Fixed type offshore photovoltaic environment load intelligent calculation method considering shielding effect

Abstract

The invention relates to the field of offshore photovoltaic technology, and discloses a fixed offshore photovoltaic environment load intelligent calculation method considering a shielding effect. Comprising the following steps: S1, constructing a multi-source mapping sample set between ocean wind wave flow environment parameters and load coefficients corrected by a shielding effect; s2, constructing a load coefficient intelligent prediction model, and capturing a complex nonlinear mapping relation on the premise of guaranteeing the rationality of mechanics; s3, the load coefficient output by the neural network is substituted into a load physical formula, and wind, wave and flow load values considering the shielding effect are rapidly and accurately calculated; and S4, inverting a real load coefficient based on actual engineering monitoring data, and performing continuous fine adjustment and optimization on the prediction model to improve the adaptability and prediction precision of the prediction model under different site conditions. According to the scheme, the key load coefficient which is difficult to accurately obtain in a traditional method is determined through deep learning, the mechanical rationality of the result is ensured by relying on a physical model, and reliable support is provided for safe and economic design of a fixed offshore photovoltaic structure.

Description

Technical Field

[0001] This invention relates to the field of marine photovoltaic technology, and specifically to an intelligent calculation method for the environmental load of fixed marine photovoltaic systems that takes into account shading effects. Background Technology

[0002] As an important direction for promoting energy transition, the large-scale commercialization of offshore photovoltaics highly depends on improving the economics of the projects. Accurate assessment and calculation of marine environmental loads (including wind, wave, and current loads) are key to optimizing structural design, reducing material usage and costs, and are crucial for controlling the total investment of the project.

[0003] Currently, the assessment of marine environmental loads for offshore photovoltaic (PV) systems primarily relies on traditional code methods and physical simulations. The wind, wave, and current load calculation methods provided in current structural design codes are typically based on ideal conditions for isolated structural members, failing to adequately consider the complex shading effects prevalent in actual PV arrays. Specifically, in wind load calculations, the code methods do not consider the airflow shading and interference effects between rows of PV panels, leading to overly conservative results when calculating wind loads on panels within the array, resulting in unnecessary structural design redundancy and material waste. In wave and current load calculations, commonly used methods such as the Morison formula fail to accurately account for the mutual shading and interference effects between pile groups caused by waves and ocean currents when calculating pile group loads. This also leads to overestimation of the loads, resulting in conservative pile foundation designs.

[0004] In addition, key parameters in load calculation, such as wind pressure distribution coefficient, wind load shape coefficient, drag force coefficient in wave load, and inertial force coefficient, are highly scene-dependent. These coefficients are difficult to give accurately through theoretical formulas and are usually determined or inverted indirectly by wind tunnel tests, water tank tests, or high-fidelity computational fluid dynamics numerical simulations. These methods are costly, time-consuming, and cannot be quickly applied to the comparison of numerous design schemes. Summary of the Invention

[0005] The present invention aims to provide an intelligent calculation method for the environmental load of fixed marine photovoltaic systems that takes into account the shading effect, in order to solve the problem that the current calculation of wind, wave and current loads does not take into account the complex shading effect of the structure, resulting in an overestimation of the load.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent calculation method for the environmental load of fixed marine photovoltaic systems considering shading effects, comprising: S1: Through high-fidelity numerical simulation, physical model tests and field monitoring, a multi-source mapping sample set between ocean wind, wave and current environmental parameters and load coefficients corrected for shielding effects is constructed; S2: Introduce a multi-layer feedforward neural network with regularization constraints, and combine small sample data augmentation and physical guidance strategies to construct an intelligent prediction model of load coefficient based on the multi-layer feedforward neural network. The model includes: a fast prediction model of wind load shape coefficient considering panel shading effect and a fast prediction model of drag force coefficient and inertial force coefficient considering pile group shading effect. S3: Based on the extreme marine wind, wave and current environmental parameters of the target sea area, predict the load factor corrected by the shielding effect, and then calculate the wind load value, wave load value and current load value considering the shielding effect based on the predicted load factor, so as to achieve the complementary advantages of intelligent prediction and physical mechanism. S4: Obtain actual marine wind, wave and current environmental parameters and structural load data of the target pile-based offshore photovoltaic structure, and adaptively correct the intelligent calculation model of environmental load; The load factor calculation model employs a generative adversarial network (GAN) based on data homeomorphism. By learning from limited historical photovoltaic data, it generates augmented data that is consistent with the real data in both statistical characteristics and physical meaning. The GAN training process is essentially a process of solving for the minimum and maximum values ​​of a bivariate function, specifically using the following formula: In the formula and Let G represent the probabilities of real data and generated data. To ensure that the generated data not only has similar statistical characteristics but also conforms to physical laws, a physical constraint term is added to the loss function of the generator G. In the formula, P is the physical equation. To weigh the parameters.

