Island reef underwater photovoltaic selection and deployment optimization method and system

Abstract

The application discloses an island reef underwater photovoltaic type selection and deployment optimization method and system, and belongs to the cross technical field of clean energy and marine optics. The method comprises the following steps: S1, acquiring the hyperspectral absorption coefficient, the beam attenuation coefficient and the finite angle volume scattering function at different water depths of a target island reef water area, calculating the backscattering ratio and extrapolating to the full spectrum; S2, taking the finite angle volume scattering function and the full spectrum backscattering ratio as input, constructing a neural network regression model guided by a physical constraint loss function, and directly outputting a reconstruction result; S3, establishing a decision mapping relationship between the optical state of the target water area and photovoltaic configuration parameters based on the reconstruction result, and compiling a calculation table; S4, pre-storing the calculation table in a central controller of an underwater photovoltaic system deployed in the target water area, and the central controller only executes the calculation table and configuration instruction output. The application is suitable for the engineering practice of remote islands.

Description

Technical Field

[0001] This invention belongs to the field of clean energy and marine optics interdisciplinary technology, specifically relating to a method and system for the selection and deployment optimization of underwater photovoltaic systems for islands and reefs. Background Technology

[0002] Among various zero-carbon energy technologies, photovoltaic (PV) power generation has become the preferred renewable energy source for islands and reefs due to its abundant solar energy resources, high modularity, and relatively simple maintenance. However, traditional land-based / floating PV faces fundamental challenges in island and reef applications: the land area is limited, and floating PV is directly exposed to strong typhoon paths, both constrained by space and environment. In contrast, underwater PV exhibits unique environmental adaptability advantages—extending power generation space to the vertical profile of the water body without occupying island and reef surface resources; the attenuation effect of water on typhoon waves protects it from mechanical damage by extreme sea conditions; and the cooling effect of seawater helps maintain stable battery operating temperatures, reducing thermally induced efficiency losses. These advantages make underwater PV irreplaceable in zero-carbon energy supply for islands and reefs.

[0003] However, when traditional onshore photovoltaic technology is directly transplanted to the underwater environment, the limited utilization of photovoltaic energy in the water, the failure of key materials, and the low coupling efficiency of multiple energy systems have become the core bottlenecks restricting the development of underwater photovoltaic technology.

[0004] Underwater photovoltaics utilizes the spectral filtering and cooling effects of water to achieve conversion efficiencies superior to those on land at specific depths. Clean waters surrounding islands and reefs (Jerlov Type I and II) offer good light transmittance and stable temperatures, making them ideal application scenarios for underwater photovoltaics.

[0005] In recent years, significant progress has been made in underwater photovoltaic technology research. While Wang et al. (Solar Energy, 2026) demonstrated through optical-electrical-thermal coupling modeling that immersion depth can significantly improve underwater photovoltaic performance, their optimal window is located in shallow water at the centimeter level, making it impractical for engineering applications in typhoon-prone areas. Furthermore, the optical inputs are static parameters from literature studies, failing to consider dynamic changes in the water's optical state, and lacking engineering configuration schemes for material selection, deployment depth, and control strategies. PNAS (2024) reported experimental verification of the optimal underwater depth for different bandgap materials (c-Si, CIGS, CdTe, a-Si), with a-Si achieving an efficiency of 59.7% at a depth of 2m. Regarding the acquisition of marine optical parameters, Piskozub et al. (Oceanologia, 2007) proposed a Fournier-Forand phase function parameterization method using the backscattering ratio Bp as the parameter variable; Optics Express (2015) further established a semi-analytical relationship between Bp and the refractive index n and the particle size distribution slope μ. Regarding MPPT technology, Solar Energy (2020) has systematically categorized MPPT techniques such as incremental conductance method, perturbation observation method, variable step size adaptive algorithm, and global scan algorithm. Zhang Xianqing et al. (2023) from the South China Sea Institute of Oceanology, Chinese Academy of Sciences, used a seven-angle volume scattering function profiler to measure the volume scattering function β(ψ) and combined it with the absorption coefficient a measured by ac-s, employing machine learning methods to estimate the single-band Kd(650) profile distribution. In this study, the particulate backscattering ratio Bp was simplified to a wavelength-independent constant (Boss et al., 2007) to meet the technical requirement of only needing to estimate single-band Kd.

[0006] However, the aforementioned existing technologies still have the following shortcomings: (i) Regarding the information abundance and technical objectives of the output target parameters, Zhang Xianqing et al. (2023) only output K. d (650) Single-band, this study serves marine optics and water color remote sensing research, but only remains at the academic estimation level; although "underwater photovoltaic power" is mentioned as a possible application direction, it does not disclose how to reduce the diffusion attenuation coefficient K. d The estimation results are then translated into specific engineering configuration plans (material selection, deployment depth, MPPT strategy, etc.).

[0007] (ii) Regarding the Bp processing method, existing technologies will B p It is simplified to a wavelength-independent constant, a simplification that requires only estimating a single-band K. d This is reasonable within the existing framework, but if the goal is to select full-spectrum photovoltaic materials, estimating only the single-band K... d The framework is unreasonable and the calculations are inaccurate.

[0008] (iii) In terms of reconstruction methods, existing technologies use pure data-driven machine learning models, which lack constraints on the physical consistency of reconstruction results.

