Intelligent photovoltaic cell health degree evaluation system and method based on deep learning

Abstract

The invention discloses an intelligent photovoltaic cell health degree evaluation system and method based on deep learning. The method comprises the steps of S1, obtaining a standardized data set meeting modeling requirements; s2, inputting the standardized data into a continuous time dynamic health state modeling system constructed based on a Shenchang differential equation; s3, performing global search and adaptive dynamic adjustment on model parameters; s4, performing real-time prediction and evaluation on the current health state of the photovoltaic cell by using the optimized photovoltaic cell health state model, and outputting a health state evaluation result reflecting the dynamic health state of the photovoltaic cell; and S5, transmitting a health state evaluation result to an evaluation feedback module, and performing health state classification and grading on the photovoltaic cell by the evaluation feedback module according to a preset health state grading standard. According to the method, the action weights of the key influence factors of the photovoltaic cell at different time points can be quantified, and state-based feature weight self-adjustment is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of batteries, and relates to an intelligent photovoltaic battery health assessment system and method based on deep learning. Background Art

[0002] With the rapid development of renewable energy, as an important part of green energy, photovoltaic power generation is being widely applied in industrial parks, residential houses and large-scale photovoltaic power stations. However, due to the long-term exposure of photovoltaic battery components to complex environments, their performance will gradually degrade or even fail. Therefore, how to timely and accurately evaluate the health status of photovoltaic batteries to achieve preventive maintenance has become one of the key technical issues in current photovoltaic operation and maintenance management.

[0003] The existing photovoltaic battery health assessment methods can be mainly divided into two categories: one is the method based on statistical analysis of electrical performance indicators, which mainly collects data such as voltage, current and power offline for trend analysis to evaluate the performance degradation of components; the other is to establish physical models such as equivalent circuit models and thermodynamic models, and combine partial real-time data to infer the internal state of the battery, which has certain reference value in engineering practice, but there are still obvious deficiencies.

[0004] First of all, traditional statistical analysis methods rely on static data, insufficiently depict the time evolution process of photovoltaic batteries, and it is difficult to achieve dynamic tracking and prediction of degradation trends. Secondly, physical modeling methods usually require more prior knowledge, with complex model parameters, long debugging cycles and weak generalization ability, and it is difficult to adapt to the diversity and complexity of equipment operation states in large-scale scenarios. In addition, most traditional methods adopt a threshold discrimination mechanism set manually, lacking self-adaptability, and it is difficult to accurately identify early degradation signals, resulting in a lag in maintenance response.

[0005] In recent years, some studies have tried to introduce machine learning algorithms into photovoltaic health assessment, but the commonly used regression or classification models still mainly process discrete time series, fail to capture the dynamic evolution characteristics of photovoltaic battery states in the continuous time domain, and the model training process lacks an efficient global optimization mechanism, and is easy to fall into local optima, affecting the evaluation accuracy and stability.

[0006] To sum up, the existing photovoltaic battery health assessment technologies have limitations in modeling accuracy, dynamics, parameter optimization ability and early degradation detection ability, and it is difficult to meet the comprehensive assessment requirements of real-time, accuracy and forward-looking of current intelligent photovoltaic operation and maintenance systems. There is an urgent need for a new assessment method that can simultaneously model continuous dynamics, has an adaptive ability and can perform global optimization to improve the intelligent level of photovoltaic battery operation and maintenance management. Summary of the Invention

[0007] An object of the present invention is to provide an intelligent photovoltaic cell health assessment system and method based on deep learning. The present invention can quantify the action weights of key influencing factors of photovoltaic cells at different time points, and realizes self-adjustment of feature weights based on states.

[0008] An intelligent photovoltaic cell health assessment method based on deep learning according to an embodiment of the present invention includes the following steps: S1. Real-time collect the original data generated during the operation of the photovoltaic cell and perform preprocessing operations to obtain a standardized data set that meets the modeling requirements; S2. Input the standardized data into a continuous-time dynamic health state modeling system constructed based on neural ordinary differential equations. The continuous-time dynamic health state modeling system generates a photovoltaic cell health state model that describes the continuous evolution of the internal state of the photovoltaic cell over time through training; S3. Introduce a jellyfish swarm optimization algorithm during the training process of the photovoltaic cell health state model to perform global search and adaptive dynamic adjustment on the model parameters, achieve the optimal setting of the model parameters, and enable the photovoltaic cell health state model to obtain a global optimal solution in the multi-dimensional parameter space; S4. Use the optimized photovoltaic cell health state model to perform real-time prediction and evaluation on the current health state of the photovoltaic cell, and output a health state evaluation result reflecting the dynamic health status of the photovoltaic cell; S5. Transmit the health state evaluation result to the evaluation feedback module. The evaluation feedback module classifies and grades the health state of the photovoltaic cell according to the pre-set health state grading standard, and generates a corresponding evaluation report and preventive maintenance suggestions.