[0007] Currently, determining the wind load shape coefficient is extremely difficult in the wind load calculation of fixed-foundation offshore photovoltaic (PV) systems. Offshore PV arrays are large-scale and complex in layout, and wind can generate complex mutual interference and shading effects. Traditional standard shape coefficients cannot accurately reflect the true wind pressure distribution in the actual three-dimensional space of the offshore PV structure. In the calculation of wave and current loads on fixed-foundation offshore PV piles, the drag force coefficient (involved by both types of loads) and the inertial force coefficient (involved only by wave loads) are semi-empirical parameters that cannot be directly derived from theory. Furthermore, due to wake and vortex interference, the piles in front of the PV array will generate a wake zone behind them, significantly altering the wave and current environment of the piles behind them and reducing the drag force on the piles.

[0008] The inventors constructed an intelligent load factor prediction model by combining wind tunnel and water tank test data with high-fidelity CFD simulation data. Then, based on the predicted load factor, they calculated the wind load, wave load, and flow load values ​​considering the shading effect. This model overcomes the various uncertainties transmitted from environmental parameters to structural loads. It not only utilizes deep learning to solve the complex calculation problems caused by the shading effect, but also ensures the rationality of the corresponding load values ​​at the mechanical level through physical equations.

[0009] Preferably, S1 specifically includes: S11: constructing a mapping sample set between wind environment parameters and photovoltaic structure parameters and wind load shape coefficient considering panel shading effect; S12: constructing a mapping sample set between wave flow environment parameters and pile structure parameters and drag force coefficient and inertial force coefficient considering pile group shading effect.

[0010] Preferably, the wind environment parameters include: wind direction angle; the photovoltaic structure parameters include: photovoltaic array spacing and photovoltaic panel tilt angle; the wave flow environment parameters include: wave height, wave period, wave direction angle and ocean current velocity; and the pile structure parameters include: pile diameter and pile spacing.

[0011] Preferably, in step S2, the data is preprocessed by standardization before model training, that is, all input features are scaled to a uniform scale, and the standardization formula is the Z-Score formula. In the formula, x is the value of a single data point. The average value of the dataset. This represents the standard deviation of the dataset. For example, scaling all input features to the range of 0 to 1 so that they follow a standard normal distribution with a mean of 0 and a variance of 1 can significantly accelerate model convergence and improve model training stability.

[0012] Preferably, in step S2, the dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0013] Preferably, in step S2, an early-stop strategy is used to control the training of the intelligent load factor prediction model, that is, when the loss function value L on the validation set is within N consecutive training cycles... val Training stops when the loss stops decreasing (or falls below a threshold). This strategy continuously monitors the model's loss on the validation set, and automatically terminates training and saves the best-performing model state once it detects that the validation set loss no longer decreases or even starts to increase over several consecutive epochs.

[0014] Preferably, the number of hidden layer neurons, batch training size, and learning rate hyperparameters of the intelligent load factor prediction model are all optimized using a Bayesian optimization method. The expected improvement function of the Bayesian optimization is... In the formula The objective function value, This represents the currently known optimal objective function value.

[0015] Preferably, the formula for calculating the wind load value is: In the formula Wind load values ​​considering shading effects for a single panel. Let z be the wind vibration coefficient at height z. To account for the wind load shape coefficient after considering the panel shading effect, This is the coefficient of wind pressure height variation. This is the basic wind pressure.

