Floating photovoltaic power generation power prediction method based on optimized Brown bear memory network

Abstract

The invention discloses a floating photovoltaic power generation power prediction method based on an optimized Brown bear memory network, and relates to the technical field of photovoltaic power generation prediction, and the method comprises the steps: collecting multi-source heterogeneous data through a sensor network; preprocessing the collected data based on the improved mathematical model, and screening characteristic variables sensitive to a prediction target; constructing an initial LSTM model, and optimizing the hyper-parameters by adopting a Brown-bear optimization algorithm to obtain an optimal LSTM model; and carrying out verification and performance evaluation on the model by adopting a multi-dimensional statistical index, and outputting a photovoltaic power generation power prediction result. According to the method, original data are preprocessed by improving a mathematical model, abnormal data caused by environmental interference, equipment faults and other factors are effectively eliminated, the reliability of the data is improved from the source, the Brown bear optimization algorithm is integrated into the construction process of the LSTM model, a traditional manual trial and error adjustment and optimization mode depending on expert experience is replaced, and the accuracy of the LSTM model is improved. And the model prediction performance is improved through the obtained optimal hyper-parameter combination of the LSTM model.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically a floating photovoltaic power generation prediction method based on an optimized Brown Bear memory network. Background Technology

[0002] With the rapid development of the world economy, global energy demand has surged, leading to a series of serious problems such as the depletion of fossil fuel resources, increased environmental pollution, ozone layer depletion, and the global greenhouse effect. To address these issues, the international community is actively promoting the transformation of the energy structure, focusing on sustainable and environmentally friendly renewable energy technologies. Solar energy, with its unlimited reserves, wide distribution, and clean, pollution-free nature, has become a core component of the global energy transition strategy, and extensive research is dedicated to improving the efficiency of solar energy conversion and utilization.

[0003] Among various solar energy utilization technologies, solar photovoltaic (PV) technology (which directly converts solar radiation energy into electrical energy through the photovoltaic effect) is the mainstream method for large-scale commercial development of green electricity. However, in practical promotion, traditional ground-mounted photovoltaic (SGPV) systems face a major bottleneck: the system requires a large amount of land resources, creating a significant contradiction with densely populated areas and high land prices. To solve this contradiction, floating solar photovoltaic (SFPV) systems have emerged. This system installs photovoltaic arrays on the surface of water bodies such as reservoirs, lakes, and coastal waters, thereby effectively utilizing vast water areas without occupying additional land, reducing the operating temperature of photovoltaic panels, and suppressing water evaporation, thus demonstrating enormous market potential and technological prospects.

[0004] Despite the significant advantages of floating solar photovoltaic systems, accurate performance prediction and efficient operation and maintenance still face technical challenges. 1. SFPV systems operate in corrosive marine environments with high humidity and high salinity. Their operating characteristics are affected by unique factors such as water surface reflection, wave load, and tidal changes, which are very different from SGPV systems. Traditional prediction models are difficult to apply. Furthermore, there is a highly nonlinear and strongly coupled complex relationship between the photovoltaic module temperature and power generation of SFPV systems and environmental meteorological conditions (solar radiation, ambient temperature, and wind speed), which is difficult to accurately describe using traditional physical models or simple mathematical models. 2. Existing prediction methods mostly rely on simplifying assumptions. When dealing with the real-time nature and complexity of SFPV systems, they suffer from insufficient prediction accuracy and weak generalization ability. Furthermore, hyperparameter tuning depends on expert experience and manual trial and error, resulting in low efficiency and difficulty in achieving optimal performance.

