A photovoltaic intelligent monitoring and management system based on Internet of Things

Abstract

The application relates to the technical field of photovoltaic intelligent operation and maintenance, in particular to a photovoltaic intelligent monitoring and management system based on the Internet of Things, which comprises the following steps: a data acquisition module acquires original data streams of a sensing array arranged on a photovoltaic panel; a quantitative evaluation module determines a composite contamination adhesion index and a chemical dissolution characteristic index; an efficiency prediction module determines a predicted cleaning success rate based on the composite contamination adhesion index, the chemical dissolution characteristic index, real-time temperature, real-time humidity and a preset cleaning strategy control variable; a strategy generation module solves an optimal cleaning strategy corresponding to minimum resource cost function; and a model correction module acquires actual cleaning efficiency after the optimal cleaning strategy is executed, and corrects the efficiency prediction module according to the deviation between the actual cleaning efficiency and the predicted cleaning success rate; the application constructs a core prediction engine connecting contamination characteristics and cleaning strategies, and greatly improves the robustness and accuracy of the prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic intelligent operation and maintenance, in particular to a photovoltaic intelligent monitoring and management system based on Internet of Things. BACKGROUND

[0002] In the field of photovoltaic power station operation and maintenance, the cleaning of photovoltaic panels is a core link to ensure power generation efficiency. In order to maintain power generation, the operation and maintenance team needs to regularly remove various types of dirt attached to the surface of the photovoltaic panel. This process involves the consumption of multiple resources such as water, electricity, and chemical cleaning agents.

[0003] The current photovoltaic panel cleaning strategy mainly relies on fixed time periods or manual experience, lacking real-time and quantitative perception of the specific physical and chemical properties of panel surface dirt. This approach cannot make fine adjustments to the actual working conditions of the adhesion strength, chemical composition, and environmental temperature and humidity of the dirt, as it lacks a prediction model for the effectiveness of the cleaning solution and a cost optimization function for resource input. The existing method is difficult to balance cleaning effectiveness and resource consumption, often leading to over-cleaning or under-cleaning, resulting in significant waste of resources such as water, electricity, and cleaning agents, and reducing the overall economic benefits of operation and maintenance. In addition, the traditional cleaning mode is an open-loop management that cannot use historical cleaning data to feedback and correct future decisions, lacking the ability to adapt and optimize itself. Therefore, how to build an intelligent management system that can perceive from multiple dimensions, quantify online, predict effectiveness, and optimize in a closed loop to achieve optimal allocation of photovoltaic cleaning resources is a technical problem that needs to be solved in this field. SUMMARY

[0004] To solve the above technical problems, the present application provides a photovoltaic intelligent monitoring and management system based on Internet of Things, specifically, the technical solution of the present application includes:

[0005] A data acquisition module is used to obtain the original data stream of a sensing array deployed on a photovoltaic panel. The original data stream includes: an attenuation signal and a phase shift signal of an acoustic impedance sensor, a multi-channel reflectivity vector of a multi-spectral sensor, and real-time temperature and real-time humidity of an environmental sensor.

[0006] A quantitative evaluation module is used to determine a composite dirt adhesion index and a chemical dissolution characteristic index based on the original data stream and preset calibration parameters.

[0007] An effectiveness prediction module is used to determine a predicted cleaning success rate based on the composite dirt adhesion index, the chemical dissolution characteristic index, the real-time temperature, the real-time humidity, and preset cleaning strategy control variables.

[0008] A strategy generation module is used to solve an optimal cleaning strategy corresponding to the minimization of a resource cost function based on the predicted cleaning success rate and a preset cleaning effectiveness threshold.

[0009] a model correction module configured to obtain an actual cleaning performance after the optimal cleaning strategy is executed, and correct the performance prediction module according to a deviation between the actual cleaning performance and the predicted cleaning success rate.

[0010] Preferably, the quantitative evaluation module is configured to determine the composite contamination adhesion index, and includes:

[0011] the decay signal and the phase shift signal in the original data stream, and a preset reference decay signal, a preset reference phase shift signal and a preset weight coefficient of a clean photovoltaic panel surface;

[0012] based on an empirical model of acoustic impedance physics, the relative change amount of the decay signal and the reference decay signal, and the relative change amount of the phase shift signal and the reference phase shift signal are weighted and summed;

[0013] the result of the weighted sum is determined as the composite contamination adhesion index.

