Neural network photovoltaic ash deposition prediction method based on ship swing characteristics

Abstract

The invention discloses a neural network photovoltaic ash deposition prediction method based on ship swing characteristics, which adds the ship swing characteristics on the basis of the traditional neural network photovoltaic ash deposition prediction so as to adapt to the working condition of ocean dynamic change. According to the method, on the basis of introducing characteristic quantities of open-circuit voltage, short-circuit current, maximum power point voltage and maximum power point current traditionally, a new characteristic quantity of ship swing is introduced, and a CNN-LSTM-Attention hybrid neural network is adopted to input the characteristic quantities to predict ship photovoltaic dust deposition. And the problems of low convergence speed, long search time, poor generalization ability and the like of photovoltaic ash deposition prediction caused by swinging of ships on the ocean are solved. A simulation model is established by analyzing the characteristics of ship swing, sample data is acquired, input and output data training is performed by using a neural network algorithm, and prediction of ship photovoltaic system dust deposition is realized.

Description

Technical Field

[0001] This invention belongs to the technical field of ship photovoltaic power generation, and mainly relates to a neural network photovoltaic dust accumulation prediction method based on ship sway characteristics. Background Technology

[0002] Photovoltaic dust accumulation prediction based on ship rolling characteristics is applicable to the operating conditions of ships at sea, and can improve the efficiency of photovoltaic power generation under these conditions, helping to optimize the operation and maintenance of photovoltaic systems. With the growth of global energy demand and the increasing awareness of environmental protection, photovoltaic power generation systems have been widely used. In particular, deploying photovoltaic systems on ships can utilize the abundant sunlight resources at sea to provide clean energy for vessels.

[0003] Traditional methods for detecting dust accumulation on photovoltaic panels mainly include electrical measurement methods, image detection methods, and intelligent system detection methods.

[0004] Traditional electrical measurement methods detect dust accumulation in photovoltaic (PV) systems by modifying circuit structures and incorporating voltage and current sensors. They analyze the characteristics of the PV module's IV curve, diagnose shading faults by combining changes in electrical parameters, and verify the feasibility of these methods through simulation and experiments. Image monitoring methods use infrared detectors to capture infrared images of the dust-accumulated areas of the PV array, diagnosing faults based on the temperature difference between the faulty and normal areas. These methods are susceptible to external environmental interference and require the installation of infrared detectors, increasing system costs. Intelligent system detection methods integrate monitoring data from PV power plants and weather stations to diagnose common dust accumulation fault types.

[0005] Harsh conditions in the marine environment, such as high humidity, salt spray, strong winds, and ship swaying, pose challenges to the stability and efficiency of photovoltaic systems. The problem of dust accumulation on the surface of photovoltaic modules is particularly prominent, affecting the light transmittance and power generation efficiency of photovoltaic panels.

[0006] Most of the methods mentioned above are designed for detecting dust accumulation in land-based photovoltaic power plants, neglecting the dynamic operating conditions at sea, which leads to instability in marine photovoltaic power generation. Therefore, adapting to these conditions and improving the efficiency of marine photovoltaic power generation is an urgent problem to be solved. Summary of the Invention

[0007] This invention addresses the problem of unstable photovoltaic power generation caused by the swaying of marine vessels by providing a neural network-based method for predicting photovoltaic dust accumulation based on the characteristics of ship swaying. In addition to traditional feature quantities, the method incorporates the characteristics of ship swaying into the neural network for dust accumulation prediction.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a neural network photovoltaic dust accumulation prediction method based on ship sway characteristics, comprising the following steps:

[0009] S1: Establish a photovoltaic array model and obtain sample data; the photovoltaic array model is specifically as follows:

[0010]

[0011] In the formula: I PH Photocurrent; I O1 I O2 For V D1 V D2 The reverse saturation current, R S The resistors are connected in series, and a1 and a2 are V. D1 V D2 The ideal factor, N S V represents the number of cells connected in series in the module. T1 V T2 R is the thermal voltage of the solar cell. SH The resistors are connected in parallel;