[0005] Based on this, we now present a floating photovoltaic power generation prediction method based on an optimized Brown Bear memory network, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the power generation of floating photovoltaic systems based on an optimized Brown Bear memory network, so as to solve the challenges of accurate performance prediction and efficient operation and maintenance of floating solar photovoltaic systems in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The floating photovoltaic power generation prediction method based on optimized Brown Bear memory networks includes the following steps: Step S1: Synchronously collect multi-source heterogeneous data from the floating photovoltaic power station through a sensor network; Step S2: Based on the improved mathematical model, preprocess the collected data, screen the feature variables that are sensitive to the prediction target, remove outliers and retain the valid data under normal operating conditions; Step S3: Construct an initial LSTM model and use the Brown Bear optimization algorithm to optimize the hyperparameters of the initial LSTM model to obtain the optimal LSTM model based on the Brown Bear optimization algorithm. Step S4: Use multidimensional statistical indicators to verify and evaluate the model, and output the photovoltaic power generation prediction results.

[0008] Preferably, step S2 specifically includes: Step S2-1: Based on the single diode model, construct an inference model for the output voltage and current of the photovoltaic array. Introduce the actual operating irradiance, component temperature, number of series modules, number of parallel modules, temperature coefficient, aging loss rate and operating years as input variables to obtain the final voltage and current inference model. Step S2-2: Collect DC voltage and current at each moment in real time, and calculate the ratio of measured data to model-inferred values; Step S2-3: Retain the data points with ratios in the interval [0.9, 1.1] and assign the data points in the interval to the normal operating status of the photovoltaic system.

[0009] Preferably, the expression for the voltage and current inference model includes: ; ; in, The calculated value of the photovoltaic array output voltage is inferred based on a single diode model. The calculated value of the photovoltaic array output current is inferred based on the single-diode model. and These are mathematical functions for voltage calculation and current calculation, respectively. Irradiance, For the temperature of photovoltaic modules, For the number of modules in series, For the number of parallel modules, The temperature coefficient of voltage, The aging loss rate of voltage. The number of years a photovoltaic module has been in operation. The temperature coefficient of the current. This represents the aging loss rate of the current.

[0010] Preferably, the expression for the ratio of the measured data to the model-inferred value includes: ; ; in, This is the ratio of the measured DC voltage to the voltage inferred from the model. This is the measured DC voltage value. The calculated value of the photovoltaic array output voltage is inferred based on a single diode model. This is the ratio of the measured DC current to the model-inferred current. This is the measured DC current value. This is the calculated value of the photovoltaic array output current inferred from the single diode model.

[0011] Preferably, step S3 specifically includes: building a long short-term memory network as the core predictor, using a gating mechanism to capture long-term temporal dependencies and nonlinear features in meteorological and power generation data, integrating the Brown Bear optimization algorithm as the optimization engine into the LSTM model construction process, searching for the optimal hyperparameter combination of the LSTM model, training the LSTM model with the optimal hyperparameter combination, and obtaining the optimal LSTM model.

[0012] Preferably, the specific operations of the Brown Bear optimization algorithm include: The set of hyperparameters to be optimized in the LSTM model is mapped to Brown Bear positions in the search space of the Brown Bear optimization algorithm. Each position vector corresponds to a set of hyperparameter configurations. A fitness function oriented towards prediction accuracy is defined, and the fitness function uses the reciprocal of the mean square error as the core calculation index. The Brown Bear Optimization Algorithm simulates the marking and sniffing behavior of a Brown Bear, and performs iterative search in the hyperparameter space to achieve global optimization. The global optimization includes population initialization, development phase and exploration phase. The termination condition of the iterative search is reaching the maximum number of iterations or the fitness value converges, and the algorithm outputs the hyperparameter combination with the highest fitness value in history. In the population initialization phase, an initial Brown Bear population is randomly generated within a preset hyperparameter range. In the development phase, the Brown Bear's walking, stepping, and twisting movements are simulated. The walking, stepping, and twisting movements correspond to extensive exploration in the early stage of iteration, approaching the optimal solution in the middle stage, and fine local search in the later stage, respectively.