[0014] Preferably, the quantitative evaluation module is configured to determine the chemical dissolution characteristic index, and includes:

[0015] the multi-channel reflectance vector in the original data stream, and a preset spectral channel index set representing an organic matter absorption peak, a preset spectral channel index set representing an inorganic matter absorption peak and a sensitivity weight coefficient;

[0016] based on an empirical proportional model, the sum of the weighted reflectances representing the organic matter in the multi-channel reflectance vector is calculated respectively;

[0017] and the sum of the weighted reflectances representing the inorganic matter is calculated respectively;

[0018] the chemical dissolution characteristic index is obtained by calculating the ratio of the sum of the weighted reflectances representing the organic matter to the sum of the weighted reflectances representing the inorganic matter.

[0019] Preferably, the performance prediction module is configured to determine the predicted cleaning success rate, and includes:

[0020] determining a contamination resistance index according to the adhesion index, real-time temperature, real-time humidity and a preset model coefficient;

[0021] determining a cleaning rate index according to the chemical dissolution characteristic index, the cleaning strategy control variable and a preset rate coefficient;

[0022] combining the contamination resistance index, the cleaning rate index, the action time in the cleaning strategy control variable and a preset gain coefficient, and determining the predicted cleaning success rate based on a logistic growth model.

[0023] Preferably, the performance prediction module is configured to determine the contamination resistance index, and includes:

[0024] calling the adhesion index, real-time humidity, real-time temperature, and preset humidity normalization reference value, temperature normalization reference value, and model coefficient;

[0025] linearly combining the adhesion index, the ratio of real-time humidity to humidity normalization reference value, and the ratio of real-time temperature to temperature normalization reference value based on an empirical linear regression model;

[0026] determining the fouling resistance index as a result of the linear combination.

[0027] Preferably, the performance prediction module is configured to determine the cleaning rate index, comprising:

[0028] calling the chemical dissolution characteristic index, and cleaning strategy control variables, the cleaning strategy control variables including: water pressure, cleaning liquid temperature, cleaning agent A concentration, and cleaning agent B concentration;

[0029] and calling preset pressure normalization reference value, temperature normalization reference value, physical cleaning rate coefficient, and chemical enhancement coefficient;

[0030] based on the physical cleaning rate coefficient, performing weighted summation on the ratio of water pressure to pressure normalization reference value, and the ratio of cleaning liquid temperature to temperature normalization reference value, to obtain a physical cleaning rate term;

[0031] based on the chemical dissolution characteristic index, cleaning agent A concentration, cleaning agent B concentration, and chemical enhancement coefficient, determining a chemical enhancement factor, wherein the chemical enhancement factor is used to dynamically adjust the effectiveness of cleaning agent A concentration and cleaning agent B concentration;

[0032] determining the cleaning rate index according to the product of the physical cleaning rate term and the chemical enhancement factor.

[0033] Preferably, the strategy generation module is configured to solve the optimal cleaning strategy, comprising:

[0034] calling the cleaning strategy control variables, real-time temperature, and preset cost coefficients;

[0035] based on a linear cost accounting model, constructing a resource cost function, wherein the resource cost function includes energy consumption cost and cleaning agent cost;

[0036] constructing a constraint optimization objective, wherein the constraint optimization objective aims to minimize the resource cost function, and is subject to the constraint condition that the predicted cleaning success rate is not lower than the cleaning performance threshold;

[0037] solving the constraint optimization objective to determine the optimal cleaning strategy.

[0038] Preferably, the energy consumption cost in the resource cost function comprises a unit temperature difference energy consumption cost,

[0039] When the cleaning liquid temperature is greater than the real-time temperature, the unit temperature difference energy consumption cost is calculated based on the difference between the cleaning liquid temperature and the real-time temperature.

[0040] When the cleaning liquid temperature is not greater than the real-time temperature, the unit temperature difference energy consumption cost is zero.

[0041] Preferably, the model correction module is used to correct the efficiency prediction module, comprising:

[0042] calling the actual cleaning efficiency and the predicted cleaning success rate;

[0043] calculating the deviation of the actual cleaning efficiency and the predicted cleaning success rate, and determining a loss function based on the deviation;

[0044] calculating the gradient of the loss function to preset model coefficients, rate coefficients and gain coefficients in the efficiency prediction module based on a gradient descent update rule;

[0045] updating the model coefficients, the rate coefficients and the gain coefficients according to the gradient and a preset learning rate.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] 1. On-line and multi-dimensional quantification of fouling characteristics is realized; the present system acquires acoustic impedance and multi-spectral data through a sensing array, on one hand, based on an acoustic impedance physical model, the relative change amount of the attenuation and phase shift signals is weighted and summed to obtain an adhesion index that can evaluate the sticking strength; on the other hand, based on an empirical proportional model, by calculating the ratio of the sum of the weighted reflectivity representing organic matter and inorganic matter, a chemical dissolution characteristic index is obtained, which reduces the complex original data to key and quantifiable indexes guiding subsequent prediction and decision-making, and realizes accurate perception of the physical and chemical characteristics of the fouling;

