thereby forming a dynamic closed-loop correction data stream.

[0036] The joint output module provides the final calculation results to the user or upper-layer application system. In one embodiment, when power forecasting is required, this module receives future weather forecast data and the latest health parameter space output by the twin autonomous correction and parameter update module. The system calls the differentiable physical digital twin model module to perform calculations and outputs the final power prediction results. Simultaneously, this module can also output a structured health parameter space. The time-varying trajectory is used as a diagnostic result for the equipment's condition.

[0037] See attached document Figure 2 , Figure 2 This is a schematic flowchart illustrating the forward computation within a differentiable physical digital twin model according to an embodiment of the present invention. This section provides a detailed description of the model construction and computation process. In one embodiment, taking a photovoltaic power generation device as an example, the detailed construction and computation process of the differentiable physical digital twin model module is as follows.

[0038] First, regarding the core of the differentiable physical digital twin model module, namely the structured health parameter space... This space is defined. It parameterizes key physical properties of the device that change slowly over time and operating conditions and are not directly measurable. In one specific embodiment, this health parameter space... This may include, but is not limited to, the following parameters: Equivalent thermal resistance This parameter characterizes the heat transfer capability between the photovoltaic module and its surrounding environment. An increase in its value indicates a decrease in the module's heat dissipation capacity, which directly affects the module's actual operating temperature and, consequently, its power generation efficiency.

[0039] Capacitor attenuation coefficient This parameter characterizes the degree of degradation of the effective capacitance of the DC-side support capacitor or AC-side filter capacitor in the inverter relative to its initial value. Capacitor aging leads to a decrease in its effective capacitance, affecting the stability of the DC voltage and the quality of the output AC power.

[0040] Inductance drift coefficient This parameter characterizes the deviation of the actual inductance value of the boost inductor or filter inductor in the inverter from its nominal value. Inductance drift affects the inverter's switching frequency, current ripple, and overall conversion efficiency.

[0041] Maximum power point tracking efficiency offset This parameter characterizes the efficiency deviation between the actual power point tracked and the theoretical maximum power point caused by factors such as sensor aging, control parameter mismatch, or non-uniform degradation of components in the maximum power point tracking (MPPT) control algorithm.

[0042] After defining the structured health parameter space, the differentiable physical digital twin model module acquires real-time meteorological data. (For example, solar irradiance) and ambient temperature and the structured health parameter space at the current moment. Then, through a series of physical formulas, sequential calculations are performed to obtain the predicted value of available power. The calculation process is completely differentiable.

[0043] The calculation process begins by determining the actual operating temperature of the photovoltaic module. This temperature is the sum of the ambient temperature and the temperature rise caused by the component absorbing solar radiation. The temperature rise is related to the equivalent thermal resistance. Directly related: ; in, For component operating temperature, For ambient temperature, Solar irradiance, This represents the equivalent thermal resistance health parameter at the current moment.

[0044] Secondly, calculate the ideal DC power output of the photovoltaic module at the current operating temperature. This power is affected by the component's operating temperature. The effect of this is calculated using the following formula: ; in, The total area of ​​the photovoltaic modules, Photoelectric conversion efficiency under reference conditions. Power temperature coefficient of the component This is a reference temperature. , , and These are all fixed parameters calibrated at the factory for the equipment.

[0045] Next, the actual DC power was calculated, taking into account the MPPT efficiency offset. This calculation process will take health parameters Introduction: ; in, The MPPT efficiency of an inverter is a fixed value. The MPPT efficiency offset health parameter is given at the current moment. Finally, the final AC output power after inverter conversion is calculated, which can be used as the power prediction value. Inverter conversion efficiency Affected by the health status of its internal components, these effects are quantified through health parameters. and This is reflected in the inverter efficiency, which can be modeled as a function of input power, DC voltage, and health parameters. ; function The specific form is pre-defined as a differentiable lookup table or a polynomial function, with the current actual DC power as its input. and health parameters and Through the above steps, the system completes a full forward calculation. Each step in this calculation process is a mathematically differentiable operation, thus ensuring the final predicted usable power value. Relative to structured health parameter space Each health parameter The gradients are all computable. This property forms the basis for subsequent gradient-based online updates.