[0016] Preferably, the formula for calculating the wave load is: In the formula Wave load values ​​considering the shielding effect of pile groups for a single pile. The density of seawater, To account for the drag force coefficient of the pile group shielding effect, Let D be the inertial force coefficient, A be the pile diameter, and u be the cross-sectional area of ​​the pile. This refers to the horizontal acceleration of the water particles.

[0017] Preferably, the formula for calculating the flow load is: In the formula The flow load value is determined for a single pile considering the shielding effect of pile groups. The density of seawater, To account for the drag force coefficient of the pile group shielding effect, This refers to the ocean current velocity.

[0018] Advantages of this solution: 1. A deep integration of physical mechanisms and data-driven methods has been achieved, constructing a hybrid architecture. This architecture utilizes deep learning models to determine key load calculation coefficients that are difficult to determine using traditional methods, and then substitutes these coefficients into a validated load physics model for final load calculation. This approach leverages the advantages of deep learning in handling complex nonlinear mappings while ensuring the mechanical validity of the results through physical equations, achieving a complementary advantage between intelligent prediction and physical mechanisms.

[0019] 2. The complex shading effects of wind, wave, and current loads are precisely quantified, overcoming the limitations of traditional methods. Through deep learning technology, the complex shading effects of wind, wave, and current loads on offshore photovoltaic arrays are accurately quantified. The trained model can dynamically reflect the influence of multiple factors such as wind direction angle, wave direction angle, array spacing, and pile foundation layout, accurately predicting the actual load reduction of components at different locations in the array. This overcomes the problem of overly conservative results caused by the simplification of shading effects in traditional methods, significantly improving the realism of calculations and engineering economy while ensuring structural safety.

[0020] 3. This method provides efficient site adaptation capabilities based on transfer learning. It fine-tunes and corrects a pre-trained general-purpose deep learning model using short-term field monitoring data, forming an efficient model adaptation mechanism. Utilizing limited small-sample field data, the general-purpose model can quickly adapt to the environmental characteristics and structural response properties of specific engineering sites, significantly improving the engineering reliability of the prediction results. The core innovation of this method lies in the first-time quantification of the shading effect between photovoltaic arrays and the integration of deep learning and physical equations to construct a hybrid computational model. This effectively solves the load calculation bias caused by neglecting complex aerodynamic and hydrodynamic interference between components in traditional methods, significantly improving the accuracy and engineering applicability of structural load prediction in complex marine environments. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the training process of the model in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of a multilayer feedforward neural network according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram illustrating an example of photovoltaic array arrangement according to an embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of the marine photovoltaic monitoring system according to an embodiment of the present invention. Detailed Implementation

[0026] The following detailed description illustrates the specific implementation method: Example: A smart calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects, such as... Figure 1 As shown, it includes: S1: Construct a multi-source mapping sample set between ocean wind, wave and current environmental parameters and load coefficients corrected for shading effects.

[0027] This step specifically includes: S11: Construct a mapping sample set between wind environment parameters and photovoltaic structure parameters and wind load shape coefficients considering panel shading effects.

[0028] Determining the wind load shape coefficient is extremely difficult in the calculation of wind loads on pile-foundation fixed offshore photovoltaic systems. Offshore photovoltaic arrays are large in scale and complex in layout, and wind can generate complex mutual interference and shading effects. The standard shape coefficients in the specifications cannot accurately reflect the true wind pressure distribution in this three-dimensional space. Furthermore, the wind load shape coefficient is greatly affected by the wind direction angle. Therefore, this sample set aims to establish a mapping relationship between wind environment parameters, structural parameters, and the wind load shape coefficient considering the shading effect. The wind environment parameters include the wind direction angle, and the structural parameters include the array spacing and panel tilt angle.

[0029] S12: Construct a mapping sample set between wave flow environment parameters and pile structure parameters and drag force coefficients and inertial force coefficients considering the shielding effect of pile groups.

[0030] In the calculation of wave and current loads for fixed offshore photovoltaic (PV) piles, the drag force coefficient and inertial force coefficient are semi-empirical parameters that cannot be directly derived from theory. Both wave and current loads involve the drag force coefficient, while wave loads involve the inertial force coefficient. Furthermore, due to wake and vortex interference, the piles in front of the PV array generate a wake zone behind them, significantly altering the wave and current environment of the piles behind and reducing the drag force on the piles. Therefore, this step aims to establish a mapping relationship between wave and current environment parameters, structural parameters, and the drag force and inertial force coefficients considering the pile group shielding effect. The wave and current environment parameters include wave height, wave period, wave direction angle, and current velocity, while the structural parameters include pile diameter and pile spacing.