[0009] Optionally, the S1 includes the following steps: S11. Real-time collect the original data generated during the operation of the photovoltaic cell, and set a sampling time window to collect data to form an original data set : Among them, represents the th photovoltaic cell operation data collected at time , represents the voltage of the photovoltaic cell collected at time , represents the current of the photovoltaic cell collected at time , represents the temperature of the photovoltaic cell collected at time , represents the light irradiance at time , is the total number of sampling times; S12. Denoise the original dataset by removing random interference and high-frequency noise in the original data. For the missing values in the original dataset, use the adjacent data interpolation method to complete them, estimate them according to the average value of two adjacent valid observation points in the time series of the data, and then perform outlier removal. Set the effective interval range of each physical parameter, and remove the observation data that is not within the effective interval range. The effective interval range is set according to the normal physical state boundary of the photovoltaic cell operation. Finally, perform normalization processing, and convert the original physical parameter values into dimensionless standard values within the interval according to a unified linear mapping rule, and construct it into a standardized dataset ; S13. Each piece of observation data in the standardized dataset consists of five dimensions: sampling time, standardized photovoltaic cell voltage, standardized photovoltaic cell current, standardized photovoltaic cell temperature, and standardized light irradiance. Each piece of standardized data is uniformly represented as an ordered data tuple containing five indicators: sampling time, photovoltaic cell voltage, photovoltaic cell current, photovoltaic cell temperature, and light irradiance.

[0010] Optionally, the S2 includes the following steps: S21. Construct a photovoltaic cell health state model based on the standardized dataset, introduce a bivariate joint modeling mechanism in the neural ordinary differential equation structure. The bivariate joint modeling mechanism includes the photovoltaic cell health state variable and the photovoltaic cell degradation state variable . The photovoltaic cell health state variable and the photovoltaic cell degradation state variable are used to describe the functional state and deterioration trend state of the photovoltaic cell respectively, and establish a simultaneous differential equation model: where, represents the input feature vector of the standardized dataset at time , , are neural network functions with parameters being the neural network trainable parameter groups , respectively, is the dynamic attention weight coefficient vector; S23. Define the dynamic weight adjustment mechanism of the degradation state, and construct the time-adaptive dynamic attention weight coefficient vector ; S24. Based on the simultaneous differential equation model constructed in step S22, input the initial state , , and perform forward integration on the continuous time interval to obtain the joint health state prediction sequence , where, represents the estimated value of the photovoltaic cell function state predicted at time ; represents the estimated value of the degradation trend state predicted at time ; S25. The predicted joint health state prediction sequence reflects the current operating state and potential aging trend of the photovoltaic cell.

[0011] Optionally, the S231 includes the following steps: S231. During the dynamic modeling of the photovoltaic cell degradation state variable , construct a time-adaptive dynamic attention weight coefficient vector for the time-varying influence relationship of the input features on the degradation state: where, , , , respectively represent the relative influence weights of the photovoltaic cell voltage, current, temperature, and light irradiance on the photovoltaic cell degradation state variable at time ; S232. The time-adaptive dynamic attention weight coefficient vector is dynamically generated by the feature importance evaluation function : where, is a non-linear mapping function constructed based on the neural network parameters . The input is the photovoltaic cell degradation state variable at the current time and the normalized input feature vector . The output is an unnormalized feature importance vector, which is normalized by Softmax to generate the dynamic attention weight coefficient vector ; S233. The parameter set of the neural network mapping function is trained end-to-end through the backpropagation mechanism during the photovoltaic cell health state modeling process. The generated dynamic attention weight coefficient vector is updated in real time at each moment to realize the dynamic recognition and weight adjustment of the influence degree of different feature parameters of the photovoltaic cell on the degradation process; S234. The dynamic attention weight coefficient vector Applied to the differential equation of the degradation state in the simultaneous differential equation model, so that the evolution process of the degradation state variable of the photovoltaic cell is simultaneously affected by the standardized input feature vector and its dynamic attention weight coefficient vector at each moment, and a differential modeling mechanism reflecting the nonlinear degradation characteristics and feature-sensitive dynamics of the photovoltaic cell is constructed.

[0012] Optionally, S3 includes the following steps: S31. Introduce the jellyfish swarm optimization algorithm to optimize the parameters in the structure of the neural ordinary differential equation during the training process of the photovoltaic cell health state model, and set the optimization parameter set ; S32. Encode the optimization parameter set into the position vector of the jellyfish individual, where , is the number of individuals in the jellyfish population, and an initial jellyfish population is constructed; S33. Based on the structure of the neural ordinary differential equation corresponding to each jellyfish individual in the current population , calculate the health state prediction error of the neural ordinary differential equation structure on the training set, and construct a loss function: where, , respectively represent the predicted functional state and degradation state of the photovoltaic cell at the th jellyfish individual at time , , are the true label data; S34. According to the behavioral mechanism of the jellyfish swarm optimization algorithm, update the position of each jellyfish individual by using the active movement and tidal drift mechanisms: where, represents the parameter vector of the th jellyfish individual at the th iteration, is the optimal individual parameter vector with the smallest prediction error in the current population, is the perturbation vector simulating tidal flow, , are the adjustment coefficients controlling the intensity of active movement and drift, , are uniformly distributed random numbers; S35. In each round of iteration, based on the loss function Sort and select all jellyfish individuals, and retain the current global optimal parameter vector : S36. Repeat the jellyfish swarm optimization process until the maximum number of iterations or the convergence condition of the loss function is met, and finally output the optimal parameter set , and apply the optimal parameter set to each module of the neural ordinary differential equation structure to construct a photovoltaic cell health state model with the global optimal solution