[0013] Preferably, the development phase specifically includes: The walking action is extensively explored in the first third of iterations to coarsely scan the region and avoid prematurely getting trapped in local optima. The action expression is: ,in, This is Brown Bear's current location. This is Brown Bear's original location. The random step size coefficient is used to introduce randomness for the search. The current iteration ratio and satisfying the formula ; The stepping action moves towards the current optimal solution during the middle 1 / 3 of the iterations, driving all Brown Bears to learn from the better Brown Bears and converge to the known optimal region. The action expression is: ,in, The updated location for Brown Bear. A factor to control the magnitude of movement. This represents the position of the historical best solution in the current population. This represents the position of the worst solution in the current population. The ankle-twisting action undergoes a more refined local search in the last 1 / 3 of iterations. This search focuses on the vicinity of the optimal solution to ultimately pinpoint the precise global optimum. The action expression is as follows: ,in, The updated location for Brown Bear. The value is random and the range of values ​​is , The first in the current population The position of the historical best solution in each dimension. The first in the current population The position of the worst solution in each dimension; The expression for the exploration phase is: ,in, For the first Brown Bear in the first The new position updated in each dimension The parameter values ​​prior to the start of this exploration phase, i.e., the first... Brown Bear in the first The original position in each dimension It is a random number. and The first one randomly selected from the population Only with the first Brown Bear's current location. This is the fitness function.

[0014] Preferably, the expression for the fitness function is: ,in, For fitness value, Mean square error, It is a very small constant used to avoid division by zero errors.

[0015] Preferably, the multi-source heterogeneous data in step S1 includes environmental data and system operation data. The environmental data includes total solar radiation intensity, ambient temperature, relative humidity and wind speed. The system operation data includes DC voltage of photovoltaic modules, DC current, front temperature of modules and back sheet temperature.

[0016] Preferably, the multidimensional statistical indicators in step S4 include the coefficient of determination, root mean square error, and mean absolute error.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention establishes a floating photovoltaic power generation prediction method based on an optimized Brown Bear memory network. It preprocesses the raw data using an improved mathematical model based on a single-diode model, and effectively eliminates abnormal data caused by environmental interference, equipment failure, and other factors by calculating the ratio of measured values ​​to model-inferred values ​​and filtering data within a reasonable range. This improves data reliability from the source. Furthermore, this invention integrates the Brown Bear optimization algorithm into the LSTM model construction process, replacing the traditional manual trial-and-error tuning method that relies on expert experience. By simulating walking, stepping, and ankle twisting movements, it finds the optimal hyperparameter combination of the LSTM model in the hyperparameter search space, thereby improving the model's predictive performance. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of the floating photovoltaic power generation prediction method of the present invention.