[0031] This embodiment specifically collects publicly available wind tunnel and flume test reports for the target sea area, as well as high-fidelity numerical simulation data under multiple operating conditions from the large-scale commercial CFD software Fluent. A total of 500 sets of operating conditions covering different environmental parameters, such as wind and wave angles from 0° to 180°, and structural parameters such as array spacing ratios from 1.5 to 3.0 and panel tilt angles from 10° to 40°, are then compiled to construct a structured sample database.

[0032] S2: Construct an intelligent prediction model for load factors based on a multi-layer feedforward neural network. The load factor calculation model includes a prediction model for wind load shape coefficient and prediction models for drag force coefficient and inertial force coefficient.

[0033] This step specifically includes: S21: Construct an intelligent prediction model for wind load shape coefficient.

[0034] In this step, environmental and structural parameters from a wind environment sample set are used as model input, and a wind load shape coefficient considering shading effects is used as output. A deep learning model is trained, and the mapping relationships are constructed to obtain an intelligent and refined calculation model for the wind load of photovoltaic panels. This model can intelligently and precisely predict the wind load shape coefficient of each individual panel in the photovoltaic array based on the input environmental and structural parameters. This coefficient also considers the load reduction caused by mutual shading between panels. Subsequently, by combining the basic wind pressure and other coefficients (wind vibration coefficient, wind pressure height variation coefficient), the final wind load value for each individual panel in the array is obtained.

[0035] The model building process is as follows: First, the previously constructed mapping sample set containing 500 working conditions is used as the basic data source. The model training process is as follows: Figure 2 As shown, before training, the dataset needs to be standardized, which means scaling all input features to a uniform scale. For example, scaling all input features to the range of 0 to 1 or making them follow a standard normal distribution with a mean of 0 and a variance of 1. This can significantly accelerate model convergence and improve training stability. The standardization formula is the Z-Score formula. .

[0036] Subsequently, the dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for model learning and parameter updates; the validation set was used to monitor model performance, perform hyperparameter tuning, and serve as the basis for triggering the "early stopping" mechanism during training; the test set did not participate in training at all, but was only used at the end to objectively evaluate the model's final generalization ability.

[0037] In this step, a model is built based on a multi-layer feedforward neural network, such as... Figure 3 As shown, the network consists of an input layer, three hidden layers, and an output layer connected sequentially. The number of neurons in the input layer is strictly equal to the dimension of the input features. The hidden layers automatically learn and extract abstract features from the original input, ranging from low to high order, that can characterize the physical laws of complex flow fields, through progressive nonlinear transformations. Modified linear units are used as activation functions in these hidden layers. Finally, the number of neurons in the output layer corresponds to the number of labels to be predicted, and this layer uses a linear activation function to ensure that the model's output is an unrestricted continuous real value.

[0038] In this step, the model's "learning" or "training" is a process of minimizing prediction error through iterative optimization. In each training epoch, data is fed into the network in mini-batches for training. The loss function is Mean Squared Error (MSE), and the Adam (Adaptive Matrix Estimation) optimizer is used to perform parameter updates. The initial learning rate is set to a small value to ensure the precision of parameter updates. The training epochs for "forward propagation → loss calculation → backpropagation → parameter update" are set to 2000.

[0039] In this step, an "early stopping" strategy is adopted during model training. Once the validation set loss L is detected, the strategy is initiated. val If the model stops declining or even starts to rise over several consecutive cycles, the training will automatically terminate and the best-performing model state will be saved to prevent the model from "overfitting" on the training data. This means that the model has over-learned the details and noise of the training data, resulting in poor performance on new data.

[0040] The optimal values ​​for hyperparameters such as the number of hidden layer neurons, batch size, and learning rate of the computational model are determined using Bayesian optimization. The expected improvement function is... .