[0009] (iv) Regarding engineering deployment, the aforementioned optical-electrical-thermal coupling research and material selection research both focus on the theoretical optimization or experimental verification of a single configuration point, without considering the impact of the spatiotemporal dynamic changes in the optical state of the water body on the system's annual efficiency, nor providing a systematic multi-scenario configuration scheme for engineering deployment. In addition, although existing technologies have proposed schemes for long-term underwater deployment of optical parameter measurement equipment to achieve real-time adaptive control, such schemes face serious engineering obstacles in island and reef environments: long-term underwater deployment of high-precision optical instruments is severely affected by biofouling and requires monthly cleaning and maintenance; the power supply and real-time data transmission of the equipment place extremely high demands on the energy and communication infrastructure of the islands and reefs; and the purchase and maintenance costs of long-term deployed equipment far exceed the affordability of remote island and reef projects.

[0010] In summary, the existing technology has the following technical contradictions: on the one hand, fixed-configuration underwater photovoltaics cannot adapt to the dynamic changes in the optical state of the water body and it is difficult to maintain optimal efficiency throughout the year; on the other hand, although the real-time closed-loop adaptive scheme can be dynamically optimized, its deployment and maintenance costs are too high and its engineering feasibility is poor in remote island and reef environments. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for the selection and deployment optimization of underwater photovoltaic systems for islands and reefs, which is suitable for the engineering practice of remote islands and reefs.

[0012] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention provides a method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs, comprising the following steps: S1. Obtain the hyperspectral absorption coefficient, beam attenuation coefficient, and finite-angle volume scattering function at different water depths in the target island / reef waters, and calculate the backscattering ratio; the backscattering ratio is then extrapolated to obtain the full-spectrum backscattering ratio. S2. Using the finite-angle volume scattering function and the full-spectrum backscattering ratio as input, construct a neural network regression model. Use the physical constraint loss function to guide the training of the neural network regression model through backpropagation, so that the neural network regression model directly outputs the reconstruction result. The reconstruction result includes the full-angle, full-spectrum volume scattering function and the diffuse attenuation coefficient. S3. Based on the reconstruction results, establish the decision mapping relationship between the optical state of the target water area and the photovoltaic configuration parameters, and compile it into a lookup table; S4. The lookup table is pre-stored in the central controller of the underwater photovoltaic system in the target water area. The central controller only executes the lookup table and outputs the configuration instructions.

[0013] Optionally, the method further includes: S5. Repeat steps S1 to S3 according to a preset cycle or after an abnormal water event occurs to update the lookup table.

[0014] Optionally, the lookup table includes the following decision information: peak wavelength λc of the transmission spectrum corresponding to different water depth ranges, photovoltaic material type, optimal deployment depth z*, MPPT strategy, and expected net output power range.

[0015] Optionally, the physical constraint loss function includes data consistency constraints based on finite angle measurements, backscattering integral consistency constraints based on the physical closure relationship of backscattering ratio, total scattering integral consistency constraints based on the physical closure relationship of total scattering coefficient, and angle domain smoothness regularization constraints.

[0016] Optionally, the backscattering ratio is: B p (λ0)=b b (λ0) / b(λ0); Among them, b b (λ0) is the backscattering coefficient at the reference wavelength λ0, which is obtained by back hemispherical integration and extrapolation correction from the measured finite-angle volume scattering function; b(λ0) is the total scattering coefficient at the reference wavelength λ0, b(λ0)=c(λ0)-a(λ0), where c(λ0) and a(λ0) are the beam attenuation coefficient and hyperspectral absorption coefficient at the reference wavelength λ0, respectively.

[0017] Optionally, the full-spectrum backscattering ratio obtained by spectral extrapolation of the backscattering ratio is: B p (λ) = B p (λ0)·(λ / λ0) (-η) ; Wherein, η is the spectral change index, which is determined by inversion based on the spectral shape information carried by the hyperspectral absorption coefficient a(λ) and the beam attenuation coefficient c(λ); the inversion adopts a multi-path fusion strategy with the attenuation spectral slope path as the main path and the absorption spectral correction path and machine learning path as the correction verification, and the results of multiple paths are fused by weighted fusion according to their credibility; λ is any wavelength in the visible light range.

[0018] Optionally, in step S3, the photovoltaic configuration parameters include the optimal deployment depth z*, which is determined by: The optical input power at various water depths is calculated based on the diffuse attenuation coefficient. The photovoltaic operating temperature heat balance equation is established and solved to obtain the operating temperature T(z). With the maximization of net output power Pelec(z) as the objective function, z* = argmax Pelec(z) is searched within the feasible water depth range, and z* is included in the lookup table.

[0019] Optionally, the MPPT strategy is recommended based on the diffuse attenuation coefficient Kd(z,λc) characteristics, including: When Kd(z,λc) is below the first threshold, the incremental conductivity method or the perturbation observation method is recommended. When Kd(z,λc) is between the first and second thresholds, a variable step size adaptive algorithm is recommended. When Kd(z,λc) is higher than the second threshold, a global scanning algorithm is recommended.

[0020] Optionally, the photovoltaic material type is selected according to the peak wavelength λc of the transmission spectrum corresponding to different water depth ranges, including: λc∈[400, 500] nm is used to select a wide bandgap material; λc∈(500, 600] nm selects a medium bandgap material; λc∈(600, 750] nm selects a narrow bandgap material; The material configuration scheme includes the following three-layer configuration: The first layer of online electrical configuration adjusts the activation / bypass status of each substring and the series-parallel topology through a substring-level electrical switch matrix. The second layer of online parameter updates synchronously updates the MPPT policy control parameters. The third-layer offline material suggestion outputs a material replacement suggestion when λc continuously deviates beyond a preset threshold and the duration exceeds the set time.