[0013] Optionally, the S4 includes the following steps: S41. According to the comprehensive evaluation index of the health degree of the photovoltaic cell , fuse the joint health state prediction sequence to calculate the health state evaluation value: Among them, represents the health state evaluation value of the photovoltaic cell at time , , are the evaluation weight coefficients of the functional state and the degradation state respectively; S43. Extract the classification criteria of different health levels of the photovoltaic cell from the historical monitoring data, and define the health level division rules according to the health state evaluation value as follows: Normal state: If , it means that the photovoltaic cell is in good operating condition and has complete functions, and it is recommended to continue normal operation; Mild degradation state: If , it means that the photovoltaic cell has initial aging phenomenon, and it is recommended to conduct an operating cycle extension test and a non-invasive optical inspection; Moderate degradation state: If , it means that the photovoltaic cell has obvious performance degradation, and it is recommended to carry out distributed component tests in combination with on-site temperature data and electrical performance analysis to estimate its remaining life; Severe degradation state: If , it means that the photovoltaic cell may have structural or material failures, and it is recommended to immediately stop using this component and arrange for manual inspection and component replacement; S44. Correlate the health state evaluation value of the photovoltaic cell at each moment to the health level classification standard, and output the health state evaluation result including the current health level and the health trend index

[0014] Optionally, the S5 includes the following steps: S51. Based on the photovoltaic cell in the full-cycle sampling period The health status evaluation trend sequence within is used to extract the health status change trajectory and calculate the degradation speed index: Among them, represents the average decline rate of the photovoltaic cell health status evaluation value per unit time, which is used to judge whether there is a risk of rapid decay; S52. Generate a health status evaluation report according to the classification result and the degradation speed index ; S53. Generate preventive maintenance suggestions based on the combination result of the health level and the degradation speed index.

[0015] Optionally, the health status evaluation report includes the current health level, the health status trend curve of the most recent complete cycle, and the key feature influence weights.

[0016] Optionally, the preventive maintenance suggestions include the recommended monitoring cycle, whether to enable manual inspection, and whether to perform replacement warning marking.

[0017] An intelligent photovoltaic cell health evaluation system based on deep learning is used to execute an intelligent photovoltaic cell health evaluation method based on deep learning, and includes the following modules: The data acquisition and preprocessing module is used to collect the original operation data of the photovoltaic cell and complete preprocessing to generate a standardized data set; The health status modeling module constructs a photovoltaic cell health status model of the photovoltaic cell health status and degradation status based on neural ordinary differential equations, which is used to characterize the continuous-time evolution of the photovoltaic cell operation status; The parameter optimization module globally optimizes the parameters in the photovoltaic cell health status model based on the jellyfish swarm optimization algorithm and outputs an optimal parameter set; The state prediction and evaluation module uses the optimized photovoltaic cell health status model to predict the current state of the photovoltaic cell and generate a health status evaluation result; The state feedback and decision-making module completes the classification and grading of the photovoltaic cell health status according to the set health status grading standard and degradation trend, and generates an evaluation report and preventive maintenance suggestions.

[0018] The beneficial effects of the present invention are: The present invention constructs a continuous-time modeling framework based on neural ordinary differential equations. By introducing a bivariate joint modeling mechanism of health state variables and degradation state variables, a set of simultaneous differential equations is established to accurately model the evolution process of the functional state and degradation trend state of photovoltaic cells. Neural ordinary differential equations can model continuously changing states over time at any resolution, avoiding the limitation of fixed time steps and improving the adaptability of the model to non-uniformly sampled data. It is applicable to photovoltaic cell component systems with complex operating environments and slow state changes.

[0019] The present invention introduces the jellyfish swarm optimization algorithm into the training process of neural ordinary differential equations to globally search and optimize a multi-dimensional parameter set including health state function parameters, degradation state function parameters, and attention mechanism network parameters. By simulating the active movement and tidal drift behavior of jellyfish, efficient search is achieved in the complex loss function space, avoiding the problem that the gradient descent method is prone to falling into local optima, and significantly improving the stability of training convergence and the consistency of model performance.

[0020] The present invention proposes a time-adaptive dynamic attention weight mechanism. By constructing a neural network mapping function with degradation state variables and input feature vectors as inputs, attention coefficients are dynamically generated at each moment, which can quantify the action weights of key influencing factors of photovoltaic cells at different time points, realizing state-based self-adjustment of feature weights, and enabling the model to have stronger discrimination and generalization abilities when facing multi-source complex perturbation factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of an intelligent photovoltaic cell health assessment system and method based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0022] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0023] Refer to Figure 1 , an intelligent photovoltaic cell health assessment method based on deep learning, including the following steps: S1. Real-time collect the original data generated during the operation of the photovoltaic cell and perform preprocessing operations to obtain a standardized data set that meets the modeling requirements; S2. Input the standardized data into the continuous-time dynamic health state modeling system constructed based on neural ordinary differential equations. The continuous-time dynamic health state modeling system generates a photovoltaic cell health state model through training, which describes the continuous evolution of the internal state of the photovoltaic cell over time. S3. Introduce the jellyfish swarm optimization algorithm during the training process of the photovoltaic cell health state model to perform global search and adaptive dynamic adjustment on the model parameters, achieve the optimal setting of the model parameters, and enable the photovoltaic cell health state model to obtain the global optimal solution in the multi-dimensional parameter space. S4. Use the optimized photovoltaic cell health state model to perform real-time prediction and evaluation on the current health state of the photovoltaic cell, and output the health state evaluation result reflecting the dynamic health condition of the photovoltaic cell. S5. Transmit the health state evaluation result to the evaluation feedback module. The evaluation feedback module classifies and grades the health state of the photovoltaic cell according to the pre-set health state grading standard, and generates the corresponding evaluation report and preventive maintenance suggestions.