[0046] Before describing the technical solution of this invention in detail, we first define the core concept of this invention—the "differentiable physical-based digital twin model." The differentiable physical-based digital twin model is a special type of digital twin model that possesses the following three core characteristics: Physics-based: The core architecture of the model is not a general "black box" (such as a pure neural network), but is built upon the physical laws and engineering equations (such as thermodynamic, electrical, and optical equations) that describe the operating mechanisms of physical entities (such as photovoltaic power generation equipment). This gives the model's internal parameters (such as thermal resistance and efficiency degradation coefficient) clear physical meaning, ensuring the model's interpretability.

[0047] Digital Twin Synchronization: This model keeps synchronized with the physical entity through real-time sensor data (such as actual power generation, temperature, irradiance, etc.), dynamically reflecting the actual operating status of the physical entity.

[0048] Fully differentiable: The entire computational process of the model, from environmental input to final output, is constructed to be mathematically continuous and differentiable. This means that the gradient (i.e., "responsibility" or "influence") of the prediction bias of the output with respect to any physical health parameter within the model can be precisely calculated. It is this characteristic that enables the use of optimization algorithms such as gradient descent to automatically and backward update and correct the health parameters within the model based on the prediction bias, thereby achieving online self-learning and self-evolution of the model.

[0049] In summary, the differentiable physical digital twin model is a "white box" model that can both explain physical processes and automatically learn and evolve from data like a neural network, by combining physical mechanisms with differentiable programming.

[0050] See attached document Figure 3 , Figure 3 This is a functional diagram of the internal state deviation decoupling and credit allocation module according to an embodiment of the present invention. This section provides a detailed description of the deviation calculation and credit allocation process.

[0051] The internal state deviation decoupling and credit allocation module calculates the internal state deviation signal and assigns credits to each health parameter in the structured health parameter space based on the characteristics of this signal. This module receives the available power prediction value output by the differentiable physical-digital twin model module. and the actual output power provided by the data acquisition module. It first obtains the internal state deviation signal by calculating the difference between the two. : ; This deviation signal In a physical sense, power fluctuations caused by changes in external weather conditions are stripped away, and the values ​​directly reflect the inherent inconsistency between the current parameterized physical digital twin model and the actual state of the physical entity.

[0052] To accurately attribute the root cause of the deviation signal, the system needs to analyze the deviation signal sequence within a preset time window. The process involves extracting a signal feature vector that characterizes the sequence properties. .

[0053] In one embodiment, the feature vector It may include the following components: Time-domain characteristics: For example, the mean, variance, kurtosis, or kurtosis of the bias signal sequence. The mean can reflect a systematic, slowly changing prediction bias, which is related to the equivalent thermal resistance. The gradual increase or maximum power point tracking efficiency shift The continuous accumulation of these factors is correlated; variance reflects the degree of fluctuation in deviation, which is related to factors such as DC bus voltage instability and may indicate capacitor attenuation. .

[0054] Frequency domain characteristics: For example, the energy distribution within a specific frequency band obtained by performing a Fast Fourier Transform (FFT) on the bias signal sequence. Energy anomalies in the low-frequency band can indicate slow performance drift, while energy enhancements in specific high-frequency bands can indicate anomalies related to the inverter switching frequency, which is related to inductor drift. Or capacitor aging They are related.

[0055] Signal feature vector After being extracted, the data will be input into a pre-defined credit allocation network (CAN). In one embodiment, this credit allocation network is a multilayer perceptron (MLP) network, denoted as... This network, trained offline, has established a model based on signal feature vectors. The network receives a complex nonlinear mapping relationship between the contribution of various health parameters to this deviation. As input, and output a structured health parameter space Intermediate vector of the same dimension .

[0056] Finally, to obtain a normalized credit assignment vector that can be directly used as weights. The system will use the intermediate vector This is processed using a Softmax function. The calculation formula is as follows: ; in, It is a credit allocation vector The Each component corresponds to a health parameter. Credit weight; It is the intermediate vector The One component; This represents the total number of health parameters. The processed credit allocation vector is... In this vector, the sum of all components is 1, and the value of each component is between 0 and 1. Each component That is, it quantifies the health parameters determined by the system. Mismatch with current internal state deviation The proportion of responsibility that should be borne.