[0012]

[0013] In the formula: k is the Boltzmann constant, k = 1.38 × 10 -23 J / K; T is the temperature of the solar cell; q is the electron charge, q = 1.602 × 10⁻⁶ -19 C;

[0014] S2: Determine the input features of the neural network. It is understandable that, considering that the dust accumulation on the photovoltaic panel will directly affect the incident light intensity of the photovoltaic module, short-circuit current, open-circuit voltage, maximum power point voltage, and maximum power point current are selected as the input features of the neural network. The type of dust accumulation is determined and labeled as the output. The dataset is constructed using the input and output features.

[0015] S3: Initialize the neural network, determine the number of nodes in each layer of the neural network, and initialize the weights and thresholds of the neural network;

[0016] S4: Based on the characteristics of ship swaying, optimize the selection of input feature quantities in step S2; the selection of ship swaying feature quantities is determined by the solar azimuth angle, solar altitude angle, and ship swaying period parameters to determine the solar radiation intensity received by the photovoltaic module surface;

[0017] S5: Train the model, calculate the percentage error of photovoltaic dust accumulation prediction, and compare the predicted value with the actual value.

[0018] Furthermore, in step S2, by building a model and observing the IV and PV curves of the photovoltaic module, it can be determined that different dust accumulation conditions on the surface of the photovoltaic module will change four characteristic quantities: short-circuit current, open-circuit voltage, maximum power point voltage, and maximum power point current. This allows the determination of the input characteristic quantities for the neural network. The equivalent light intensity affected by dust accumulation is:

[0019] E′=η sc E

[0020] In the formula: η sc E represents the shading coefficient marked on the dust-covered shadow of the photovoltaic panel, where E is the original light intensity.

[0021] Furthermore, in step S4, based on the characteristics of ship swaying, a simulation model of solar radiation energy characteristic of ship swaying is established. The simulation model of solar radiation energy characteristic of ship swaying is as follows:

[0022] cosi=cosβsinh+sinβcoshcos(α-γ)

[0023] I n =ξ0I sc P m sinh

[0024]

[0025] I g =ρ(I n +I h )

[0026]

[0027] Where: β is the tilt angle; h is the solar altitude angle; α is the solar azimuth angle; γ is the slope azimuth angle; ξ0 is the Earth orbit eccentricity correction coefficient; I sc Where is the solar constant, P is the atmospheric transparency coefficient, m is the atmospheric optical mass, and θ is the atmospheric constant. max Let θ be the maximum angle of the ship's roll, T be the ship's roll period, and θ0 be the initial phase.

[0028] Furthermore, step S5 specifically includes the following steps:

[0029] S51: The input data is processed through a one-dimensional convolutional layer, which includes convolutional layers and pooling layers. The convolutional layer is used to extract local features of the data, and the pooling layer is used to reduce the dimensionality of the features. It can be understood that the role of this convolutional layer is to extract important features from the original time-series data, such as periodic changes, trends, and possible outliers. This step is crucial because it provides the LSTM layer with preprocessed and refined input data. The core advantage of CNNs lies in their ability to effectively extract features from the input data through convolutional and pooling layers. In the convolutional layer, the network extracts local features of the data by applying a series of convolutional operations, while the pooling layer reduces the dimensionality of the features, thereby reducing computational complexity and extracting the most important features. In the application of dust accumulation prediction, CNNs can process dust accumulation prediction data through their convolutional layers. These layers automatically detect key features in the data through filters, such as periodic changes, long-term trends, seasonal factors, and possible outliers. These features are crucial for understanding the historical behavior and future trends of freight volume.

[0030] S52: The processed data is fed into an LSTM layer, which is used to capture long-term dependencies and complex temporal dynamics in the data, and a first dropout layer is introduced. Understandably, the first dropout layer helps prevent overfitting and ensures that the model still performs well when faced with new, unseen data.