[0019] Figure 2 This is a schematic flowchart of the floating photovoltaic power generation prediction method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0021] With the rapid development of artificial intelligence (AI) technology, AI models, especially deep learning models, have a powerful ability to process complex nonlinear data relationships. They can then automatically learn from massive amounts of historical data and uncover deep patterns that affect system performance. Time series prediction capabilities, such as Long Short-Term Memory (LSTM) networks, have unique gating mechanisms and are adept at processing time-series related data. They can effectively capture long-term dependencies in meteorological and power generation data, making them suitable for short-term and ultra-short-term forecasting of photovoltaic power generation. This provides a new solution to the problem of accurate forecasting of SFPV systems. In this embodiment, as Figures 1-2 As shown, this invention proposes a floating photovoltaic power generation prediction method based on an optimized Brown Bear Memory Network. It utilizes an advanced AI prediction model (LSTM) combined with the Brown Bear Optimization Algorithm (BBOA) to construct a novel hybrid intelligent prediction framework to solve the challenge of accurate performance prediction in the floating photovoltaic application scenario. The Brown Bear Optimization Algorithm is a swarm intelligent optimization algorithm that simulates the behavior of Brown bears in their natural environment, such as marking and sniffing, to find food resources. This algorithm simulates the cooperative search behavior of a Brown bear population, efficiently finding the optimal solution in a complex solution space. It then maps the parameters to be solved in the optimization problem to the position of the Brown bear in the search space. By defining a series of update rules inspired by its behavior (such as walking, stepping, and twisting movements in the development phase, and random sniffing in the exploration phase), it balances global exploration and local development capabilities, thereby effectively finding the globally optimal solution or a near-optimal solution. Specifically, the following steps are included: Step S1: Synchronously collect multi-source heterogeneous data from the floating photovoltaic power station through a sensor network; Step S2: Based on the improved mathematical model, the collected data is preprocessed, the feature variables that are sensitive to the prediction target are screened, outliers are removed and valid data under normal operating conditions are retained. The feature variables that are sensitive to the prediction target are the data points with ratios in the interval [0.9, 1.1]. These data points are the input variables of the optimal LSTM model. The improvement of the improved mathematical model lies in the enhancement of the traditional single diode physical model. The traditional model usually uses irradiance and temperature as basic inputs, while this invention introduces the actual operating structural parameters of the photovoltaic array (number of series modules and number of parallel modules) and key loss factors, which can be adapted to photovoltaic arrays with different electrical connection methods and directly infer the total output voltage and current of the entire array. Step S3: Construct an initial LSTM model and use the Brown Bear optimization algorithm to optimize the hyperparameters of the initial LSTM model to obtain the optimal LSTM model based on the Brown Bear optimization algorithm. The input of the optimal LSTM model is the measured environmental data and system parameters, and the output is the predicted photovoltaic power generation. Step S4: Use multidimensional statistical indicators to verify and evaluate the model, and output the photovoltaic power generation prediction results; In this embodiment, the present invention uses a data preprocessing method based on an improved mathematical model, mainly reflected in the standardized screening process of the original photovoltaic system monitoring data. This method effectively improves the prediction accuracy of the subsequent prediction model by combining the physical model with the measured data. The present invention also proposes an LSTM model based on the Brown Bear optimization algorithm, which is deeply embedded into the construction process of the LSTM model. The Brown Bear optimization algorithm is applied to solve the hyperparameter optimization problem in photovoltaic power generation prediction, especially floating photovoltaic prediction, making it an inseparable optimization engine of the model. This provides an automated general framework for building high-performance AI prediction models. The Brown Bear optimization algorithm simulates the marking and sniffing behavior of Brown Bear. Its unique search mechanism (divided into three stages: walking, stepping, and twisting) can effectively balance global exploration and local development capabilities, avoiding the defects of traditional methods that are prone to getting trapped in local optima. This makes it more likely to find the global optimal parameter configuration of the LSTM model, so that the LSTM model shows stronger stability and generalization ability when facing complex scenarios such as sudden weather changes and noisy data. Among them, such as Figure 2 As shown, step S2 specifically includes: Step S2-1: Based on the single-diode model, construct an inference model for the output voltage and current of the photovoltaic array. Introduce the actual operating irradiance, module temperature, number of series modules, number of parallel modules, temperature coefficient, aging loss rate, and operating years as input variables to obtain the final voltage and current inference model. The data preprocessing method based on the improved mathematical model does not directly use the measured voltage and current values ​​for model training. Instead, it uses an improved mathematical model, namely the single-diode model with input variables, to calculate the inferred voltage and current values ​​under specific irradiance and module temperature. The improvement of this mathematical model lies in incorporating key loss factors in actual engineering for enhanced operation. Key loss factors include temperature coefficient, array configuration, aging loss, etc. The temperature coefficient is used to correct the change of voltage and current with temperature. The array configuration considers the number of photovoltaic modules connected in series and parallel. The aging loss is