[0041] In this step, the wind load shape coefficient calculation model was constructed based on the above content. A precise mapping relationship was successfully established between the input (environmental and structural parameters) and the output (wind load shape coefficient considering shading effects). It can complete the calculation tasks that would take hours or even days to complete using traditional methods with extremely high efficiency. The specific parameters of the basic network model used are shown in Table 1: Table 1 Network Model Parameter Table Parameter type Parameter settings Number of hidden layers 3 Number of neurons 64 Activation function ReLU Batch training size 64 loss function MSE Learning rate 0.001 Optimizer Adam Training times Early Stopping In Table 1, ReLU represents the linear rectification function. Based on the above model and the input environmental and structural parameters, the wind load shape coefficient of each individual panel in the photovoltaic array after considering the shading effect can be quickly predicted.

[0042] In this scheme, the intelligent prediction model for the load factor of wind load shape coefficient applies a generative adversarial network (GAN) based on data homeomorphism. A load factor prediction GAN model is constructed based on the aforementioned basic network model. By learning from limited historical photovoltaic data, it generates augmented data that is consistent with the real data in both statistical characteristics and physical meaning. The GAN training process is essentially a process of solving for the minimum and maximum values ​​of a bivariate function, with the specific formula being: In the formula and Let D represent the probability of real data and generated data, and G represent the generator. To ensure that the generated data not only have similar statistical characteristics but also conform to physical laws, a physical constraint term is added to the loss function of the generator G. In the formula, P is the physical equation. To weigh the parameters.

[0043] The specific process of generating and training the GAN model for wind load shape coefficient is as follows: Initialization: Randomly initialize the neural network parameters of generator G and discriminator D, define the noise vector z (Gaussian distribution), learning rate (lr=0.0002), and batch size (64). The tradeoff parameter is set to a positive scalar. =0.2 to balance the effects of damage and physical consistency. Sample of actual wind load shape coefficient. Normalized to the [-1, 1] interval, matching the Tanh activation function of the generator output layer.

[0044] Iterative training loop phase: First, train the discriminator D. With the parameters of the generator G fixed, only the discriminator D is updated. A batch of real data samples is sampled from the real wind load dataset, and a batch of noise vectors is sampled from the noise distribution and input into the generator G to obtain generated data G(z). Finally, the loss function L of the discriminator D is calculated. D Secondly, train the generator G, fix the parameters of the discriminator D, update only the generator G, resample the noise vector, generate fake data G(z), and finally calculate the total loss L of the generator G. G This includes adversarial losses and physical constraints: , .

[0045] Convergence criterion: The fluctuation of the physical constraint term P is less than the threshold, i.e. .

[0046] The final generator and discriminator are as follows: Generator G: ① Input: Noise vector z with dimension 128, sampled from Gaussian distribution.

[0047] ② Core structure: The input layer is a linear transformation, followed by BatchNorm and LeakyReLU activation function (slope = 0.2); the hidden layer consists of 3 fully connected layers, each followed by BatchNorm and LeakyReLU; the output layer maps features to the output dimension and uses Tanh activation function to restrict the output to the [-1,1] interval to match the standardized data.

[0048] Discriminator D ① Input: Actual wind load shape coefficient data x or generated data G(z).

[0049] ② Core structure: The input layer flattens the data into a vector, followed by LeakyReLU (slope=0.2) and Dropout (dropout rate=0.3) to prevent overfitting; the hidden layer consists of 3 fully connected layers, each followed by LeakyReLU; the output layer uses the Sigmoid activation function, and the output sample is the true probability value (between 0 and 1).

[0050] S22: Construct intelligent prediction models for drag force coefficient and inertial force coefficient.

[0051] In this step, environmental and structural parameters from the wave-current environment sample set are used as model inputs, and drag force coefficients and inertial force coefficients, considering the pile group shielding effect, are used as outputs. A deep learning model is trained, and the mapping relationships are constructed to obtain an intelligent and refined calculation model for pile wave loads and flow loads. This model can intelligently and precisely predict the drag force coefficients and inertial force coefficients of each row of individual piles in the photovoltaic array based on the input environmental and structural parameters. The drag force coefficients also consider the load reduction caused by the pile group effect. Subsequently, the final wave-current load values ​​of each individual pile in the array are calculated using the Morison equation.