[0021] Secondly, the present invention provides a system for selecting and optimizing the deployment of underwater photovoltaic systems for islands and reefs, implementing the method described in any of the preceding claims, comprising: The field survey subsystem includes a hyperspectral absorption attenuator and a volume scattering function profiler, used to collect survey data in the target water area. Offline modeling and guide generation subsystem: includes a computing platform and storage device to run a neural network inference module and a physical constraint loss function module, reconstruct optical parameters from the survey data, and generate a lookup table for the target water area; The underwater photovoltaic subsystem includes a sub-string array composed of various bandgap materials, a sub-string electrical switch matrix, an MPPT control unit, and a central controller. Configuration execution unit: This is the lookup decision module within the central controller. It pre-sets the lookup table and outputs electrical switch matrix control commands and MPPT parameters based on the water depth deployed on-site and the auxiliary information input by sensors.

[0022] Compared with the prior art, the advantages of this invention are as follows: (I) System Architecture of Offline Guideline Mode. The complete technical process is "site survey → offline reconstruction → guideline lookup table generation → lookup table application → periodic review". The on-site survey equipment is withdrawn from the target waters immediately after data collection. The underwater photovoltaic system only requires a low-power MCU to execute the guideline lookup table, without incorporating complex optical reconstruction algorithms or dedicated NPU / GPU hardware. Compared to real-time closed-loop solutions requiring long-term underwater monitoring, this invention has extremely low maintenance requirements, making it suitable for engineering projects on remote islands and reefs. This architecture is not publicly disclosed in existing underwater photovoltaic and marine optics literature.

[0023] (II) Physically Constrained Neural Network Reconstruction Method. Using the finite-angle volume scattering function and full-spectrum backscattering ratio measured in the field as input, a neural network regression model (referred to as a physically constrained neural network) is constructed. Four physical constraint loss functions are used for end-to-end backpropagation-guided training, enabling the model to directly output the volume scattering function and diffuse attenuation coefficient across the entire angle (0-180°, angular resolution ≤2°) and the entire spectrum (350-750nm). Unlike existing technologies that use pure data-driven training and only output a single Kd(650) band, the model of this invention outputs a β(θ,λ) spectrum covering the 0-180° range, and the training process is constrained by physical closure relationships, resulting in a physically self-consistent reconstruction result.

[0024] (III) Bp spectral processing. This differs from existing techniques that process Bp... p The simplified approach, which uses a wavelength-independent constant, is achieved by combining the Fournier-Forand phase function parameterization framework with power-law extrapolation to realize B. p (λ) Full-spectrum reconstruction, for full-spectrum reconstruction and transmission spectrum peak wavelength λ c Precise calculations provide the physical basis.

[0025] (iv) Structured lookup table guidelines and three-layer material configuration. Based on the reconstruction results, a table containing the water depth range -λ is generated. c -Material type-z*-MPPT strategy-A lookup table for the six-dimensional mapping relationship of expected power is used to achieve engineering decision mapping between optical state and photovoltaic configuration. Furthermore, a three-layer material configuration scheme is adopted: the first layer is online electrical configuration, adjusting the activation / bypass state and series-parallel topology of each substring through a substring-level electrical switch matrix; the second layer is online parameter updating, synchronously updating MPPT control parameters; the third layer is offline material suggestion, when λ c If the deviation continues to exceed the preset threshold, a material replacement suggestion will be output.

[0026] (v) Low-maintenance mechanism for regular review. The guidelines are updated at a preset cycle (six months to one year for clean water bodies) or after an abnormal water event. The survey equipment is withdrawn after data collection, and the underwater photovoltaic system does not carry complex optical reconstruction algorithms or dedicated NPU / GPU hardware. Attached Figure Description

[0027] Figure 1 A flowchart illustrating the method for selecting and optimizing underwater photovoltaic systems for islands and reefs, provided in this application embodiment; Figure 2 A schematic diagram illustrating the composition of the underwater photovoltaic selection and deployment optimization system for islands and reefs provided in this application embodiment; Figure 3 This is an overall flowchart illustrating an application scenario example of this application; Figure 4 Flowchart for optical parameter reconstruction; Figure 5 Diagram of the physical constraint loss function mechanism; Figure 6 Flowchart for deep optimization of optical-electrical-thermal coupling; Figure 7 The purpose of this guide is to calculate the intended meaning. Detailed Implementation

[0028] Example: The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] See Figure 1 As shown in this embodiment, the method for selecting and optimizing the deployment of underwater photovoltaic systems for islands and reefs includes the following steps: S1. Obtain the hyperspectral absorption coefficient a(λ), beam attenuation coefficient c(λ), and finite-angle volume scattering function β(θi, λ0) at different water depths in the target island / reef area to obtain the particulate backscattering ratio B. p (λ0); λ0 is the single-band operating wavelength of the finite-angle volume scattering function β(θi, λ0), and λ is any wavelength in the visible light range (350≤λ≤750nm).