[0048] 2. A core prediction engine connecting fouling characteristics and cleaning strategies is constructed; the present system innovatively uses a logistic growth model to predict the cleaning success rate; the model is driven by two core indexes: a fouling resistance index that can dynamically reflect environmental influences is constructed by linearly combining the adhesion index and real-time temperature and humidity; a cleaning rate index that can intelligently integrate water pressure, temperature and the efficiency of symptomatic cleaning agent is constructed by the product of the physical cleaning rate term and the chemical enhancement factor, which accurately fits the nonlinear cleaning process and greatly improves the robustness and accuracy of the prediction model;

[0049] 3. The constraint optimization and cost precision control of cleaning decision are realized; the generation of the cleaning strategy is converted into an optimization problem with the minimum of the resource cost function as the goal and the predicted cleaning success rate not lower than the efficiency threshold as the constraint, the cost function is based on a linear model, accurately calculates the cleaning agent cost and energy consumption cost, and especially makes non-negative correction to the energy consumption cost, only when the cleaning liquid temperature is higher than the ambient temperature, the temperature difference is charged, which ensures that the system can automatically solve the most economical strategy combination under the premise of ensuring the cleaning effect;

[0050] 4. The adaptive correction closed loop based on actual effect feedback is established; the system has online learning ability, after the execution of the cleaning strategy, the loss function is determined by calculating the deviation of the actual efficiency and the predicted success rate, the system calculates the gradient of the loss function to the preset model, rate and gain coefficient in the efficiency prediction module based on the gradient descent update rule, and iteratively updates combined with the learning rate. This builds a complete machine learning closed loop, so that the system can continuously optimize itself, ensuring that its prediction model always maintains long-term high accuracy and adaptability when facing new pollution or environmental changes. BRIEF DESCRIPTION OF DRAWINGS

[0051] The present application will be further explained in conjunction with the accompanying drawings and examples:

[0052] Figure 1 is a structural diagram of the system of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further explained in conjunction with specific examples.

[0054] Example 1

[0055] Please refer to Figure 1 A photovoltaic intelligent monitoring and management system based on Internet of Things, comprising:

[0056] A data acquisition module for acquiring the original data stream of the sensing array deployed on the photovoltaic panel, the original data stream comprising: the attenuation signal and the phase shift signal of the acoustic impedance sensor, the multi-channel reflectivity vector of the multi-spectral sensor, and the real-time temperature and real-time humidity of the environmental sensor;

[0057] A quantitative evaluation module for determining the composite pollution adhesion index and the chemical dissolution characteristic index based on the original data stream and the preset calibration parameters;

[0058] An efficiency prediction module for determining the predicted cleaning success rate based on the composite pollution adhesion index, the chemical dissolution characteristic index, the real-time temperature, the real-time humidity, and the preset cleaning strategy control variable;

[0059] The strategy generation module is configured to determine an optimal cleaning strategy corresponding to a minimum resource cost function based on the predicted cleaning success rate and a preset cleaning efficiency threshold;

[0060] The model correction module is configured to obtain an actual cleaning efficiency after the optimal cleaning strategy is executed, and correct the efficiency prediction module according to a deviation between the actual cleaning efficiency and the predicted cleaning success rate.

[0061] The embodiment provides a photovoltaic intelligent monitoring and management system based on an Internet of Things; the system aims to solve the technical problems of existing photovoltaic cleaning strategies that rely on manual experience or fixed cycles, resulting in resource waste and low cleaning efficiency; the system realizes accurate management of photovoltaic panel contamination and optimal allocation of resources by constructing an intelligent closed loop from multi-dimensional perception, online quantification, efficiency prediction to closed loop optimization;

[0062] The data acquisition module is configured to obtain original contamination state data of the surface of the photovoltaic panel in real time and in multiple dimensions, to provide original input for subsequent quantitative evaluation; in the embodiment, the module is realized by a sensing array deployed at key positions of the photovoltaic panel; the sensing array specifically includes:

[0063] An acoustic impedance sensor is configured to obtain an attenuation signal representing physical adhesion characteristics of the contamination layer And a phase shift signal ; the attenuation signal and the phase shift signal refer to the amount of physical change of the acoustic wave when passing through the interface between the contamination layer and the photovoltaic panel, and the change degree is highly related to the physical adhesion strength and thickness of the contamination layer and other characteristics;