[0031] S53: An attention mechanism was added to the model. The role of the attention layer is to weight the output of the LSTM layer, enabling the model to focus on the most critical parts of the sequence. Understandably, this weighting makes the model more accurate in analysis and prediction because it can distinguish which information is most important for the final prediction.

[0032] S54: The attention-processed data is fed into the first fully connected layer, which aggregates information and maps it to an intermediate representation space. By applying the ReLU activation function, the model can enhance its ability to capture nonlinear relationships and further enhance the accuracy of prediction.

[0033] S55: In the final data processing stage, the model introduces a second dropout layer and a second fully connected layer to enhance its generalization ability. The second fully connected layer maps the processed features to the prediction output space.

[0034] S56: The second fully connected layer maps the processed information to the prediction output space. The role of the second fully connected layer is to transform all the learning results of the model into specific prediction values, providing support for the final decision.

[0035] Furthermore, in step S52, the unit structure of the LSTM layer (Long Short-Term Memory network) comprises four main parts: cell state, forget gate, input gate, and output gate. The forget gate uses a sigmoid activation function to determine the information that needs to be retained from the cell state. The forget gate receives the hidden state (part of the memory cell) from the previous time step and the input from the current time step, and outputs a value between 0 and 1, representing the degree of information retention at the corresponding position. 0 represents complete forgetting, and 1 represents complete retention. The mathematical expression for the forget gate is as follows:

[0036] f t =σ|W f ·[H t-1 ,x t ]+b f |

[0037] Among them, f t For the forget gate, σ represents the Sigmoid function, W f and b f These are the forget gate weights and bias coefficients, respectively; H t-1 For the cell output at time t-1, x t Input the information at time t;

[0038] The input gate is used to control the input of new information, allowing the network to selectively update the cell state to adapt to the currently input information. The mathematical expression for the input gate is as follows:

[0039] i t =σ|W t ·[H t-1 ,x t ]+b t |

[0040]

[0041] Where: i t The input gate determines how many new candidate values ​​are generated. It will be added to the cellular state. It is a new candidate value for the current time step, which may be added to the cell state, C. t It represents the cell state at the current time step, W. C It is the weight matrix of the candidate values, b C W is the bias term for the candidate values. t It is the weight matrix of the input gate, b t is the bias term of the input gate, and tanh is the hyperbolic tangent activation function, which compresses any real value to the interval [-1, 1].

[0042] The output gate determines whether the information in the current hidden state will be output to the next layer of the network or used for final prediction. The output gate uses a sigmoid activation function to control which parts of the hidden state will be activated. The mathematical expression of the output gate is as follows:

[0043] O t =σ|W O ·[H t-1 ,x t ]+b O

[0044] H t =O t ×tanh(C t )

[0045] Among them: O t It is the output of the output gate, W O It is the weight matrix of the output gate, b O It is the bias term of the output gate.

[0046] Furthermore, in step S53, the attention mechanism model network is used to capture the information features output by the LSTM layer in the prediction model, divide the weight ratio, extract key information, and improve the accuracy of the prediction model. The hidden layer output of the LSTM is used as the input of the attention mechanism. The hidden layer output of the LSTM first passes through a first fully connected layer. The output of the first fully connected layer is normalized by the softmax function to generate a set of weights. The weights are used to show the relative importance of the hidden state at each time step to the final prediction. The attention mechanism calculation process is as follows:

[0047] S i =tanh(WH i +b i )

[0048] α i =softmax(S i )

[0049]

[0050] Where: α i The weights are obtained by applying the softmax function to the attention scores; H i S is the output of the last LSTM hidden layer. i The score output for each hidden layer; C i b is the value after weighted average summation. t It is a bias term.