[0052] In this step, the input parameters in the wave-current environment sample set are used as features, and the drag force coefficient and inertial force coefficient considering the pile blockage effect are used as labels to train a deep neural network. The specific training process and parameter settings are the same as those of the wind load shape coefficient calculation model.

[0053] In this step, based on the input environmental and structural parameters, the drag force coefficient and inertial force coefficient of each individual pile in the photovoltaic array, considering the shading effect of the group piles, are quickly predicted.

[0054] The generator and discriminator of the GAN model for drag force coefficient and inertia force coefficient are as follows: Generator G: ① Input: Noise vector z with dimension 128, sampled from Gaussian distribution.

[0055] ② Core structure: The input layer is a linear transformation, followed by BatchNorm and LeakyReLU activation function (slope = 0.2); the hidden layer consists of 3 fully connected layers, each followed by BatchNorm and LeakyReLU; the output layer maps features to the output dimension and uses Tanh activation function to restrict the output to the [-1,1] interval to match the standardized data.

[0056] Discriminator D ① Input: Actual drag force coefficient and inertial force coefficient data x or generated data G(z).

[0057] ② Core structure: The input layer flattens the data into a vector, followed by LeakyReLU (slope=0.2) and Dropout (dropout rate=0.4) to prevent overfitting; the hidden layer consists of 3 fully connected layers, each followed by LeakyReLU; the output layer uses the Sigmoid activation function, and the output sample is the true probability value (between 0 and 1).

[0058] S3: Based on the extreme marine wind, wave and current environmental parameters and load factor intelligent prediction model of the target sea area, the load factor is predicted, and then the wind load, wave load and current load values ​​considering the shielding effect are calculated based on the predicted load factor.

[0059] The formula for calculating wind load is as follows: (1) In the formula, Wind load values ​​considering shading effects for a single panel. Let z be the wind vibration coefficient at height z. To account for the wind load shape coefficient after considering the panel shading effect, This is the coefficient of wind pressure height variation. This is the basic wind pressure.

[0060] The formula for calculating wave load is as follows: (2) In the formula, Wave load values ​​considering the shielding effect of pile groups for a single pile. The density of seawater, To account for the drag force coefficient of the pile group shielding effect, Let D be the inertial force coefficient, A be the pile diameter, and u be the cross-sectional area of ​​the pile. This refers to the horizontal acceleration of the water particles.

[0061] The formula for calculating flow load is as follows: (3) In the formula, The flow load value is determined for a single pile considering the shielding effect of pile groups. The density of seawater, To account for the drag force coefficient of the pile group shielding effect, This refers to the ocean current velocity.

[0062] In this embodiment, based on current standards such as the "Design Code for Offshore Photovoltaic Power Generation Systems" (NB / T 11744-2024) and site conditions, the design environmental parameters are as follows: wind speed v = 25 m / s, prevailing wind direction is N (0°); significant wave height H = 3.6 m, wave period T = 7.5 s, prevailing wave direction is S (180°), and current velocity... =1.2m / s. The pile foundations are spaced 36m north-south and 19.5m east-west, with a pile diameter of 1.1m and a panel tilt angle of 15°. In this embodiment, the photovoltaic array is arranged as follows: Figure 4 As shown.

[0063] 1. Substitute environmental parameters and design array parameters into the intelligent prediction model of wind load shape coefficient. Taking the photovoltaic panel in the third column as an example, the wind load shape coefficient of each single panel is shown in Table 2.

[0064] Table 2 Wind load shape coefficients for each individual panel considering shading effects Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Wind load shape coefficient considering shading effect 0.14 0.14 0.14 0.10 0.26 The wind loads of each individual panel are calculated according to formula (1) as shown in Table 3.

[0065] Table 3 Wind load values ​​for each individual panel considering shading effects Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Wind load (kN) 5.24 5.24 5.24 3.72 9.53 2. Substituting the wave flow environment parameters and design array parameters into the intelligent prediction model of drag force coefficient and inertial force coefficient, the drag force coefficient and inertial force coefficient of each individual pile in the photovoltaic array considering the shading effect of the group piles are shown in Table 4.