[0024] In this embodiment, S1 includes the following steps: S11. Real-time collect the original data generated during the operation of the photovoltaic cell, and set the sampling time window to collect data to form the original data set : Among them, represents the th photovoltaic cell operation data collected at time represents the voltage of the photovoltaic cell collected at time represents the current of the photovoltaic cell collected at time represents the temperature of the photovoltaic cell collected at time represents the irradiance of the light collected at time represents the time at which the temperature of the photovoltaic cell is collected, represents the time at which the irradiance of the light is collected, is the total number of sampling times; S12. Perform data denoising processing on the original data set to remove the random interference and high-frequency noise in the original data, and use the adjacent data interpolation method to fill in the missing values in the original data set. Estimate it according to the average value of two adjacent valid observation points of the data in the time series, and then perform outlier removal processing. Set the effective interval range of each physical parameter, and remove the observation data that is not within the effective interval range. The effective interval range is set according to the normal physical state boundary of the operation of the photovoltaic cell. Finally, perform normalization processing to convert the original physical parameter values into the interval according to the unified linear mapping rule The dimensionless standard values within are constructed into a standardized data set ; S13. Each piece of observed data in the standardized data set consists of five dimensions: sampling time, standardized photovoltaic cell voltage, standardized photovoltaic cell current, standardized photovoltaic cell temperature, and standardized light irradiance. Each piece of standardized data is uniformly represented as an ordered data tuple containing five indicators: sampling time, photovoltaic cell voltage, photovoltaic cell current, photovoltaic cell temperature, and light irradiance.

[0025] In this embodiment, S2 includes the following steps: S21. Based on the standardized data set, construct a photovoltaic cell health state model, and introduce a bivariate joint modeling mechanism into the neural ordinary differential equation structure. The bivariate joint modeling mechanism includes a photovoltaic cell health state variable and a photovoltaic cell degradation state variable . The photovoltaic cell health state variable and the photovoltaic cell degradation state variable are respectively used to characterize the functional state and deterioration trend state of the photovoltaic cell, and establish a simultaneous differential equation model: Among them, represents the input feature vector of the standardized data set at time , , are neural network functions with parameters being the neural network trainable parameter groups , respectively, is the dynamic attention weight coefficient vector; , The construction specifically includes the following steps: The neural network function is used to construct the differential change relationship of the photovoltaic cell health state variable . The input is the health state , degradation state , standardized input feature vector and the current time , and the output is the change rate of the health state variable: The neural network function is composed of a multi-layer feedforward neural network, and the parameter set is , where , are respectively the weight matrix and bias vector of the th layer, is the number of network layers; The said neural network function Used to construct photovoltaic cell degradation state variables The differential change relationship of , standardize the input feature vector , dynamic attention weight coefficient vector and the current moment , the output is the rate of change of the degraded state variable: Neural Network Functions It is also composed of a multi-layer feedforward neural network, and the parameter set is ,in , Respectively The weight matrix and bias vector of the layer, the number of network layers and the structure can be adaptively set according to the training data set.

[0026] Neural Network Functions , In both cases, nonlinear activation functions are used for feature transformation, and activation functions are used after each layer of neurons. Processing linear output, the activation function is: in, is the hyperbolic tangent function, To modify the linear unit function, the selected activation function is used to enhance the nonlinear modeling capability.

[0027] Two neural network functions , Collaborative training is performed through the back-propagation mechanism, and its outputs are used as differential control terms of the healthy state variables and the degradation state variables of the photovoltaic cells in the neural ordinary differential equation model.

[0028] S23. Define the dynamic weight adjustment mechanism of the degradation state and construct the time-adaptive dynamic attention weight coefficient vector ; S24. Based on the simultaneous differential equation model constructed in step S22, input the initial state , , for continuous time intervals Perform forward integral solution to obtain the joint health status prediction sequence ,in, Indicates at time The estimated value of the photovoltaic cell functional state obtained by prediction, Indicates at time The estimated value of the degradation trend state obtained by prediction; S25. The predicted joint health status prediction sequence Reflect the current operating state and potential aging trend of the photovoltaic cell.

[0029] In this embodiment, S231 includes the following steps: S231. During the dynamic modeling process of the degradation state variable of the photovoltaic cell, construct a time-adaptive dynamic attention weight coefficient vector for the time-varying influence relationship of the input features on the degradation state: Among them, , , , respectively represent the relative influence weights of the voltage, current, temperature, and light irradiance of the photovoltaic cell on the degradation state variable of the photovoltaic cell at time ; S232. The time-adaptive dynamic attention weight coefficient vector is dynamically generated by the feature importance evaluation function : Among them, is a non-linear mapping function constructed based on the neural network parameters . The input is the degradation state variable of the photovoltaic cell at the current time and the normalized input feature vector . The output is an unnormalized feature importance vector, which is normalized by Softmax to generate the dynamic attention weight coefficient vector ; S233. The parameter set of the neural network mapping function is trained end-to-end through the backpropagation mechanism during the health state modeling process of the photovoltaic cell. The generated dynamic attention weight coefficient vector is updated in real time at each moment to realize the dynamic recognition and weight adjustment of the influence degree of different characteristic parameters of the photovoltaic cell on the degradation process; S234. Apply the dynamic attention weight coefficient vector to the degradation state differential equation in the simultaneous differential equation model, so that the evolution process of the degradation state variable of the photovoltaic cell is jointly affected by the normalized input feature vector and its dynamic attention weight coefficient vector at each moment, and construct a differential modeling mechanism that reflects the non-linear degradation characteristics and feature-sensitive dynamics of the photovoltaic cell.