[0051] Understandably, this mechanism allows the model to allocate more attention to key time steps during prediction, thereby improving overall prediction accuracy and efficiency. This dynamic weight allocation process is adaptive, ensuring that the model can identify and focus on the most informative parts of complex sequence data.

[0052] This method proposes a new feature quantity related to the ship's rolling motion on the ocean. The angle of the ship's rolling motion corresponds to the angle between sunlight and the photovoltaic panel. Building upon traditional methods that introduce open-circuit voltage, short-circuit current, maximum power point voltage, and maximum power point current, this new feature quantity of ship rolling motion is introduced. A CNN-LSTM-Attention hybrid neural network is used to input these features for predicting dust accumulation in ship photovoltaic systems. This improves upon the slow convergence speed, long search time, and poor generalization ability issues caused by the ship's rolling motion on the ocean. A simulation model is established by analyzing the characteristics of ship rolling motion, sample data is obtained, and the neural network algorithm is used to train the input and output data to predict dust accumulation in the ship's photovoltaic system. Attached Figure Description

[0053] Figure 1 This is a block diagram of the overall structure of the ship photovoltaic power generation system of the present invention.

[0054] Figure 2 This is a flowchart illustrating the structure of the present invention.

[0055] Figure 3 This is a diagram of the CNN-LSTM-Attention neural network prediction model structure.

[0056] Figure 4 This is a comparison chart of prediction results for the test set.

[0057] Figure 5 This is a regression plot of the predicted results.

[0058] Figure 6 This is a comparison chart of the prediction results for the test set and the prediction set. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0061] like Figure 1 As shown, the overall structure of the ship photovoltaic power generation system includes a solar radiation energy module that takes into account the ship's swaying characteristics, a photovoltaic panel module that incorporates dust accumulation characteristics, an oscilloscope module that reads voltage and current characteristics, and a neural network prediction module.

[0062] A neural network-based method for predicting photovoltaic dust accumulation based on ship sway characteristics includes the following steps:

[0063] S1: A simulation model is established by incorporating ship swaying characteristics into the traditional photovoltaic array model, and sample data is obtained.

[0064] Based on formulas (1)-(8), a solar radiation energy simulation model considering the ship's swaying characteristics is built using Simulink. The equivalent light intensity is given by the following formula, where the direct radiation energy I is obtained by considering the angle i between the incident ray and the normal of the tilted photovoltaic panel. n Scattered radiation energy I h and reflected radiation energy I g The calculation formula for solar radiation energy I is constructed. t The simulation model uses angle θ to mathematically describe the ship's rolling characteristics:

[0065] E′=η sc E(1)

[0066]

[0067] cosi=cosβsinh+sinβcoshcos(α-γ)(4)

[0068] I n =ξ0I sc P m sinh(5)

[0069]

[0070] I g =ρ(I n +I h (7)

[0071]

[0072] In the formula: I PH For photocurrent, I O1 I O2 For V D1 V D2 The reverse saturation current, R S The resistors are connected in series, and a1 and a2 are V. D1 V D2 The ideal factor, N SV represents the number of cells connected in series in the module. T1 V T2 R is the thermal voltage of the solar cell. SH For resistors connected in parallel, η sc The shading coefficient marked on the dust accumulation shadow of the photovoltaic panel is E, the original light intensity is β, the tilt angle is h, the solar altitude angle is α, the solar azimuth angle is γ, the slope azimuth angle is ξ0, the Earth orbit eccentricity correction factor is I. sc Where is the solar constant, P is the atmospheric transparency coefficient, m is the atmospheric optical mass, and θ is the atmospheric constant. max Let T be the maximum angle of ship sway, θ0 be the ship sway period, and θ0 be the initial phase. S2: Determine the input features of the neural network. Considering that dust accumulation on photovoltaic panels directly affects the incident light intensity of the photovoltaic module, short-circuit current, open-circuit voltage, maximum power point voltage, and maximum power point current are ultimately selected as the input features of the neural network. The type of dust accumulation is determined and labeled as the output. A dataset is constructed using the input and output features. The structure diagram of this invention for photovoltaic dust accumulation prediction based on ship sway characteristics is shown below. Figure 2 As shown.