[0066] Table 4. Drag force coefficients for each individual pile considering the shielding effect of pile group. and inertial force coefficient Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 drag coefficient 0.96 0.98 1.01 1.06 1.10 Inertia coefficient 1.82 1.82 1.82 1.82 1.82 The wave loads of each single pile are calculated according to formulas (2) to (3) as shown in Table 5.

[0067] Table 5 Wave loads on individual piles and flow load Value Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Wave load (kN) 15.96 16.18 16.49 17.01 17.41 Flow load (kN) 4.08 4.17 4.30 4.51 4.68 This approach utilizes deep learning models to determine key load calculation coefficients that are difficult to ascertain using traditional methods. These coefficients are then substituted into a validated load physics model for final load calculation. This approach leverages the advantages of deep learning in handling complex nonlinear mappings while ensuring the mechanical validity of the results through physical equations, achieving a complementary advantage between intelligent prediction and physical mechanisms.

[0068] S4: Obtain actual marine wind, wave, and current environmental parameters and structural load data for the target pile-based offshore photovoltaic structure, and make adaptive corrections to the load factor calculation model.

[0069] Set up a monitoring system at the engineering site, such as Figure 5As shown, the equipment includes a marine wind speed and direction instrument, a wave observation buoy, an acoustic Doppler current profiler, and a load pressure sensor installed on the demonstration photovoltaic structure. Based on existing technologies, it continuously collects actual environmental parameters and structural load data for one month.

[0070] Using collected field data, the pre-trained model was fine-tuned. Transfer learning was employed to freeze the model's underlying network, retraining only the last few layers. This allowed the model to quickly adapt to the site characteristics, resulting in a refined and accurate intelligent calculation model for environmental loads.

[0071] Based on a month-long monitoring period at the case site, under a wind direction angle of 0°, the wind load shape coefficients of each individual panel, calculated from actual measurements, are shown in Table 6 (the values ​​in the table have taken into account the shading effect). The initial model's prediction results showed a maximum error of 12.23% compared to the measured values. Subsequently, the pre-trained model was fine-tuned using measured data, and the revised high-precision model's prediction results are shown in Table 7, with its maximum error reduced to 5.37%.

[0072] Table 6 Initial model predictions of wind load shape coefficients for each individual panel Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Initial calculation of body size coefficient 0.14 0.14 0.14 0.10 0.26 Site measured shape coefficient 0.125 0.126 0.125 0.092 0.248 error(%) 11.64 11.49 12.23 9.27 4.99 Table 7. Wind load shape coefficients for each individual panel predicted by the modified model. Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Correcting the calculation of body size coefficient 0.131 0.131 0.131 0.096 0.255 Site measured shape coefficient 0.125 0.126 0.125 0.092 0.248 error(%) 4.22 4.05 4.69 5.37 2.79 Similarly, under the condition of a wave angle of 180°, the drag force coefficient and inertia force coefficient of each individual pile calculated from actual measurements are shown in Table 8 (the values ​​in the table have taken into account the shielding effect of the pile group). Compared with the measured values, the maximum error of the drag force coefficient in the initial model was 14.12%, and the error of the inertia force coefficient was 13.65%. Subsequently, the pre-trained model was fine-tuned using measured data. The prediction results of the corrected high-precision model are shown in Table 9, where the maximum error of the drag force coefficient was reduced to 5.21%, and the error of the inertia force coefficient was reduced to 4.76%.

[0073] Table 8 Initial model prediction of drag force coefficients for each individual pile and inertial force coefficient Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Initial calculation of drag coefficient 0.96 0.98 1.01 1.06 1.10 Site-measured drag coefficient 0.84 0.87 0.89 0.98 1.02 error(%) 14.12 12.44 12.96 8.52 7.93 Initial calculation of inertial force coefficients 1.82 1.82 1.82 1.82 1.82 Measured inertial force coefficient at the site 1.60 1.60 1.60 1.60 1.60 error(%) 13.65 13.65 13.65 13.65 13.65 Table 9. Predicted drag force coefficients for each individual pile using the modified model. and inertial force coefficient Arrangement Number A1-3 A2-3 A3-3 A4-3 A5-3 Correction of the calculated drag coefficient 0.92 0.93 0.97 1.02 1.06 Site-measured drag coefficient 0.84 0.87 0.89 0.98 1.02 error(%) 4.10 5.21 4.64 3.52 3.48 Corrected calculation of inertial force coefficient 1.68 1.68 1.68 1.68 1.68 Measured inertial force coefficient at the site 1.60 1.60 1.60 1.60 1.60 error(%) 4.76 4.76 4.76 4.76 4.76 This approach utilizes on-site measured data to effectively validate the model's predictions, ensuring its applicability and reliability in the current site environment. Furthermore, the newly acquired field data enriches the training dataset, enhancing the model's generalization ability. The resulting refined and improved high-precision intelligent environmental load calculation model demonstrates higher prediction accuracy and greater practicality in subsequent similar projects, providing more reliable data support for engineering design and safety assessment.