[0030] In this embodiment, S3 includes the following steps: S31. During the training process of the photovoltaic cell health state model, introduce the jellyfish swarm optimization algorithm to optimize the parameters in the neural ordinary differential equation structure, and set the optimization parameter set ; S32. Encode the optimization parameter set as the position vector of jellyfish individuals , where , is the number of individuals in the jellyfish population, and construct the initial jellyfish population ; S33. Based on the neural ordinary differential equation structure corresponding to each jellyfish individual in the current population , calculate the health state prediction error of the neural ordinary differential equation structure on the training set, and construct the loss function: where, , respectively represent the predicted photovoltaic cell functional state and degradation state at time by the th jellyfish individual, , are the true label data; S34. According to the behavioral mechanism of the jellyfish swarm optimization algorithm, update the position of each jellyfish individual using the active movement and tidal drift mechanisms: where, represents the parameter vector of the th jellyfish individual at the th iteration, is the optimal individual parameter vector with the minimum prediction error in the current population, is the perturbation vector simulating tidal flow, , are the adjustment coefficients controlling the intensity of active movement and drift, , are random numbers uniformly distributed; S35. In each iteration, sort and select all jellyfish individuals based on the loss function , and retain the current global optimal parameter vector : S36. Repeat the jellyfish swarm optimization process until the maximum number of iterations or the loss function convergence condition is met, and finally output the optimal parameter set , apply the optimal parameter set to each module of the neural ordinary differential equation structure to construct a globally optimal photovoltaic cell health state model.

[0031] In this embodiment, S4 includes the following steps: S41. According to the comprehensive evaluation index of photovoltaic cell health , fuse the joint health state prediction sequence to calculate the health state evaluation value: Among them, represents the health state evaluation value of the photovoltaic cell at time , , are the evaluation weight coefficients of the functional state and the degradation state respectively; S43. Extract the classification criteria of different health levels of the photovoltaic cell from the historical monitoring data, and define the health level division rules according to the health state evaluation value as follows: Normal state: If , it means that the photovoltaic cell is in good operating condition and has complete functions, and it is recommended to continue normal operation; Mild degradation state: If , it means that the photovoltaic cell has initial aging phenomenon, and it is recommended to conduct an operating cycle extension test and a non-invasive optical inspection; Moderate degradation state: If , it means that the photovoltaic cell has obvious performance decay, and it is recommended to carry out distributed component tests in combination with on-site temperature data and electrical performance analysis to estimate its remaining life; Severe degradation state: If , it means that the photovoltaic cell may have structural or material failures, and it is recommended to immediately stop using this component and arrange for manual inspection and component replacement; S44. Correlate the health state evaluation value of the photovoltaic cell at each moment to the health level classification standard, and output the health state evaluation result including the current health level and the health trend index.

[0032] In this embodiment, S5 includes the following steps: S51. Based on the health state evaluation trend sequence of the photovoltaic cell during the full-cycle sampling period , extract the health state change trajectory and calculate the degradation speed index: Among them, represents the average decline rate of the health state evaluation value of the photovoltaic cell per unit time, which is used to judge whether there is a risk of rapid decay; S52. According to the classification result and the degradation speed index Generate a health status assessment report; S53. Generate preventive maintenance suggestions based on the combination result of the health level and the degradation speed index.

[0033] In this embodiment, the health status assessment report includes the current health level, the health status trend curve of the most recent complete cycle, and the key feature impact weights.

[0034] In this embodiment, the preventive maintenance suggestions include the recommended monitoring cycle, whether to enable manual inspection, and whether to perform replacement warning annotation.

[0035] An intelligent photovoltaic cell health assessment system based on deep learning, which is used to execute an intelligent photovoltaic cell health assessment method based on deep learning, includes the following modules: The data acquisition and preprocessing module is used to collect the original operation data of the photovoltaic cell and complete preprocessing to generate a standardized data set; The health status modeling module constructs a photovoltaic cell health status model of the health status and degradation status of the photovoltaic cell based on the neural ordinary differential equation, which is used to characterize the continuous-time evolution of the operating state of the photovoltaic cell; The parameter optimization module globally optimizes the parameters in the photovoltaic cell health status model based on the jellyfish swarm optimization algorithm and outputs an optimal parameter set; The state prediction and evaluation module uses the optimized photovoltaic cell health status model to predict the current state of the photovoltaic cell and generates a health status evaluation result; The state feedback and decision-making module classifies and grades the health status of the photovoltaic cell according to the set health status classification standard and degradation trend, and generates an evaluation report and preventive maintenance suggestions. Example

[0036] During the routine data collection process of the operation monitoring system of a large-scale ground photovoltaic power station in City A, it was found that the output power of the 14th group of components in the 7th column of the east area (number: D7G14) showed slight fluctuations compared with the other components in this string, and the power fluctuation frequency increased abnormally (more than 7 times with an offset of more than 3% within the past 2 hours). This offset did not trigger the inverter protection mechanism but entered the gray area abnormal window of operation and maintenance monitoring.