[0073] S3: Initialize the neural network, determine the number of nodes in each layer of the neural network, and initialize the weights and thresholds of the neural network;

[0074] The structure diagram of a 5-input, 1-output CNN-LSTM-Attention neural network prediction model is constructed using the collected sample data, as shown below. Figure 3 As shown. Open-circuit voltage, short-circuit current, maximum power point voltage, maximum power point current, and ship roll angle are used as the input layer of the neural network with 4 nodes, while the type of ash accumulation is used as the output layer of the neural network with 1 node.

[0075] A CNN-LSTM-Attention photovoltaic dust accumulation prediction model was established, combining CNN, which has the ability to extract data features, with LSTM, which performs well in processing time series data. The attention mechanism is used to allocate probability weights to make up for the shortcoming of CNN, which can only capture local information. The prediction accuracy of the model was verified.

[0076] The LSTM dropout mechanism process includes input dropout, LSTM cell computation, loop dropout, and output.

[0077] Before the input data enters the LSTM unit, Dropout is applied to the input data. Let's assume the input data is x. t Dropout will randomly select x t Some elements in the array are set to 0.

[0078] x t = Dropout(xt ,rate)(10)

[0079] Here, rate is the proportion of Dropout.

[0080] Use the input x after Dropout t 'Calculate each gate state of the LSTM.'

[0081] In the hidden state h t Before passing it to the next time step, for h t Use Dropout.

[0082] h t = Dropout(h t ,rate)(11)

[0083] The final output is the hidden state h after Dropout. t '.

[0084] The key to LSTM is its cell structure, with each cell containing four main parts: a cell state and three "gates"—the forget gate, the input gate, and the output gate. The forget gate's role is to determine, through a sigmoid activation function, which information needs to be retained from the cell state and which needs to be forgotten. Specifically, the forget gate receives the hidden state from the previous time step (part of the memory cell) and the input from the current time step, outputting a value between 0 and 1, representing the degree to which information at the corresponding position is retained. 0 represents complete forgetting, and 1 represents complete retention. The mathematical expression for the forget gate is as follows:

[0085] f t =σ|W f ·[H t-1 ,x t ]+b f | (12)

[0086] Among them, f t Forget gate; σ represents the Sigmoid function; W f and b f These are the forget gate weights and bias coefficients, respectively; H t-1 The unit output at time t-1; x t Input the information at time t.

[0087] The input gate is a key component in LSTM, its main function being to control the input of new information, i.e., to determine which parts of the cell state are updated. The input gate allows the network to selectively update the cell state to adapt to the currently input information. The mathematical expression for the input gate is as follows:

[0088] i t =σ|Wt ·[H t-1 ,x t ]+b t | (13)

[0089]

[0090] Where: i t The input gate determines how many new candidate values ​​are generated. It will be added to the cellular state. It is a new candidate value for the current time step, which may be added to the cell state, C. t It represents the cell state at the current time step, W. C It is the weight matrix of the candidate values, b C W is the bias term for the candidate values. t It is the weight matrix of the input gate, b t is the bias term of the input gate, and tanh is the hyperbolic tangent activation function, which compresses any real value to the interval [-1, 1].

[0091] The output gate is another key component in LSTM, its main function being to determine which information from the current hidden state will be output to the next layer of the network or used for the final prediction. The output gate uses a sigmoid activation function to control which parts of the hidden state are activated. The mathematical expression for the output gate is as follows:

[0092] O t =σ|W O ·[H t-1 ,x t ]+b O |(16)

[0093] H t =O t ×tanh(C t (17)

[0094] Among them: O t It is the output of the output gate, W O It is the weight matrix of the output gate, b O It is the bias term of the output gate.