[0057] In the method and system for predicting the power generation of photovoltaic power generation equipment according to embodiments of the present invention, real cloud images are used to obtain the light transmittance of clouds at the initial moment, and physical equations considering cloud advection and cloud source-sink are used to describe cloud changes. However, the essential difference between the present invention and the prior art lies in that it further constructs a dynamically evolving, differentiable physical digital twin model capable of closed-loop self-correction. This model not only uses external variables predicted by numerical weather prediction models, but more importantly, it also incorporates the physical health parameters inside the equipment as part of the model, and utilizes the deviation between the predicted power and the actual collected power to continuously optimize these health parameters online through the characteristic of full differentiability. This allows the model to "age" along with the physical equipment, thereby maintaining high prediction accuracy over a long period.

[0058] This method also includes a fault warning step, which can be implemented in at least one of the following two ways: Early warning based on health parameter thresholds: This involves setting normal baseline ranges or safety thresholds for one or more key health parameters in the model under healthy conditions. During system operation, the real-time values ​​of these parameters, updated online by the model, are continuously monitored. When the value of a parameter (such as the thermal resistance of the cooling system) continuously deviates from its healthy baseline and exceeds the preset safety threshold, the system automatically generates and issues a fault warning signal, indicating potential specific problems such as blocked cooling channels or fan failure.

[0059] Early warning based on prediction deviation trend: Monitor the time series of power prediction deviation. Under healthy equipment and good model fit, this deviation should fluctuate randomly and slightly around zero. If the system encounters a new fault mode that the model has not learned, it will cause the prediction deviation to increase continuously and unidirectionally or oscillate violently. By calculating the moving average, variance, or growth gradient of the deviation, the system can also issue a "model mismatch" warning when these statistical indicators exceed the normal range, prompting manual inspection.

[0060] Implementation of Remaining Useful Life (RUL) prediction function: This method may also include a lifetime prediction step. Since the model can output a trajectory of key health parameters (such as the photoelectric conversion efficiency decay coefficient) changing continuously over time, this trajectory accurately records the performance degradation process of the device. This step includes: trend modeling of selected time-series data characterizing aging key health parameters, for example, using a multinomial regression or ARIMA model to fit their degradation curves. Then, this degradation curve is extrapolated into the future to predict the future time point at which the parameter will reach its preset "failure threshold" (e.g., efficiency dropping to 80% of its initial value). The difference between this predicted time point and the current time is the device's predicted remaining useful life (RUL).

[0061] This invention can be extended from the application of a single device to a power generation system consisting of multiple devices: Multiple power generation devices.

[0062] Multiple prediction and diagnostic units are included, each a differentiable physical digital twin model constructed based on the aforementioned embodiments, corresponding one-to-one with a power generation device. Each unit operates independently, generating power prediction data and internal health status parameters of its corresponding device in real time.

[0063] A central collaborative optimization platform. This platform is configured as follows: Data aggregation: Power prediction curves and quantified health status data from all prediction and diagnostic units are received in real time via the network.

[0064] Intelligent decision-making and collaborative control: Based on aggregated global information, execute system-level optimization strategies. For example: Optimal load allocation: Under the premise of meeting the grid dispatch requirements, priority is given to the equipment units with the best health and the strongest expected power generation capacity to operate at full load, while the load of equipment units that have shown early signs of decline is appropriately reduced to delay their aging. In this way, while ensuring the current total power generation, the long-term health of the entire system and the cumulative power generation over the total life cycle are maximized.

[0065] Predictive maintenance scheduling: Based on the fault warning information and remaining service life prediction results of all equipment units, maintenance tasks are prioritized and an optimal maintenance plan is generated. Limited operation and maintenance resources (manpower, spare parts) are allocated to the most urgent equipment first, realizing the transformation from "scheduled maintenance" to "on-demand maintenance", thereby minimizing operation and maintenance costs and downtime losses.

[0066] See attached document Figure 4 , Figure 4 This is a flowchart illustrating the twin autonomous correction and parameter update process according to an embodiment of the present invention. This section provides a detailed description of the online update process for the structured health parameter space. The twin autonomous correction and parameter update module functions to update the structured health parameter space based on internal state deviations and its credit allocation results. Perform online updates. This module receives the internal state deviation signal output by the internal state deviation decoupling and credit allocation module. and credit allocation weight vector .