[0037] The intelligent health assessment system starts the model of the present invention to analyze the data of component D7G14 in the most recent 24 hours. The system retrieves the continuous data of this component in real time, and a total of 288 data points are collected. The data format is (timestamp, voltage V, current A, temperature °C, irradiance W / m²). The following is part of the original record: Table 1 Partial original record of component D7G14 Timestamp Voltage (V) Current (A) Temperature (°C) Irradiance (W / m²) 2024-08-13 08:45:00 38.72 4.96 47.6 845 2024-08-13 12:10:00 38.21 4.52 53.1 916 2024-08-13 15:05:00 37.35 4.01 55.2 888 2024-08-13 18:35:00 38.02 4.37 49.8 691 2024-08-14 05:50:00 36.90 3.94 46.4 402 The data is sent into the continuous-time modeling module constructed by neural ordinary differential equations. The system automatically loads the initial values of the historical state variables of component number D7G14, h0 = 0.81 and z0 = 0.23, and constructs a differential prediction path based on the current input feature vector. It is predicted that the health state score of this component will rapidly drop to 0.39 in the next 3 hours, and the degradation variable shows exponential growth.

[0038] Meanwhile, the dynamic attention mechanism analyzes the influencing factors of the component state. The attention weights of the input features at the current moment calculated by the system are: αV(t)=0.16 (voltage); αI(t)=0.11 (current); αT(t)=0.48 (temperature); αG(t)=0.25 (irradiance). The system determines that the abnormal temperature has a significant impact on the degradation trend.

[0039] The system refers to the constructed health state level threshold standard, determines that the current state has changed from "mild degradation" to "moderate degradation", and automatically generates a health state assessment report, the content of which is as follows: The current health score S(t)=0.41; the degradation rate Rdeg in the last 12 hours = 0.026 / h; recommended maintenance level: second level; recommended measures: component thermal imaging detection + electrical performance test.

[0040] The system pushes the above report and suggestions to the PDA device of front-line inspection personnel through the internal operation and maintenance platform in the main control room.

[0041] The front-line inspection personnel view the system evaluation suggestions and go to the site to start the inspection. After arriving at the site of component D7G14, an infrared thermal imager is used to find non-uniform hot spot patches in the lower left corner area of this component, and the temperature is 12.7℃ higher than the surrounding area. Subsequently, an open-circuit voltage test is carried out: the actual voltage is 37.22V, about 1.3V lower than the average voltage of 38.56V of adjacent components in the same group, and the current is 4.08A. It is initially judged that there are problems such as local cracks or welding delamination in this component.

[0042] The front-line inspection personnel complete the on-site test and check "adoption of suspension of use suggestion" in the system. The system automatically marks this component as "high-risk to be replaced" and records the fault identification time as "2024081406:29".

[0043] To verify the advantages of the method of the present invention in the actual scenario, the project team synchronously deployed a comparative test of the traditional LSTM model and the empirical rule method on the same day, selected the same input data, and the test results are as follows: Table 2 Test results of the model of the present invention and the comparative model Model type Fault identification time Initial health score Health degradation warning time Alarm lead time The method of the present invention 06:29 0.81 06:29 - LSTM prediction method 08:10 0.79 08:06 Lag by 97 minutes Based on empirical threshold method No alarm No score None Fault not identified In addition, a comparative analysis was conducted on 25 degraded component events identified by the model of the present invention and the comparative model within one week of continuous deployment: Table 3 Comparative Analysis of Evaluation Metrics between the Model of the Present Invention and the Comparative Model Evaluation index The method of the present invention LSTM model Empirical method Fault identification accuracy 96.2% 83.1% 68.7% Early warning time (average) 2.3 hours 0.6 hours 0 hours Health score error (mean square) 0.014 0.029 Unable to evaluate Response time (model calculation time consumption) 5.4 seconds 19.8 seconds 0.3 seconds The chief operation and maintenance engineer of the power station pointed out at the monthly summary meeting that since the deployment of the intelligent health assessment system, a total of 38 early-degraded components have been identified, reducing unplanned outages by 11 times, avoiding the series of low-efficiency problems caused by hidden component failures, and the system has triggered a total of 65 effective maintenance suggestions, with an adoption rate of 94%.

[0044] The present invention constructs a continuous-time modeling framework based on neural ordinary differential equations. By introducing a bivariate joint modeling mechanism of health state variables and degradation state variables, a set of simultaneous differential equations is established to accurately model the evolution process of the functional state and degradation trend state of photovoltaic cells. Neural ordinary differential equations can model the continuously changing state over time at any resolution, avoiding the limitation of fixed time steps and improving the adaptability of the model to non-uniformly sampled data. It is applicable to photovoltaic cell component systems with complex operating environments and slow state changes.

[0045] The present invention introduces the jellyfish swarm optimization algorithm into the training process of neural ordinary differential equations to globally search and optimize a multi-dimensional parameter set including health state function parameters, degradation state function parameters, and attention mechanism network parameters. By simulating the active movement and tidal drift behavior of jellyfish, an efficient search is achieved in the complex loss function space, avoiding the problem that the gradient descent method is prone to falling into local optima, and significantly improving the stability of training convergence and the consistency of model performance.