[0030] Specifically, the backscattering ratio B p (λ0)=b b (λ0) / b(λ0); where b b (λ0) is the backscattering coefficient at the reference wavelength λ0, which is obtained by back hemispherical integration and extrapolation correction from the measured finite angle volume scattering function; b(λ0) is the total scattering coefficient at the reference wavelength λ0, b(λ0)=c(λ0)-a(λ0), where c(λ0) and a(λ0) are the beam attenuation coefficient and hyperspectral absorption coefficient at the reference wavelength λ0, respectively; By extrapolating the backscattering ratio Bp(λ0) at the reference wavelength λ0 to the spectral variation exponent η according to a power law relationship, the full-spectrum backscattering ratio Bp(λ) is obtained: B p (λ) = Bp(λ0)·(λ / λ0) (-η) ; Wherein, η is the spectral change index, which is determined by inversion based on the spectral shape information carried by the hyperspectral absorption coefficient a(λ) and the beam attenuation coefficient c(λ). The inversion adopts a multi-path fusion strategy with the attenuation spectral slope path as the main path and the absorption spectral correction path and machine learning path as the correction and verification paths. The results of the multiple paths are fused by weighting according to their credibility.

[0031] Thus, unlike existing technologies, B p The simplified approach, which reduces the input to a wavelength-independent constant, is achieved by combining the Fournier-Forand phase function parameterization framework with power-law extrapolation to realize B. p (λ) Full-spectrum conversion, for subsequent steps of full-spectrum reconstruction and transmission spectrum peak wavelength λ c Precise calculations provide the physical basis.

[0032] S2. Using the finite-angle volume scattering function and the full-spectrum backscattering ratio as input, construct a neural network regression model. Use a physical constraint loss function to guide the training of the neural network regression model, so that the neural network regression model directly outputs the reconstruction result. The reconstruction result includes the full-angle, full-spectrum volume scattering function and the diffuse attenuation coefficient.

[0033] Specifically, the physical constraint loss function includes: data consistency constraint based on finite angle measurements, backscattering integral consistency constraint based on the physical closure relationship of backscattering ratio, total scattering integral consistency constraint based on the physical closure relationship of total scattering coefficient, and angle domain smoothness regularization constraint; the weights of the four constraints are determined through a validation set during the offline calibration phase.

[0034] Thus, using the finite-angle volume scattering function and full-spectrum backscattering ratio measured in the field as input, a neural network regression model (referred to as a physically constrained neural network) is constructed. Training is guided by four physically constrained loss functions, enabling the model to directly output the volume scattering function and diffuse attenuation coefficient across the entire angle (0-180°, angular resolution ≤2°) and the entire spectrum (350-750nm). Unlike existing technologies that only output a single Kd(650) band, this invention's model outputs a β(θ,λ) spectrum covering the 0-180° range, rather than simply outputting phase function parameters or backscattering ratios.

[0035] S3. Based on the reconstruction results, establish a decision mapping relationship between the optical state of the target water area and the photovoltaic configuration parameters, and compile it into a lookup table; the lookup table shall contain at least the following decision information: peak wavelength λc of transmission spectrum corresponding to different water depth ranges, photovoltaic material type, optimal deployment depth z*, MPPT (Maximum Power Point Tracking) strategy and expected net output power range, as a guiding basis for the design, selection and operation and maintenance of underwater photovoltaic system in the island and reef water area.

[0036] Specifically, the optimal deployment depth z* is determined as follows: the optical input power Popt(z) at each water depth is calculated based on the reconstructed diffuse attenuation coefficient, the photovoltaic operating temperature heat balance equation is established and the operating temperature T(z) is obtained by solving it, and the net output power Pelec(z) is maximized as the objective function. Within the feasible water depth range, z* = argmax Pelec(z) is searched, and z* is included in the lookup table.

[0037] MPPT strategies are recommended based on the diffuse attenuation coefficient Kd(z,λc): when Kd(z,λc) is below the first threshold, the incremental conductance method or perturbation observation method is recommended; when Kd(z,λc) is between the first and second thresholds, the variable step size adaptive algorithm is recommended; when Kd(z,λc) is above the second threshold, the global scan algorithm is recommended.

[0038] Photovoltaic materials are selected based on the peak wavelength λc of the transmission spectrum: wide bandgap materials are selected for λc∈[400, 500] nm; medium bandgap materials are selected for λc∈(500, 600] nm; and narrow bandgap materials are selected for λc∈(600, 750] nm. The material configuration scheme is implemented in three layers: the first layer is online electrical configuration, which adjusts the activation / bypass status and series-parallel topology of each substring through the substring-level electrical switch matrix; the second layer is online parameter updating, which updates the MPPT control parameters synchronously; and the third layer is offline material suggestion, which outputs material replacement suggestions when λc continuously deviates beyond the preset threshold and the duration exceeds the set time.

[0039] Thus, based on the reconstruction results, a range containing the water depth interval -λ is generated. c -Material type-z*-MPPT strategy-A lookup table for the six-dimensional mapping relationship of expected power is used to achieve engineering decision mapping between optical state and photovoltaic configuration. Furthermore, a three-layer material configuration scheme is adopted: the first layer is online electrical configuration, adjusting the activation / bypass state and series-parallel topology of each substring through a substring-level electrical switch matrix; the second layer is online parameter updating, synchronously updating MPPT control parameters; the third layer is offline material suggestion, when λ c If the deviation continues to exceed the preset threshold, a material replacement suggestion will be output.

[0040] S4. The lookup table is pre-stored in the central controller of the underwater photovoltaic system in the target water area. The central controller only executes the lookup table and outputs the configuration instructions.

[0041] Specifically, the central controller periodically reads data from the optional lightweight sensor. When the PAR sensor reading is lower than a certain percentage of the expected value, it prompts "Water body abnormal, review recommended," triggering a temporary review process. The expected value is derived from the underwater light field simulation results during the guideline generation phase or from the reference measurement calibration during the initial system deployment.