based on the operating years of the photovoltaic modules to simulate the normal degradation of their performance. Step S2-2: Collect DC voltage and current at each moment in real time, and calculate the ratio of measured data to model-inferred values; Step S2-3: Retain data points with ratios in the interval [0.9, 1.1] and assign the data points in the interval to the normal operating status of the photovoltaic system. This interval is obtained based on historical data and testing experience, referencing the national standard for the operating error range of the photovoltaic system and previous experimental data. If the error is too large, it is considered invalid data. Specifically, the expressions for the voltage and current inference model include: ; ; in, The calculated value of the photovoltaic array output voltage is inferred based on a single diode model. The calculated value of the photovoltaic array output current is inferred based on the single-diode model. and These are mathematical functions for voltage calculation and current calculation, respectively. Irradiance, For the temperature of photovoltaic modules, For the number of modules in series, For the number of parallel modules, The temperature coefficient of voltage, The aging loss rate of voltage. The number of years a photovoltaic module has been in operation. The temperature coefficient of the current. The aging loss rate of the current; Specifically, the expression for the ratio of measured data to model-inferred values ​​includes: ; ; in, This is the ratio of the measured DC voltage to the voltage inferred from the model. This is the measured DC voltage value. The calculated value of the photovoltaic array output voltage is inferred based on a single diode model. This is the ratio of the measured DC current to the model-inferred current. This is the measured DC current value. The calculated value of the photovoltaic array output current is inferred based on the single diode model; In this embodiment, this step uses the calculation results of the mathematical model to standardize and filter the measured data, thereby eliminating outliers and retaining valid data under normal operating conditions; Among them, such as Figure 2 As shown, step S3 specifically includes: building a long short-term memory network as the core predictor, using a gating mechanism to capture long-term temporal dependencies and nonlinear features in meteorological and power generation data, integrating the Brown Bear optimization algorithm as the optimization engine into the LSTM model construction process, searching for the optimal hyperparameter combination of the LSTM model, training the LSTM model with the optimal hyperparameter combination, and obtaining the optimal LSTM model. Specifically, the gating mechanism includes an input gate, a forget gate, and an output gate, and the network structure is designed as follows: Forget Gate: Determines which historical information to discard from the cell state. It outputs a value between 0 and 1 using the sigmoid function, which is then used to calculate the value based on the previous hidden state and the current input. The expression is: ; Input Gate: Determines which new information will be stored in the cell state. It consists of a sigmoid layer to decide which values ​​to update and a tanh layer to create new candidate value vectors. The expression is: Candidate value ; Cell state update: Multiply the old state by the result of the forget gate, add the result of the input gate and the candidate value, and form the new cell state, expressed as: ; Output Gate: Based on the updated cell state, it determines the value of the next hidden state (i.e., the output at this time step). The expression is: ; The final output is ; Specifically, the Brown Bear optimization algorithm includes the following operations: The set of hyperparameters to be optimized in the LSTM model is mapped to Brown Bear positions in the search space of the Brown Bear optimization algorithm. Each position vector corresponds to a set of hyperparameter configurations. An example is: Brown Bear positions. This means that the number of hidden layer neurons is 120, the learning rate is 0.01, the training batch size is 32, and a fitness function oriented towards prediction accuracy is defined, with the reciprocal of the mean squared error as the core calculation index. The Brown Bear Optimization Algorithm simulates the marking and sniffing behavior of a Brown Bear, and performs iterative search in the hyperparameter space to achieve global optimization. Global optimization includes population initialization, development phase and exploration phase. The termination condition of the iterative search is reaching the maximum number of iterations or the fitness value converges. It outputs the hyperparameter combination with the highest fitness value in history. During the population initialization phase, within the preset hyperparameter range, the parameters of the Brown Bear optimization algorithm, such as population size, maximum number of iterations, search step size coefficient, and search space boundary parameters, are set, and an initial Brown Bear population is randomly generated. During the development phase, the walking, stepping, and twisting movements of the Brown Bear are simulated. The walking, stepping, and twisting movements correspond to extensive exploration in the early stage of iteration, approaching the optimal solution in the middle stage, and fine local search in the later stage, respectively. The hyperparameters include the number of hidden layer neurons, learning rate, training batch size, and number of iterations. The preset range for the number of hidden layer neurons is [50, 200], the learning rate adopts a logarithmic scale and the preset range is [0.001, 0.1], the preset range for the training batch size is [16, 128], and the preset range for the number of iterations is [100, 500]. Population Size refers to the number of Brown Bear individuals during the optimization process. This parameter affects the global search capability and computational cost, and its value ranges from 30 to 100. In this embodiment, 50 is preferred. Maximum Iterations refers to the maximum number of search rounds before the Brown Bear optimization algorithm terminates. This parameter affects the convergence of the optimization, and its value ranges from 100 to 500. In this embodiment, 200 is preferred. Step Size Coefficient is used to control the search step size of Brown Bear in the "walking", "stepping", and "twisting" actions. Its value is usually related to the dimension of the search space, and its value ranges