[0046] The present invention proposes a time-adaptive dynamic attention weight mechanism. By constructing a neural network mapping function with degradation state variables and input feature vectors as inputs, attention coefficients are dynamically generated at each moment, which can quantify the weight of the key influencing factors of photovoltaic cells at different time points, realizing state-based self-regulation of feature weights, and enabling the model to have stronger discrimination and generalization abilities in the face of multi-source complex disturbance factors.

[0047] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A smart photovoltaic battery health assessment method based on deep learning, characterized in that: The steps include: S1. Real-time collection of raw data generated by photovoltaic cells during operation and preprocessing operations to obtain a standardized data set that meets modeling requirements; S2. Input the standardized data into a continuous-time dynamic health state modeling system based on a neural ordinary differential equation, and the continuous-time dynamic health state modeling system generates a photovoltaic cell health state model that describes the continuous evolution of the internal state of the photovoltaic cell over time through training; S3. In the training process of the photovoltaic cell health status model, the jellyfish swarm optimization algorithm is introduced to perform global search and adaptive dynamic adjustment of the model parameters to achieve the optimal setting of the model parameters, and the photovoltaic cell health status model obtains the global optimal solution in the multi-dimensional parameter space; S4. Use the optimized photovoltaic cell health status model to predict and evaluate the current health status of the photovoltaic cell in real time, and output a health status evaluation result reflecting the dynamic health status of the photovoltaic cell; S5. The health status assessment result is transmitted to the assessment feedback module. The assessment feedback module classifies and grades the health status of the photovoltaic cells according to a preset health status classification standard, and generates a corresponding assessment report and preventive maintenance recommendations.

2. According to claim 1, a method for evaluating the health of an intelligent photovoltaic battery based on deep learning is characterized in that: The S1 comprises the following steps: S11. Real-time collection of raw data generated by photovoltaic cells during operation, setting the sampling time window to collect data to form the original data set : in, Indicates at time The collected PV cell operation data, Indicates at time The collected photovoltaic cell voltage, Indicates at time The collected photovoltaic cell current, Indicates at time The collected photovoltaic cell temperature, Indicates at time The light irradiance, is the total number of sampling moments; S12. Original dataset Data denoising is performed to remove random interference and high-frequency noise in the original data, and adjacent data interpolation is used to fill in the missing values ​​in the original data set. The data is estimated based on the average value of two adjacent valid observation points in the time series, and then outliers are eliminated. The effective interval range of each physical parameter is set, and the observation data that is not within the effective interval range is eliminated. The effective interval range is set according to the normal physical state boundary of the photovoltaic cell operation. Finally, normalization is performed to convert the original physical parameter values ​​into interval values ​​according to a unified linear mapping rule. The dimensionless standard value within is constructed as a standardized data set ; S13. Each observation data in the standardized data set consists of five dimensions: sampling time, standardized photovoltaic cell voltage, standardized photovoltaic cell current, standardized photovoltaic cell temperature and standardized light irradiance. Each standardized data is uniformly represented as an ordered data tuple containing five indicators: sampling time, photovoltaic cell voltage, photovoltaic cell current, photovoltaic cell temperature and light irradiance.

3. According to claim 2, a method for evaluating the health of an intelligent photovoltaic battery based on deep learning is characterized in that: The S2 comprises the following steps: S21. A photovoltaic cell health status model is constructed based on a standardized data set. A two-variable joint modeling mechanism is introduced into the structure of the neural ordinary differential equation. The two-variable joint modeling mechanism includes the photovoltaic cell health status variable and photovoltaic cell degradation state variables , the photovoltaic cell health state variable and the photovoltaic cell degradation state variable are used to characterize the functional state and degradation trend state of the photovoltaic cell, respectively, and establish a simultaneous differential equation model: in, Indicates time The input feature vector of the standardized dataset, , The parameters are the neural network trainable parameter groups , The neural network function, is the dynamic attention weight coefficient vector; S23. Define the dynamic weight adjustment mechanism of the degradation state and construct the time-adaptive dynamic attention weight coefficient vector ; S24. Based on the simultaneous differential equation model constructed in step S22, input the initial state , , for continuous time intervals Perform forward integral solution to obtain the joint health status prediction sequence ,in, Indicates at time The estimated value of the photovoltaic cell functional state obtained by prediction, Indicates at time The estimated value of the degradation trend state obtained by prediction; S25. The predicted joint health status prediction sequence Reflects the current operating status and potential aging trend of photovoltaic cells.

4. According to claim 3, a method for evaluating the health of an intelligent photovoltaic battery based on deep learning is characterized in that: The S231 includes the following steps: S231. Photovoltaic cell degradation state variables In the dynamic modeling process, the time-varying influence of input features on the degradation state is considered, and a time-adaptive dynamic attention weight coefficient vector is constructed. : in, , , , Respectively, in time The effects of photovoltaic cell voltage, current, temperature and light irradiance on photovoltaic cell degradation state variables The relative weight of influence; S232. Time-adaptive dynamic attention weight coefficient vector By feature importance evaluation function Dynamically generated: in, Based on the neural network parameters The constructed nonlinear mapping function takes the current time as input The degradation state variables of photovoltaic cells With the normalized input feature vector , the output is the unnormalized feature importance vector, which is normalized by Softmax to generate the dynamic attention weight coefficient vector ; S233. Neural network mapping function The parameter set In the process of photovoltaic cell health state modeling, end-to-end training is performed through the back-propagation mechanism, and the dynamic attention weight coefficient vector generated At every moment The internal parameters are updated in real time to achieve dynamic identification and weight adjustment of the influence of different characteristic parameters of photovoltaic cells on the degradation process; S234. Dynamic attention weight coefficient vector Applied to the degenerate state differential equation in the simultaneous differential equation model, the degradation state variable of the photovoltaic cell is The evolution of is simultaneously affected by the normalized input feature vector and its dynamic attention weight coefficient vector A differential modeling mechanism that reflects the nonlinear degradation characteristics and characteristic sensitive dynamics of photovoltaic cells is constructed based on the combined influence of the factors affecting the performance of photovoltaic cells.