[0042] Thus, the underwater photovoltaic system only requires a low-power MCU to execute lookup tables, without incorporating complex optical reconstruction algorithms or dedicated NPU / GPU hardware. Compared to real-time closed-loop solutions that require long-term underwater monitoring, this invention has extremely low maintenance requirements, making it suitable for engineering projects on remote islands and reefs.

[0043] As a preferred method for the selection and deployment optimization of underwater photovoltaic systems on islands and reefs, the method further includes: S5. Repeat steps S1 to S3 according to a preset cycle or after an abnormal water event occurs to update the lookup table.

[0044] Specifically, the review cycle is determined based on the optical stability of the target water body: Jerlov Type I and Type II clean water bodies are reviewed every six months to one year; the abnormal water events include at least one of the following: typhoon passage, internal wave activity, and river distillation water inflow, which triggers a temporary review upon occurrence.

[0045] Thus, updates are performed according to a preset cycle (six months to one year for clean water bodies) or after an abnormal water event. The survey equipment is withdrawn after data collection, and the underwater photovoltaic system does not carry complex optical reconstruction algorithms or dedicated NPU / GPU hardware, establishing a low-maintenance mechanism for regular review.

[0046] Accordingly, such as Figure 2 As shown, this embodiment also provides an underwater photovoltaic selection and deployment optimization system for islands and reefs that implements the above method, including: (a) Field survey subsystem: includes a hyperspectral absorption attenuator and a volume scattering function profiler, used to collect survey data in the target water area. An optional backscattering measurement instrument can be used for cross-validation. This subsystem is carried by a ship or underwater robot and the equipment is removed from the target water area after the survey is completed.

[0047] (b) Offline Modeling and Guideline Generation Subsystem: This subsystem includes a high-performance computing platform and storage devices. It runs a neural network inference module and a physical constraint loss function module to reconstruct optical parameters from survey data and generate lookup tables for the target water area. Specifically, the neural network model employs a two-stage strategy: in the pre-training stage, it uses a large-scale dataset generated by radiative transfer simulation for supervised training and joint training with physical constraints; in the field modeling stage, it uses survey data from the target water area to perform transfer fine-tuning on the pre-trained model before inference application, generating a mapping relationship library and lookup tables specific to that water area.

[0048] (c) Underwater photovoltaic subsystem: includes a sub-string array composed of various bandgap materials, a sub-string electrical switch matrix, an MPPT control unit and a central controller; (d) Configuration Execution Unit: Located in the central controller, it incorporates the lookup table and outputs electrical switch matrix control commands and MPPT parameters based on the water depth deployed on-site and auxiliary information input from optional lightweight sensors. It does not carry a real-time optical parameter reconstruction algorithm or dedicated NPU hardware. Specifically, the central controller uses an industrial-grade low-power MCU, only executing the guideline lookup table and outputting configuration commands, with an online response cycle ≤1 second and an average system power consumption ≤2W.

[0049] The following example application scenario will be used to further verify and illustrate the method for selecting and optimizing underwater photovoltaic systems for islands and reefs in this application.

[0050] Taking a Jerlov I type clean water body surrounding a certain island reef as an example. The chlorophyll concentration in this sea area is <0.3 mg / m³ (NASAOcean Color Web remote sensing statistics). The time scale for changes in optical state is from days to weeks, and the long-term trend changes are from seasons to grades, such as... Figure 3 As shown, the overall workflow includes the following steps: I. Site Survey Timing of the survey: A baseline survey should be conducted before the initial system deployment, ideally covering different seasons (such as spring and autumn) to obtain the interannual variation range of optical conditions. Each survey should last 1-3 days and be carried out during a window of clear skies and favorable sea conditions.

[0051] Measurement equipment configuration: AC-S hyperspectral absorption attenuation meter was used to measure a(λ) and c(λ) (350-750nm, resolution 5nm); VSAM-17 volume scattering function profiler was used to measure β(θ) at 17 angles (10°-170°, 10° intervals). i (659nm); optional ECO-BB9 backscatterer for measuring b bp and b p To calculate B p , used for cross-validation.

[0052] Measurement profile requirements: water depth range 0-30m (covering potential photovoltaic deployment area), vertical sampling interval ≤1m, sampling time for each depth point ≥30 seconds to ensure data stability.

[0053] Data preprocessing: Remove obvious outliers (such as bubble interference and instrument startup transients), perform temperature and salinity correction and dark current subtraction, and take the average of repeated measurements.

[0054] II. Offline Optical Parameter Reconstruction (I) Bp spectral acquisition

[0055] like Figure 4 As shown, bb(λ0) is calculated from the measured β(θi,λ0): First, the measured angle range of 10°-170° is fitted with the least squares form of the Fournier-Forand phase function. The volume scattering function values ​​of the 0°-10° and 170°-180° ranges are obtained by extrapolation of the fitted phase function. Then, bb(λ0) is obtained by integrating over the backward hemisphere (90°-180°). b(λ)=c(λ)-a(λ) is calculated from a(λ) and c(λ), resulting in Bp(λ0)=bb(λ0) / b(λ0).

[0056] Full spectrum B p (λ) Extrapolated using the power-law relationship through the spectral variation exponent η: B p (λ) = B p (λ0)·(λ / λ0) (-η) η is determined by inversion based on the spectral shape information carried by hyperspectral a(λ) and c(λ).