[0074] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A smart calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects, characterized in that, include: S1: Construct a multi-source mapping sample set between ocean wind, wave and current environmental parameters and load coefficients corrected for shielding effects; S2: Construct an intelligent load factor prediction model based on a multi-layer feedforward neural network. The model includes: a wind load shape factor prediction model considering panel shading effects and a drag force and inertial force factor prediction model considering pile group shading effects. The intelligent load factor prediction model applies a generative adversarial network (GAN) based on data homeomorphism. The GAN training process involves solving for the minimum and maximum values ​​of a bivariate function, with the specific formula being: In the formula and This represents the probabilities of real and generated data; where a physical constraint term is added to the loss function of the generator G. In the formula, P is the physical equation. To weigh the parameters; S3: Predict the load factor corrected for the shielding effect based on the extreme marine wind, wave and current environmental parameters of the target sea area, and then calculate the wind load, wave load and current load considering the shielding effect based on the predicted load factor. S4: Obtain actual marine wind, wave, and current environmental parameters and structural load data for the target pile-based offshore photovoltaic structure, and adaptively modify the intelligent calculation model of environmental load.

2. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: S1 specifically includes: S11: constructing a mapping sample set between wind environment parameters and photovoltaic structure parameters and wind load shape coefficient considering panel shading effect; S12: constructing a mapping sample set between wave flow environment parameters and pile structure parameters and drag force coefficient and inertial force coefficient considering pile group shading effect.

3. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 2, characterized in that: The wind environment parameters include: wind direction angle; the photovoltaic structure parameters include: photovoltaic array spacing and photovoltaic panel tilt angle; the wave and current environment parameters include: wave height, wave period, wave direction angle and ocean current velocity; and the pile structure parameters include: pile diameter and pile spacing.

4. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: In step S2, the data is preprocessed by standardization before model training, which involves scaling all input features to a uniform scale. The standardization formula is the Z-Score formula. In the formula, x is the value of a single data point. The average value of the dataset. is the standard deviation of the dataset.

5. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: The S2 section employs an early-stop strategy to control the training of the intelligent load factor prediction model, meaning that within N consecutive training cycles, the loss function value L on the validation set... val Training stops when the price stops falling or falls below a threshold.

6. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: The number of hidden layer neurons, batch size, and learning rate hyperparameters of the intelligent load factor prediction model were all optimized using Bayesian optimization. The expected improvement function of Bayesian optimization is... In the formula The objective function value, This represents the currently known optimal objective function value.

7. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: The formula for calculating the wind load value is: In the formula Wind load values ​​considering shading effects for a single panel. Let z be the wind vibration coefficient at height z. To account for the wind load shape coefficient after considering the panel shading effect, This is the wind pressure height variation coefficient. This is the basic wind pressure.

8. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: The formula for calculating the wave load is: In the formula Wave load values ​​considering the shielding effect of pile groups for a single pile. The density of seawater, To account for the drag force coefficient of the pile group shielding effect, Let D be the inertial force coefficient, A be the pile diameter, and u be the cross-sectional area of ​​the pile. This refers to the horizontal acceleration of the water particles.

9. The intelligent calculation method for environmental loads of fixed marine photovoltaic systems considering shading effects according to claim 1, characterized in that: The formula for calculating the flow load is: In the formula The flow load value is determined for a single pile considering the shielding effect of pile groups. The density of seawater, To account for the drag force coefficient of the pile group shielding effect, This refers to the ocean current velocity.

Images