[0095] Adding an attention mechanism after the LSTM network can capture the information features output by the LSTM layers in the prediction model, allocate weight ratios, extract key information, and improve the accuracy of the photovoltaic dust accumulation prediction model. The hidden layer outputs of the LSTM are used as input to the attention mechanism. These outputs first pass through a fully connected layer whose parameters are learned during training to capture features at different time steps. Subsequently, the outputs of these fully connected layers are normalized using a softmax function, generating a set of weights that reveal the relative importance of the hidden states at each time step to the final prediction. Through this mechanism, the model can allocate more attention to those key time steps during prediction, thereby improving the overall accuracy and efficiency of the prediction. This dynamic weight allocation process is adaptive, ensuring that the model can identify and focus on the most informative parts in complex sequence data. The attention mechanism calculation process is as follows:

[0096] S i =tanh(WH i +b i (18)

[0097] α i =softmax(S i (19)

[0098]

[0099] Where: α i The weights, H, are obtained by applying the softmax function to the attention scores. i S is the output of the last LSTM hidden layer. i The score output for each hidden layer, C i b is the value after weighted average summation. t It is a bias term.

[0100] After passing through the Attention mechanism, the computation proceeds to the fully connected layer, where the output C of the Attention mechanism is input.

[0101] Z = W·C + b(21)

[0102] Where W is the weight matrix, b is the bias term, and Z is the output of the fully connected layer.

[0103] S4: Based on the characteristics of ship swaying, optimize the selection of characteristic quantities. The selection of ship swaying characteristic quantities is determined by parameters such as solar azimuth angle, solar altitude angle, and ship swaying period to determine the solar radiation intensity received by the photovoltaic module surface.

[0104] S5: Train the model, calculate the percentage error of photovoltaic dust accumulation prediction and compare the predicted value with the actual value;

[0105] like Figure 4 The predictive performance of the model is quantitatively analyzed using the coefficient of determination (R-squared), root mean square error (RMSE), mean squared error (MSE), and ratio of standard deviation of the prediction set (PRD). The predictive performance is observed by comparing the prediction results of the test set.

[0106] Due to various uncertainties in the prediction process of photovoltaic dust accumulation, the prediction results contain a certain degree of error. The performance of different prediction algorithms can be assessed through comparative analysis using relevant evaluation indicators. Generally, the coefficient of determination (R-squared), root mean square error (RMSE), mean squared error (MSE), and ratio of standard deviations of the prediction set (PRD) are used to quantitatively analyze the prediction performance of the model. 2 The range is [0-1], when R 2 When R = 1, it perfectly explains all data variations. 2 The smaller the value, the worse the model's prediction performance. Root Mean Square Error (RMSE) and Relative Mean Error (MSE) reflect the relative deviation between predicted and actual values; the square of the error value better reflects the role of larger errors in the overall error assessment. Predictive performance is observed by comparing prediction results on the test set. The Prediction Set Standard Deviation Ratio (PRD) assesses the model's uncertainty or accuracy on new data points, especially considering potential variability during prediction; a higher PRD value indicates better model performance.

[0107]

[0108] Among them: SS res Let y represent the sum of squared residuals. i These are observed values. It is a predicted value, SS tot Represents the total sum of squares. It is the average of the observed values.

[0109]

[0110] Where: y true This represents the test set, i.e., the actual values, y. pre This represents the value predicted by the model, where n is the number of samples in the test set;

[0111]

[0112] Where: RRMSEP represents the relative squared error for each individual model;

[0113] The sample data obtained from the photovoltaic dust accumulation simulation model considering ship sway characteristics were divided into training and testing sets at an 80% / 20% ratio. Initial parameters were set, and the Adam gradient descent algorithm was used with a stride of [1,1] for the convolutional layers, 64 channels, and a maximum number of iterations t. max =100, initial learning rate 0.01, learning rate decay factor 0.5. A CNN-LSTM-Attention neural network is used to train the training set data. After training, predictions are made using the input test set, and the predicted dust accumulation type values ​​are compared and analyzed with the actual values ​​on the output test set based on corresponding error evaluation metrics.