[0067] First, the module uses internal state deviation signals. Based on the square of the sum, construct an instantaneous loss function. This loss function is used to quantify the degree of inconsistency between the model's prediction and the actual output at the current moment. In one embodiment, the loss function can be defined as: ; Minimize this loss function In other words, aligning the model output with the actual output is physically based on correcting the health parameters within the model. This is to make it closer to the real state of a physical entity. Next, the system utilizes the inherent differentiability of the differentiable physical-digital twin model module described in Part Two to calculate the loss function. For the current structured health parameter space Each health parameter The partial derivatives of these derivatives form the gradient vector. This calculation can be achieved by performing automatic differentiation or backpropagation algorithms throughout the entire forward computation chain.

[0068] ; This gradient vector This indicates that the loss function can be increased most quickly in a multidimensional health parameter space. Therefore, adjusting the parameters in the opposite direction of this gradient can effectively reduce the loss.

[0069] Finally, to achieve precise control over the update process, the system uses the credit allocation weight vector calculated in Part III. For gradient vector Weighting is applied. Credit weights are added to the gradient using the Hadamard product (element-wise multiplication) operation, combined with a pre-defined learning rate. Then, perform an update to the health parameter space. The update rules are as follows: ; in, This is the updated health parameter space; It is the health parameter space at the current moment; It is a scalar learning rate used to control the step size for each update; It is the credit allocation weight vector; © represents the Hadamard product operation.

[0070] Through this weighted update mechanism, only when health parameters... Corresponding credit weight When the value is large, its corresponding gradient component is allowed to significantly influence parameter updates. If the credit weight is close to zero, the health parameter remains essentially unchanged in this iteration. This ensures that update pressure is accurately applied to the health parameter identified as the main source of bias, avoiding invalid or erroneous parameter adjustments. The calculated updated health parameter space... This will be passed back to the differentiable physical digital twin model module for calculating the predicted available power at the next moment, thus forming a continuously running, autonomously corrected closed-loop system.

[0071] See attached document Figure 5 , Figure 5 This is a schematic diagram of the combined output results according to an embodiment of the present invention. This section provides a detailed description of the final output of the system: the health status sensing power prediction result and the quantified interpretable state diagnosis result.

[0072] The joint output module, serving as the system's output interface, utilizes a structured health parameter space that is continuously updated by the twin autonomous correction and parameter update module. It provides two coupled output information: power prediction based on health status awareness and quantified interpretable state diagnosis.

[0073] In one application scenario, this system is used to predict the future power output of power generation equipment. This is necessary when it's required to predict power output for a specific future time period. When determining the output power, the joint output module first obtains the meteorological forecast data vector corresponding to that time period. Simultaneously, this module obtains the latest calculated structured health parameter space, which reflects the current actual health status of the device, from the twin autonomous correction and parameter update module. Subsequently, the joint output module will output the weather forecast data. and health parameter space Both inputs are fed into the differentiable physical digital twin model module, which then executes the forward computation process. The calculated future power prediction sequence is then obtained. This is output as the final power prediction result. This prediction process uses model parameters... It is corrected online with real-time data, so the output prediction results can take into account the actual decline in the current health status of the equipment.

[0074] In another application scenario, the system is used to perform condition diagnostics on power generation equipment. The combined output module continuously records and outputs a structured health parameter space. Time series data that evolves over time. For example... Figure 5 As shown, the output can be a set of curves, each curve representing a specific health parameter (e.g., equivalent thermal resistance). Capacitor attenuation coefficient The trajectory of the change of values ​​(etc.) over time.