[0057] The η inversion adopts a multi-path fusion strategy: the attenuation spectral slope path is the main one, while the absorption spectral correction path and the machine learning path are used for correction and verification. The results from multiple paths are fused according to their confidence levels. Specifically: (a) Attenuation spectral slope path: the initial η estimate η is calculated using the spectral slope of c(λ). c (b) Absorption spectral correction path: using the non-aqueous absorption component of a(λ) to correct η c After correction, η is obtained. ac (c) Machine learning validation path: Using the spectral shape features of a(λ) and c(λ) as input, the output η is obtained through a pre-trained lightweight regression model. ML As an independent verification value; ultimately η = w1·η c + w2·η ac + w3·η ML The weights are determined based on the inverse ratio of the backscatter integral deviation of each path on the independent validation set.

[0058] (II) Training Data and Model Architecture The training dataset was generated using HydroLight radiative transfer simulations, with parameter ranges taken from publicly available literature (B). p ∈[0.005,0.03], n∈[1.05,1.12], μ∈[3.5,4.0]; Mobley 1994; IOCCG 2015; Twardowski et al. 2001; Boss et al. 2004).

[0059] The neural network model employs a fully connected regression network: the input is a 179-dimensional feature vector (β(θi, 659nm) for 17 angles, and a(λ) and c(λ) for 81 wavelengths, 350-750nm, spaced 5nm apart), mapped to the output layer through several fully connected hidden layers (256-512 neurons per layer, using ReLU or SiLU activation functions); the output layer, after non-negative transformation, outputs β(θ,λ) values ​​for 81 wavelengths × 180 angles (2° resolution), totaling 14580 dimensions, and an additional output branch outputs the diffuse attenuation coefficient Kd(z,λ) for each water depth. Training uses the Adam optimizer, where the angle integral and second-order difference in the physical constraint loss function are implemented in discrete summation form, exhibiting analytical gradients for each output dimension, which can be optimized end-to-end through backpropagation; for high-dimensional outputs, in practice, an autoencoder can also be used to reduce the dimensionality of the output, performing regression in the low-dimensional latent space before decoding to recover the full-angle, full-spectrum results, thereby improving model accuracy and inference efficiency.

[0060] (III) Definition and weight configuration of physical constraint loss function like Figure 5 As shown, to guide the neural network model to stably reconstruct a physically consistent full-spectrum volume scattering function based on finite angle measurement data, this invention introduces the following four physically constrained loss functions during model training: Data consistency loss Ldata measures the deviation between reconstructed values ​​and measured values ​​at a finite number of angles. Ldata = (1 / N)Σ[β(θi,λ0) - βrecon(θi,λ0)]²

[0061] The backscattering integral consistency loss Lbb, constructed based on the physical closure relationship between the backscattering ratio Bp(λ) and the volume scattering coefficient b(λ), measures the deviation between the back hemispherical integral of the reconstructed β(θ,λ) and Bp(λ)·b(λ): Lbb = [2π∫90°180° βrecon(θ,λ)sinθdθ - Bp(λ)·b(λ)]²

[0062] The total scattering integral consistency loss Lb is constructed based on the physical closure relationship c(λ) = a(λ) + b(λ) between the spectral attenuation coefficient c(λ) and the absorption coefficient a(λ) and the bulk scattering coefficient b(λ). It measures the deviation between the full-angle integral of the reconstructed β(θ,λ) and the measured total scattering coefficient b(λ). Lb = [2π∫0°180° βrecon(θ,λ)sinθdθ - b(λ)]²

[0063] The angle-domain smoothness loss Lsmooth is a regularization term for the second-order partial derivative of the reconstructed β(θ,λ) angle, used to suppress non-physical oscillations at the extrapolated angle:

[0064] The weighting of the above four losses follows the following physical logic: L bb and L b The weighting prioritizes ensuring physical integral consistency (used to compensate for systematic bias caused by inaccurate instrument χ factor under finite angle measurements), L data Centering the weights ensures data fit, L smooth The lowest weight only serves a regularization function (avoiding over-smoothing loss λ). c Calculate the required spectral shape characteristics.

[0065] Weight calibration method: The four weights are determined in the offline calibration stage through a validation set grid search. The search range is for each weight ∈ [0.1, 1.0], with a step size of 0.1. Using SeaBASS / NOMAD measured data as the validation benchmark, the weight combination that minimizes the relative deviation between the reconstructed β(θ,λ) and the measured backscattering integral is selected. In this embodiment, the weight combination used is L. bb =0.35, L b =0.35, L data =0.20, L smooth =0.10, the recommended weight range is L bb =0.3~0.5, L b =0.3~0.5, L data =0.1~0.3, L smooth =0.05~0.15.

[0066] (iv) Reconstructing the output The reconstruction yields β(θ,λ) across the entire angle (2° resolution) and the entire spectrum (350-750 nm, 5 nm resolution), as well as the diffuse attenuation coefficient K. d (z,λ). By K d (z,λ) Calculate the peak wavelength λ of the transmission spectrum at each water depth. c (z), thus obtaining λ c Curve showing the change with water depth.