[0114] like Figure 4 As shown, the test set prediction results obtained after training with the CNN-LSTM-Attention neural network are compared, and the determination coefficient R is shown for each indicator. 2 The mean square error (RMSE) is 0.97526, the relative mean error (MSE) is 0.2, the prediction set standard deviation ratio (PRD) is 0.128, and the prediction set standard deviation ratio (PRD) is 8.1431.

[0115] The predicted regression plot is as follows Figure 5 As shown, by observing the predicted regression plot, we can see that the data points are close to the regression line, the data points fit the fitted line well, and the prediction model performs well.

[0116] The comparison chart of prediction results between the test set and the prediction set is shown below. Figure 6 As shown, the accuracy of the training results on the test set is 98.7351%, and the accuracy of the results on the training set is 98.6973%. It can be seen that the accuracy of both the test set and the training set is high, and the model performs well.

[0117] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0118] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A neural network-based soot deposition prediction method for photovoltaic systems based on ship sway characteristics, characterized by, The method comprises the following steps: S1: establishing a photovoltaic array model and obtaining sample data; the photovoltaic array model is specifically: In the formula: I PH For photocurrent, I O1 I O2 For V D1 V D2 The reverse saturation current, R S The resistors are connected in series, and a1 and a2 are V. D1 V D2 The ideal factor, N S V represents the number of cells connected in series in the module. T1 V T2 R is the thermal voltage of the solar cell. SH The resistors are connected in parallel; where k is the Boltzmann constant, k = 1.38 x 10 -23 J / K; T is the temperature of the solar cell; q is the electronic charge, q = 1.602 x 10 -19 C; S2: determining input features of the neural network, selecting short-circuit current, open-circuit voltage, maximum power point voltage and maximum power point current as input feature quantities of the neural network, determining the type of dust accumulation, and pasting a label as output, so as to construct a data set with the input and output feature quantities; S3: initializing the neural network, determining the number of nodes of each layer of the neural network, and initializing the weight and threshold value of the neural network; S4: optimizing the selection of the input feature quantities in step S2 based on the characteristics of ship rolling, wherein the ship rolling characteristics include a solar azimuth angle, a solar elevation angle and a ship rolling period parameter; S5: training the model, calculating a photovoltaic dust accumulation prediction percentage error and comparing a predicted value with an actual value.

2. The neural network-based soot deposition prediction method for PV panels based on ship's roll characteristics according to claim 1, characterized in that, In step S2, by establishing a photovoltaic array model, I-V and P-V curves of the photovoltaic module are observed, four feature quantities of the photovoltaic module surface under different dust accumulation conditions, i.e., short-circuit current, open-circuit voltage, maximum power point voltage and maximum power point current, are obtained, so as to determine the input feature quantities of the neural network, and the equivalent light intensity affected by dust accumulation is: E' = η sc E where: η sc is the shading coefficient identified on the soiling shadow of the photovoltaic panel, E is the original light intensity.

3. The neural network-based soot deposition prediction method for PV panels based on ship's roll characteristics according to claim 1, characterized in that, In step S4, based on the characteristics of marine ship rolling, a ship rolling characteristic solar radiation energy simulation model is established, and the ship rolling characteristic solar radiation energy simulation model is: cosi=cosβsinh+sinβcoshcos(α-γ) I n = ξ0I sc P m sinh I g = p(I n + I h ) where β is the tilt angle, h is the solar elevation angle, a is the solar azimuth angle, γ is the slope azimuth angle, ξ0is the earth orbit eccentricity correction factor, I sc is the solar constant, P is the atmospheric transparency factor, m is the atmospheric optical mass, θ max is the maximum angle of ship sway, T is the ship sway period, θ0is the initial phase.