from 0.1 to 1.0. In this embodiment, 0.5 is preferred. Search Space Boundaries include defining the value range of each LSTM hyperparameter to be optimized, the number of neurons in the LSTM hidden layer, the learning rate, etc. Specifically, the development phase includes: The walking action is explored extensively in the first third of iterations to coarsely scan the region and avoid getting trapped in local optima too early. The action expression is: ,in, This is Brown Bear's current location. This is Brown Bear's original location. The random step size coefficient is used to introduce randomness for the search. The current iteration ratio and satisfying the formula In the initial 1 / 3 iterations, The value gradually decreases from 1, and the walking action indicates that in the early stages of the search, Brown Bear uses a larger stride (affected by...). (Impact) Conduct a relatively extensive exploration around its current location, which is an exploration strategy from large to small; The stepping action moves towards the current optimal solution during the middle third of the iterations, driving all Brown Bears to learn from the better Brown Bears and converge to the known optimal region. The action expression is: ,in, The updated location for Brown Bear. The factor controlling the movement amplitude is not a fixed value that needs to be preset, but a value that is dynamically calculated during the algorithm's execution. This represents the position of the historical best solution in the current population. The position of the worst solution in the current population is represented by the stepping action: In the middle of the search, the Brown Bear begins to move towards the known best region in the population. The expression defines the direction and distance of movement by calculating the "gap" between the best and worst solutions, which drives all the bears to learn from their excellent peers and quickly converge to the promising region. The ankle-twisting action performs a more refined local search in the last third of iterations, targeting the vicinity of the optimal solution to ultimately pinpoint the precise global optimum. The action expression is: ,in, The updated location for Brown Bear. The value is random and the range of values ​​is , It is usually set to a smaller random value. After extensive experimental verification, its most effective and classic empirical range is... Within this range, the algorithm typically maintains a good balance between global exploration and local exploitation capabilities. In this embodiment, 0.729 is preferred, as this value is a commonly used typical value in this field. The first in the current population The position of the historical best solution in each dimension. The first in the current population The position of the worst solution in each dimension is represented by the foot-twisting action: in the later stages of the search, the algorithm is already near the optimal solution. At this time, Brown Bear's action becomes very subtle ("twisting his foot"), performing a fine local search near the optimal solution to finally lock in the precise global optimal solution and avoid missing the true best point. This expression is not related to the position of Brown Bear mentioned above, but only to the number of iterations of the system. The optimal solution at this time is the value obtained by updating the previous step. The expression for the exploration phase is: ,in, For the first Brown Bear in the first The new position updated in each dimension The parameter values ​​prior to the start of this exploration phase, i.e., the first... Brown Bear in the first The original position in each dimension It is a random number. and The first one randomly selected from the population Only with the first Brown Bear's current location. The fitness function has a higher value, indicating a better location. Specifically, the expression for the fitness function is: ,in, For fitness value, Mean square error, The fitness function is a tiny constant used to avoid division by zero errors. It is prediction accuracy oriented and performs the following operations for each set of hyperparameters (each Brown Bear): initialize the LSTM model with the set of parameters, train the model with the training set, calculate the prediction performance of the model on the validation set, and use the reciprocal of the mean squared error (MSE) as the fitness value of the set of hyperparameters. Specifically, this invention automates and optimizes the hyperparameter tuning process through three stages: population initialization, development, and exploration, thereby effectively solving the current technical challenges in the field of photovoltaic power generation prediction. Among them, such as Figure 2 As shown, the multi-source heterogeneous data in step S1 includes environmental data and system operation data. The environmental data includes total solar radiation intensity, ambient temperature, relative humidity and wind speed. The system operation data includes DC voltage of photovoltaic modules, DC current, front temperature of modules and back sheet temperature. Among them, such as Figure 2 As shown, the multidimensional statistical indicators in step S4 include the coefficient of determination, root mean square error, and mean absolute error. In this embodiment, the optimal LSTM model, i.e., the optimized Brown Bear Memory Network prediction model, is trained using training data. Multidimensional statistical metrics are used to validate the model's performance on the test set, ensuring its prediction accuracy and generalization ability. To verify the superiority of the Brown Bear Optimization Algorithm (BBOA) in optimizing LSTM hyperparameters for floating photovoltaic prediction, a comparative experiment is designed: Experimental setup: The dataset consists of operational data from a floating photovoltaic power station over a specific time period, divided into training and test sets in an 8:2 ratio. The Brown Bear Optimization Algorithm (BBOA) of this invention is compared with Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Whale Optimization Algorithm (WOA). All comparison algorithms use the same LSTM network structure as the prediction model, are optimized within the same hyperparameter search space, and run on the same computing hardware and software platform. The multidimensional statistical indicators and convergence iterations of the optimal LSTM model obtained by each algorithm on the test set are recorded. The specific performance data are shown in Table 1 below. Table 1 - Comparison of Performance Data for Different Algorithms