5. According to claim 3, a method for evaluating the health of an intelligent photovoltaic battery based on deep learning is characterized in that: The S3 comprises the following steps: S31. In the training process of the photovoltaic cell health status model, the jellyfish swarm optimization algorithm is introduced to optimize the parameters in the Neural Ordinary Differential Equation structure and set the optimization parameter set. ; S32. Optimize the parameter set Encoded as the position vector of the jellyfish individual ,in , is the number of individuals in the jellyfish population, constructing the initial jellyfish population ; S33. Based on the current population Each jellyfish The corresponding neural ordinary differential equation structure is used to calculate the health status prediction error of the neural ordinary differential equation structure on the training set and construct the loss function: in, , Respectively represent A jellyfish individual at a time The predicted functional and degradation states of photovoltaic cells, , is the real label data; S34. Based on the behavior mechanism of the jellyfish swarm optimization algorithm, the position of each jellyfish individual is updated using active motion and tidal drift mechanisms ; S35. In each iteration, based on the loss function Sort and select all jellyfish individuals and retain the current global optimal parameter vector : S36. Repeat the jellyfish swarm optimization process until the maximum number of iterations or the loss function convergence condition is met, and finally output the optimal parameter set , the optimal parameter set is applied to each module of the neural ordinary differential equation structure to construct a photovoltaic cell health status model with a global optimal solution.

6. The method for evaluating the health of a smart photovoltaic cell based on deep learning according to claim 5, characterized in that: The S4 comprises the following steps: S41. Based on comprehensive evaluation indicators of photovoltaic cell health , and the health status prediction sequence is integrated to calculate the health status assessment value: in, Indicates that the photovoltaic cell is at the time The health status assessment value, , are the evaluation weight coefficients of functional status and degradation status, respectively; S43. Extract the classification standards of different health levels of photovoltaic cells from historical monitoring data, and use the health status assessment value The health level classification rules are defined as follows: Normal state: If , indicating that the photovoltaic cell is in good operating condition and fully functional, and it is recommended to continue normal operation; Mild degenerative state: If , indicating that the photovoltaic cells have initial aging phenomenon, and it is recommended to conduct operating cycle extension test and non-invasive optical inspection; Moderately degraded state: If , indicating that the photovoltaic cells have experienced significant performance degradation. It is recommended to conduct distributed component testing in combination with on-site temperature data and electrical performance analysis to estimate their remaining life; Severe degeneration: If , indicating that the photovoltaic cell may have structural or material failures. It is recommended to immediately stop using the component and arrange manual inspection and component replacement; S44. The health status evaluation value of the photovoltaic cell at each moment Corresponding to the health level classification standard, the output includes the health status assessment results of the current health level and health trend indicators.

7. The method for evaluating the health of a smart photovoltaic battery based on deep learning according to claim 6, characterized in that: The S5 comprises the following steps: S51. Based on the photovoltaic cell sampling period in the whole cycle The health status assessment trend sequence in the system is extracted to extract the health status change trajectory and calculate the degradation speed index: in, Indicates the average rate of decrease of the health status assessment value of the photovoltaic cell per unit time, which is used to determine whether there is a risk of rapid attenuation; S52. Based on the classification results and degradation rate indicators Generate health status assessment reports; S53. Generate preventive maintenance recommendations based on the combined results of the health level and degradation rate indicators.

8. The method for evaluating the health of a smart photovoltaic battery based on deep learning according to claim 7, characterized in that: The health status assessment report includes the current health level, the health status trend curve of the most recent complete cycle, and the impact weights of key features.

9. The method for evaluating the health of a smart photovoltaic cell based on deep learning according to claim 7, characterized in that: The preventive maintenance recommendations include recommended monitoring cycles, whether to enable manual inspections, and whether to perform replacement warning markings.

10. A deep learning-based intelligent photovoltaic battery health assessment system, used to execute a deep learning-based intelligent photovoltaic battery health assessment method according to any one of claims 1 to 9, characterized in that: Includes the following modules: The data acquisition and preprocessing module is used to collect the raw operating data of the photovoltaic cells and complete preprocessing to generate a standardized data set; The health state modeling module builds a photovoltaic cell health state model based on the Neural Ordinary Differential Equation to characterize the continuous time evolution of the photovoltaic cell operating state; The parameter optimization module performs global optimization on the parameters in the photovoltaic cell health status model based on the jellyfish swarm optimization algorithm and outputs the optimal parameter set; The state prediction and evaluation module uses the optimized photovoltaic cell health state model to predict the current state of the photovoltaic cell and generate a health state evaluation result; The status feedback and decision-making module classifies and grades the health status of photovoltaic cells according to the set health status classification standards and degradation trends, and generates evaluation reports and preventive maintenance recommendations.

Images