[0067] III. Guideline Generation (a) Lookup table structure The reconstruction result is discretized into the following six-dimensional lookup table:

[0068] The expected net output power is an estimate based on simulation, calculated using a noon PAR irradiance of approximately 500 W / m² under clear sea conditions, an underwater simulation efficiency of 20%-30% for candidate materials, and a packaging and line loss reduction factor of 0.85. It also includes the light attenuation due to the diffuse attenuation coefficient Kd(z,λc) at various depth ranges. When turbidity events (such as sediment resuspension after a typhoon) occur with Kd(z,λc) ≥ 0.15 m⁻¹, the MPPT strategy is recommended to switch to a global scanning algorithm, triggering a temporary review. In actual deployment, the calculation needs to be recalculated based on site reconstruction results, taking into account factors such as cable loss and MPPT efficiency reduction. The peak transmission wavelength at depths of 15–20 m has shifted to the shortwave band (λc < 450 nm), and conventional deployment is not recommended for this depth range. If configuration is required, it should be determined after specialized optical and thermal calibration.

[0069] (II) Material selection and three-layer configuration scheme Photovoltaic materials are determined according to λc: λ c ∈[400,500]nm selects a wide bandgap material (InGaP, wide bandgap perovskite); λ c ∈(500,600]nm, select medium bandgap materials (CdTe, a-Si); λ c For the ∈(600,750]nm bandgap material (c-Si, CIGS), select the narrow bandgap material.

[0070] The material configuration scheme includes the following three-layer configuration: The first layer (online electrical configuration) adjusts the activation / bypass state and series-parallel topology of each substring (InGaP, CdTe, c-Si) through a substring-level electrical switch matrix to achieve adaptive material switching in different water depth ranges without the need for physical component replacement; The second layer (online parameter update): synchronously updates MPPT control parameters (step size, sampling frequency, etc.) to adapt to changes in IV properties after material switching; Third layer (offline material suggestion): When λ c A sustained offset exceeding a preset threshold (e.g., ±20nm) and a duration exceeding a set time (e.g., T) delay When the maintenance period reaches 72 hours, a material replacement recommendation will be provided so that physical material replacement can be implemented during subsequent planned maintenance.

[0071] (III) Calculation of Optimal Deployment Depth like Figure 6As shown, a photoelectric-thermal coupling model is used. The thermal balance equation for the photovoltaic operating temperature is: T(z) = [Tamb·UL + Popt(z)·(1-ηref) + h(z)·Tw] / [UL + h(z)] Net output power objective function: Pelec(z) = Popt(z)·ηSTC·[1 + βT·(T(z)-TSTC)] Search for z* = argmax P within the feasible water depth range (e.g., 1-30m). elec (z). In this embodiment, the thermodynamic parameters (T) amb =30°C, T w (10m)=24°C, h=50-100 W / (m²·K); (referencing NOAA WOD data), considering both optical attenuation and temperature benefits, z*≈5-10m, and is included in the recommended lookup table.

[0072] (iv) MPPT Strategy Recommendations MPPT strategies are classified based on the dynamic characteristics of Kd(z,λc): K d (z,λc) < 0.05 m -1 (Clear): Employing incremental conductivity or perturbation observation methods, it offers fast response and high tracking efficiency; 0.05 ≤ K d (z,λc) < 0.15 m -1 (Medium): Employs a variable step-size adaptive algorithm to balance tracking speed and steady-state accuracy; K d (z,λc) ≥ 0.15 m -1 (Turbid): A global scanning algorithm is used to avoid local optima traps under multi-peak power-voltage curves.

[0073] IV. Field Application of Guidelines The generated lookup table is embedded in the central controller of the underwater photovoltaic system (using an STM32L series or equivalent industrial-grade low-power MCU).

[0074] During on-site deployment, such as Figure 7 As shown, maintenance personnel set the following values ​​according to the guidelines: array deployment depth (z * recommended value); initial activation material substring (λ). c Recommended values); MPPT initial parameters (by K) d (Category Recommendations).

[0075] During system operation, the central controller does not need to perform any optical parameter reconstruction calculations. It only needs to: (1) periodically (e.g., every minute) read the data of the optional lightweight sensors (temperature sensor, PAR sensor); (2) make simple status judgments according to the pre-stored thresholds in the guidelines, such as prompting "water body abnormality, it is recommended to check" if the PAR is lower than 80% of the expected value; (3) output switch matrix control instructions and MPPT parameters.

[0076] V. Regular review The validity period of the guidelines depends on the stability of the optical state of the water body. In this embodiment, the optical state changes of the Jerlov I type clean water body on a certain island reef are mainly on seasonal and interannual scales, and a review cycle of six months to one year is recommended.

[0077] Events that trigger a temporary review include: typhoon passage (causing sediment resuspension, B) p It could increase by an order of magnitude; internal wave activity (which could carry deep eutrophic water to shallower layers); major engineering activities in the vicinity (such as dredging or land reclamation).

[0078] During the review, repeat the entire process of "site survey → optical parameter reconstruction → guideline update" to generate a new version of the lookup table to replace the old version.

[0079] In this application scenario, the reconstruction results are comparable to the backscattering integral deviation of NASA SeaBASS / NOMAD independent field measurement data and similar methods such as the QAA algorithm.

[0080] In terms of engineering effects, compared with the "on-site real-time deployment" scheme, the proposed scheme has the following advantages: (1) Equipment investment is reduced by about 70% (no need to deploy multiple sets of high-precision optical instruments underwater for a long time, only need to rent survey equipment periodically); (2) The power consumption of the underwater photovoltaic system is reduced from 3-5W to ≤2W (no need to run complex algorithms such as RF inference); (3) The maintenance cycle is extended from monthly to more than half a year (the photovoltaic end has no optical window that needs to be cleaned, and the survey equipment is not submerged in water for a long time); (4) The guidelines can be extended to multiple islands and reefs in the same sea area, with low marginal cost.