4. The neural network-based soot deposition prediction method for PV panels based on ship's roll characteristics according to claim 1, characterized in that: The step S5 specifically comprises the following steps: S51: processing input data through a one-dimensional convolution layer, wherein the one-dimensional convolution layer comprises a convolution layer and a pooling layer, the convolution layer is used for extracting local features of data, and the pooling layer is used for reducing feature dimensions; S52: processed data is sent to an LSTM layer, the LSTM layer is used for capturing long-term dependencies and complex time dynamics in data, and a first dropout layer is introduced; S53: an attention mechanism is added to the model, the attention layer is used for weighting the output of the LSTM layer, so that the model can focus on the most critical part of the sequence; S54: data processed through the attention mechanism is sent to a first full connection layer, the first full connection layer aggregates and maps information to an intermediate representation space, through application of a ReLU activation function, the model can enhance the ability to capture nonlinear relationships, and further enhance the prediction accuracy; S55: in the last data processing stage, a second dropout layer and a second full connection layer are introduced to the model, and the second full connection layer maps processed features to a prediction output space; S46: the second full connection layer maps processed information to a prediction output space, and the function of this layer is to convert all learning achievements of the model into specific prediction values, thereby providing support for the final decision.

5. The neural network-based soot deposition prediction method for PV panels based on ship's roll characteristics according to claim 4, characterized in that, In step S52, the unit structure of the LSTM layer includes four main parts: cell state, forget gate, input gate and output gate. The forget gate determines the information to be retained from the cell state through a sigmoid activation function, receives the hidden state of the previous time (part of the memory cell) and the input of the current time, and outputs a value between 0 and 1, indicating the degree of information retention at the corresponding position; 0 represents complete forgetting, and 1 represents complete retention. The mathematical expression of the forget gate is as follows: f t = σ | W f · [H t-1 , x t ] + b f | wherein f t is the forget gate, σ denotes the Sigmoid function, W f and b f are the forget gate weight and bias coefficient, respectively, H t-1 is the cell output at time t-1, and x t is the information input at time t. The input gate is used to control the input of new information, and the input gate allows the network to selectively update the cell state to adapt to the current input information. The mathematical expression of the input gate is as follows: i t = σ | W t · [H t-1 , x t ] + b t | where: i t is the input gate, which determines how many new candidate values will be added to the cell state, is the new candidate value at the current time step, which can be added to the cell state, C t is the cell state at the current time step, W C is the weight matrix for the candidate value, b C is the bias term for the candidate value, W t is the weight matrix for the input gate, b t is the bias term for the input gate, tanh is the hyperbolic tangent activation function, which compresses any real value to the interval [-1, 1]; The output gate is used to determine whether the information in the hidden state at the current time will be output to the next layer of the network or used for final prediction. The output gate controls which part of the hidden state will be activated through a sigmoid activation function. The mathematical expression of the output gate is as follows: O t = σ | W O · [H t-1 , x t ]+b O | H t =O t x tanh(C t ) where: O t is the output of the output gate, W O is the weight matrix of the output gate, b O is the bias term of the output gate.

6. The neural network-based soot deposition prediction method for PV panels based on ship's roll characteristics according to claim 4, characterized in that, In step S53, the attention mechanism model is used to capture the information features output by the LSTM layer in the prediction model, divide the weight proportion, extract key information, and improve the prediction model accuracy. The hidden layer output of the LSTM is used as the input of the attention mechanism. The hidden layer output of the LSTM is first passed through a first fully connected layer. The output of the first fully connected layer is normalized by a softmax function to produce a set of weights. The weights are used to show the relative importance of each time step hidden state to the final prediction. The attention mechanism calculation process is as follows: S i = tanh(WH i + b i ) a i = softmax(S i ) where: a i is the weight obtained by applying a softmax function to the attention scores, H i is the output of the last LSTM hidden layer, S i is the score of each hidden layer output, C i is the weighted average summed value, b t is the bias term.

Images