[0022] As shown in Table 1, the Brown Bear Optimization Algorithm (BBOA) used in this invention exhibits significant advantages in the floating photovoltaic prediction task. The LSTM model optimized by BBOA has a high determination coefficient, a low mean absolute error, and a prediction accuracy that is significantly better than traditional optimization algorithms. Furthermore, the convergence speed of BBOA is much faster than some traditional optimization algorithms, demonstrating better search efficiency. This indicates that applying BBOA to the hyperparameter optimization of LSTM models can solve the floating photovoltaic prediction problem and improve the accuracy and reliability of the prediction model. In summary, this invention combines an improved mathematical model with an LSTM model based on the Brown Bear optimization algorithm. Under the verification of multidimensional statistical indicators, it effectively overcomes the problems of poor adaptability to the special environment of floating photovoltaics, model hyperparameter dependence on manual tuning, and insufficient prediction accuracy and generalization ability of existing technologies. Ultimately, it achieves high-precision prediction of floating photovoltaic power generation and has good application prospects.

[0023] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting floating photovoltaic power generation based on an optimized Brown Bear memory network, characterized in that, Specifically, the following steps are included: Step S1: Synchronously collect multi-source heterogeneous data from the floating photovoltaic power station through a sensor network; Step S2: Based on the improved mathematical model, preprocess the collected data, screen the feature variables that are sensitive to the prediction target, remove outliers and retain the valid data under normal operating conditions; Step S3: Construct an initial LSTM model and use the Brown Bear optimization algorithm to optimize the hyperparameters of the initial LSTM model to obtain the optimal LSTM model based on the Brown Bear optimization algorithm. Step S4: Use multidimensional statistical indicators to verify and evaluate the model, and output the photovoltaic power generation prediction results.

2. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 1, characterized in that, Step S2 specifically includes: Step S2-1: Based on the single diode model, construct an inference model for the output voltage and current of the photovoltaic array. Introduce the actual operating irradiance, component temperature, number of series modules, number of parallel modules, temperature coefficient, aging loss rate and operating years as input variables to obtain the final voltage and current inference model. Step S2-2: Collect DC voltage and current at each moment in real time, and calculate the ratio of measured data to model-inferred values; Step S2-3: Retain the data points with ratios in the interval [0.9, 1.1] and assign the data points in the interval to the normal operating status of the photovoltaic system.

3. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 2, characterized in that, The expressions for the voltage and current inference model include: ； ； in, The calculated value of the photovoltaic array output voltage is inferred based on a single diode model. The calculated value of the photovoltaic array output current is inferred based on the single-diode model. and These are mathematical functions for voltage calculation and current calculation, respectively. Irradiance, For the temperature of photovoltaic modules, For the number of modules in series, For the number of parallel modules, The temperature coefficient of voltage, The aging loss rate of voltage. The number of years a photovoltaic module has been in operation. The temperature coefficient of the current. This represents the aging loss rate of the current.

4. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 2, characterized in that, The expression for the ratio of the measured data to the model-inferred value includes: ； ； in, This is the ratio of the measured DC voltage to the voltage inferred from the model. This is the measured DC voltage value. The calculated value of the photovoltaic array output voltage is inferred based on a single diode model. This is the ratio of the measured DC current to the model-inferred current. This is the measured DC current value. This is the calculated value of the photovoltaic array output current inferred from the single diode model.

5. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 1, characterized in that, Step S3 specifically includes: building a long short-term memory network as the core predictor, using a gating mechanism to capture long-term temporal dependencies and nonlinear features in meteorological and power generation data, integrating the Brown Bear optimization algorithm as the optimization engine into the LSTM model construction process, searching for the optimal hyperparameter combination of the LSTM model, training the LSTM model with the optimal hyperparameter combination, and obtaining the optimal LSTM model.

6. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 5, characterized in that, The specific operations of the Brown Bear optimization algorithm include: The set of hyperparameters to be optimized in the LSTM model is mapped to Brown Bear positions in the search space of the Brown Bear optimization algorithm. Each position vector corresponds to a set of hyperparameter configurations. A fitness function oriented towards prediction accuracy is defined, and the fitness function uses the reciprocal of the mean square error as the core calculation index. The Brown Bear Optimization Algorithm simulates the marking and sniffing behavior of a Brown Bear, and performs iterative search in the hyperparameter space to achieve global optimization. The global optimization includes population initialization, development phase and exploration phase. The termination condition of the iterative search is reaching the maximum number of iterations or the fitness value converges, and the algorithm outputs the hyperparameter combination with the highest fitness value in history. In the population initialization phase, an initial Brown Bear population is randomly generated within a preset hyperparameter range. In the development phase, the Brown Bear's walking, stepping, and twisting movements are simulated. The walking, stepping, and twisting movements correspond to extensive exploration in the early stage of iteration, approaching the optimal solution in the middle stage, and fine local search in the later stage, respectively.

7. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 6, characterized in that, The development phase specifically includes: The walking action is extensively explored in the first third of iterations to coarsely scan the region and avoid prematurely getting trapped in local optima. The action expression is: ,in, This is Brown Bear's current location. This is Brown Bear's original location. The random step size coefficient is used to introduce randomness for the search. The current iteration ratio and satisfying the formula ; The stepping action moves towards the current optimal solution during the middle 1 / 3 of the iterations, driving all Brown Bears to learn from the better Brown Bears and converge to the known optimal region. The action expression is: ,in, The updated location for Brown Bear. A factor to control the magnitude of movement. This represents the position of the historical best solution in the current population. This represents the position of the worst solution in the current population. The ankle-twisting action undergoes a more refined local search in the last 1 / 3 of iterations. This search focuses on the vicinity of the optimal solution to ultimately pinpoint the precise global optimum. The action expression is as follows: ,in, The updated location for Brown Bear. The value is random and the range of values ​​is , The first in the current population The position of the historical best solution in each dimension. The first in the current population The position of the worst solution in each dimension; The expression for the exploration phase is: ,in, For the first Brown Bear in the first The new position updated in each dimension The parameter values ​​prior to the start of this exploration phase, i.e., the first... Brown Bear in the first The original position in each dimension It is a random number. and The first one randomly selected from the population Only with the first Brown Bear's current location. This is the fitness function.

8. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 6, characterized in that, The expression for the fitness function is: ,in, For fitness value, Mean square error, It is a very small constant used to avoid division by zero errors.

9. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 1, characterized in that, The multi-source heterogeneous data in step S1 includes environmental data and system operation data. The environmental data includes total solar radiation intensity, ambient temperature, relative humidity and wind speed. The system operation data includes DC voltage of photovoltaic modules, DC current, front temperature of modules and back sheet temperature.

10. The floating photovoltaic power generation prediction method based on optimized Brown Bear memory network according to claim 1, characterized in that, The multidimensional statistical indicators in step S4 include the coefficient of determination, root mean square error, and mean absolute error.

Images