[0081] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs, characterized in that, Includes the following steps: S1. Obtain the hyperspectral absorption coefficient, beam attenuation coefficient, and finite-angle volume scattering function at different water depths in the target island / reef waters, and calculate the backscattering ratio; the backscattering ratio is then extrapolated to obtain the full-spectrum backscattering ratio. S2. Using the finite-angle volume scattering function and the full-spectrum backscattering ratio as input, construct a neural network regression model. Use the physical constraint loss function to guide the training of the neural network regression model through backpropagation, so that the neural network regression model directly outputs the reconstruction result. The reconstruction result includes the full-angle, full-spectrum volume scattering function and the diffuse attenuation coefficient. S3. Based on the reconstruction results, establish the decision mapping relationship between the optical state of the target water area and the photovoltaic configuration parameters, and compile it into a lookup table; S4. The lookup table is pre-stored in the central controller of the underwater photovoltaic system in the target water area. The central controller only executes the lookup table and outputs the configuration instructions.

2. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 1, characterized in that, The method further includes: S5. Repeat steps S1 to S3 according to a preset cycle or after an abnormal water event occurs to update the lookup table.

3. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 1, characterized in that, The lookup table includes the following decision information: peak wavelength λc of the transmission spectrum corresponding to different water depth ranges, photovoltaic material type, optimal deployment depth z*, MPPT strategy, and expected net output power range.

4. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 1, characterized in that, The physical constraint loss function includes data consistency constraints based on finite angle measurements, backscatter integral consistency constraints based on the physical closure relationship of backscatter ratio, total scatter integral consistency constraints based on the physical closure relationship of total scattering coefficient, and angle domain smoothness regularization constraints.

5. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 1, characterized in that, The backscattering ratio is: B p (λ0)=b b (λ0) / b(λ0); Among them, b b (λ0) is the backscattering coefficient at the reference wavelength λ0, which is obtained by back hemispherical integration and extrapolation correction from the measured finite-angle volume scattering function; b(λ0) is the total scattering coefficient at the reference wavelength λ0, b(λ0)=c(λ0)-a(λ0), where c(λ0) and a(λ0) are the beam attenuation coefficient and hyperspectral absorption coefficient at the reference wavelength λ0, respectively.

6. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 5, characterized in that, The full-spectrum backscattering ratio obtained by spectral extrapolation of the backscattering ratio is: B p (λ) = B p (λ0)·(λ / λ0) (-η) ; Wherein, η is the spectral change index, which is determined by inversion based on the spectral shape information carried by the hyperspectral absorption coefficient a(λ) and the beam attenuation coefficient c(λ); the inversion adopts a multi-path fusion strategy with the attenuation spectral slope path as the main path and the absorption spectral correction path and machine learning path as the correction verification, and the results of multiple paths are fused by weighted fusion according to their credibility; λ is any wavelength in the visible light range.

7. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 1, characterized in that, In step S3, the photovoltaic configuration parameters include the optimal deployment depth z*, which is determined as follows: The optical input power at various water depths is calculated based on the diffuse attenuation coefficient. The photovoltaic operating temperature heat balance equation is established and solved to obtain the operating temperature T(z). With the maximization of net output power Pelec(z) as the objective function, z* = argmax Pelec(z) is searched within the feasible water depth range, and z* is included in the lookup table.

8. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 3, characterized in that, The MPPT strategy is recommended based on the diffuse attenuation coefficient Kd(z,λc) characteristics, including: When Kd(z,λc) is below the first threshold, the incremental conductivity method or the perturbation observation method is recommended. When Kd(z,λc) is between the first and second thresholds, the variable step size adaptive algorithm is recommended. When Kd(z,λc) is higher than the second threshold, a global scanning algorithm is recommended.

9. The method for selecting and optimizing the deployment of underwater photovoltaic systems on islands and reefs as described in claim 3, characterized in that, The photovoltaic material type is selected according to the peak wavelength λc of the transmission spectrum corresponding to different water depth ranges, including: λc∈[400, 500] nm is used to select a wide bandgap material; λc∈(500, 600] nm selects a medium bandgap material; λc∈(600, 750] nm selects a narrow bandgap material; The material configuration scheme includes the following three-layer configuration: The first layer of online electrical configuration adjusts the activation / bypass status of each substring and the series-parallel topology through a substring-level electrical switch matrix. The second layer of online parameter updates synchronously updates the MPPT policy control parameters. The third-layer offline material suggestion outputs a material replacement suggestion when λc continuously deviates beyond a preset threshold and the duration exceeds the set time.

10. A system for selecting and optimizing the deployment of underwater photovoltaic systems for islands and reefs, implementing the method of any one of claims 1 to 9, characterized in that, include: The field survey subsystem includes a hyperspectral absorption attenuator and a volume scattering function profiler, used to collect survey data in the target water area. Offline modeling and guide generation subsystem: includes a computing platform and storage device to run a neural network inference module and a physical constraint loss function module, reconstruct optical parameters from the survey data, and generate a lookup table for the target water area; The underwater photovoltaic subsystem includes a sub-string array composed of various bandgap materials, a sub-string electrical switch matrix, an MPPT control unit, and a central controller. Configuration execution unit: This is the lookup decision module within the central controller. It pre-sets the lookup table and outputs electrical switch matrix control commands and MPPT parameters based on the water depth deployed on-site and the auxiliary information input by sensors.

Images