[0075] These time series curves provide quantitative and interpretable diagnostic information about the internal state of the device. For example, a monotonically rising curve... The curve, in its physical sense, represents the continuous decline in the heat dissipation capacity of photovoltaic modules. The slope of this curve quantifies the rate of degradation in heat dissipation capacity. Similarly, a monotonically decreasing curve... The curve directly characterizes the decay process of the effective capacity of a certain capacitor element inside the inverter. This quantitative diagnostic result provides objective data for formulating equipment maintenance plans.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0077] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A power prediction and condition monitoring system for power generation equipment based on real-time data fusion, characterized in that, include: processor; The memory stores instructions that can be executed by the processor, which, when executed, cause the processor to perform: A differentiable physical digital twin model is constructed, wherein a structured health parameter space is defined within the differentiable physical digital twin model for quantitatively characterizing the health status of the power generation equipment; Calculate the internal state deviation signal, which is the deviation between the predicted available power output by the differentiable physical digital twin model and the actual output power of the power generation equipment; Based on the signal characteristics of the internal state deviation signal, credit allocation weights are assigned to the health parameters in the structured health parameter space. The structured health parameter space is updated online based on the credit allocation weights.

2. The power prediction and condition monitoring system for power generation equipment based on real-time data fusion according to claim 1, characterized in that, The structured health parameter space includes health parameters, which are selected from at least one of the following: equivalent thermal resistance, capacitance attenuation coefficient, inductance drift coefficient, or maximum power point tracking efficiency offset.

3. The power prediction and condition monitoring system for power generation equipment based on real-time data fusion according to claim 1, characterized in that, The process of calculating the internal state deviation signal includes: Real-time meteorological data and the current structured health parameter space are input into the differentiable physical digital twin model to obtain the predicted value of available power; The internal state deviation signal is obtained by subtracting the synchronously acquired actual output power from the predicted available power value.

4. The power prediction and condition monitoring system for power generation equipment based on real-time data fusion according to claim 1, characterized in that, The process of allocating credit allocation weights includes: The internal state deviation signal sequence within the time window is analyzed in the time domain, frequency domain, or time-frequency domain to extract the signal features; The signal features are input into a preset credit allocation network to output the normalized credit allocation weights, which are of the same dimension as the structured health parameter space.

5. The power prediction and condition monitoring system for power generation equipment based on real-time data fusion according to claim 1, characterized in that, The process of updating the structured health parameter space online includes: The function of the internal state deviation signal is used as the loss function; Calculate the gradient of the loss function with respect to the structured health parameter space; The gradient is weighted using the credit allocation weights to update the structured health parameter space.

6. The power prediction and condition monitoring system for power generation equipment based on real-time data fusion according to claim 1, characterized in that, When the instruction is executed, it causes the processor to perform: Future weather forecast data and the updated structured health parameter space are input into the differentiable physical digital twin model to generate power generation prediction results; and, The temporal variation trajectory of the structured health parameter space is output as the condition diagnosis result of the power generation equipment.

7. A method for power prediction and condition monitoring of power generation equipment based on real-time data fusion, executed by the power prediction and condition monitoring system for power generation equipment based on real-time data fusion as described in any one of claims 1-6, characterized in that, Includes the following steps: Step 1: Construct a differentiable physical digital twin model, wherein the differentiable physical digital twin model defines a structured health parameter space for quantitatively characterizing the health status of the power generation equipment; Step 2: Calculate the internal state deviation signal, which is the deviation between the predicted available power output by the differentiable physical digital twin model and the actual output power of the power generation equipment. Step 3: Based on the signal characteristics of the internal state deviation signal, assign credit allocation weights to the health parameters in the structured health parameter space; Step 4: Update the structured health parameter space online according to the credit allocation weight.

8. The method for power prediction and condition monitoring of power generation equipment based on real-time data fusion according to claim 7, characterized in that, Step three includes: The internal state deviation signal sequence within the time window is analyzed in the time domain, frequency domain, or time-frequency domain to extract the signal features; The signal features are input into a preset credit allocation network to output the normalized credit allocation weights, which are of the same dimension as the structured health parameter space.

9. The method for power prediction and condition monitoring of power generation equipment based on real-time data fusion according to claim 7, characterized in that, Step four includes: The function of the internal state deviation signal is used as the loss function; Calculate the gradient of the loss function with respect to the structured health parameter space; The gradient is weighted using the credit allocation weights to update the structured health parameter space.

10. The method for power prediction and condition monitoring of power generation equipment based on real-time data fusion according to claim 7, characterized in that, The method further includes: Step 5: Input future weather forecast data and the updated structured health parameter space into the differentiable physical digital twin model to generate power generation prediction results; and, Step 6: Output the time variation trajectory of the structured health parameter space as the status diagnosis result of the power generation equipment.

Images