[0064] A multispectral sensor is configured to obtain a multi-channel reflectivity vector representing chemical composition of the contamination layer ; the multi-channel reflectivity vector refers to reflectivity readings of the contamination layer under multiple specific spectral channels, and the proportion of chemical composition can be inferred by analyzing the reflectivity of specific absorption peaks;

[0065] An environmental sensor is configured to obtain real-time temperature And real-time humidity of the environment where the photovoltaic panel is located; the two parameters will affect the adhesion characteristics of the contamination and the chemical reaction rate of the cleaning process;

[0066] The quantitative evaluation module is configured to convert the heterogeneous and original data stream collected by the sensing array into a structured contamination characteristic index that can be used for engineering decision-making; based on the original data stream and preset calibration parameters such as reference signals of clean surfaces and weight coefficients, the module decouples and quantifies the physical and chemical characteristics of the contamination, and finally determines two core indicators:

[0067] A composite contamination adhesion index : composite adhesion index refers to a dimensionless parameter for quantifying the physical adhesion strength of the fouling layer to the surface of the photovoltaic panel; The higher the value, the more firmly the fouling adheres, and the more difficult it is to physically remove;

[0068] chemical dissolution characteristic index : chemical dissolution characteristic index refers to a dimensionless parameter for representing the relative proportion of organic matter and inorganic matter in the fouling; The value will directly guide the subsequent proportioning of chemical cleaning agents;

[0069] The performance prediction module aims to simulate the removal effect of different cleaning strategies on the current specific fouling before the actual execution of the cleaning action. This module is a core prediction engine, and its inputs include: 1. The composite adhesion index and the chemical dissolution characteristic index determined by the quantitative evaluation module; 2. The real-time temperature and real-time humidity obtained by the data acquisition module; and 3. The preset cleaning strategy control variable to be evaluated; the cleaning strategy control variable refers to the parameter combination describing a cleaning action, such as water pressure , cleaning liquid temperature , cleaning agent A / B concentration and action time ; based on the above inputs, the module determines the predicted cleaning success rate ; the predicted cleaning success rate is a probability value between 0 and 1, representing the expected effect of a specific strategy in removing specific fouling under the current environment ;

[0070] The strategy generation module aims to find the optimal solution that takes into account cleaning effect and resource cost from a large number of possible cleaning strategy combinations. This module constructs a constrained optimization problem: it constructs constraint conditions based on the predicted cleaning success rate output by the performance prediction module and the preset cleaning performance threshold , and constructs a resource cost function ; the cleaning performance threshold refers to the minimum acceptable cleaning standard preset by the operation and maintenance team according to the actual working conditions or power generation contract requirements, for example, ; the module takes the minimization of the resource cost function as the optimization objective and as the constraint condition to solve the optimization problem, and finally determines the optimal cleaning strategy ;

[0071] The model correction module aims to build a closed-loop feedback mechanism, enabling the system to adapt to environmental changes, variations in fouling types, or deviations in initial model parameters, continuously improving the accuracy of the prediction model; and ensuring the system executes the optimal cleaning strategy. Then, the module obtains the actual cleaning efficiency after executing the optimal cleaning strategy. For example, the efficiency can be assessed using a light transmittance sensor or power generation reading of the photovoltaic panels after cleaning; subsequently, the module calculates the actual cleaning efficiency. Compared with the predicted cleaning success rate Deviation between Based on this deviation, the module uses online learning algorithms in machine learning, such as gradient descent, to correct the internal model coefficients in the performance prediction module, so that the model is closer to the actual situation in the next prediction.

[0072] This embodiment constructs a complete intelligent monitoring and management closed loop, from pollution perception to characteristic quantification, strategy prediction, optimal decision-making, and closed-loop correction. Compared with traditional cleaning methods that rely on experience or fixed cycles, this system can dynamically generate targeted cleaning strategies with the lowest resource consumption based on the real-time, in-situ physicochemical characteristics of photovoltaic panels. Its significant technical effect is that, while ensuring that the power generation efficiency is restored to the standard, it greatly reduces the consumption of water, electric heating, pressurization, and chemical cleaning agents, realizing refined, automated, and resource-optimized photovoltaic operation and maintenance.

[0073] Example 2

[0074] The quantitative assessment module is used to determine the composite fouling adhesion index, including:

[0075] Call the attenuation signal and phase shift signal in the original data stream, as well as the preset reference attenuation signal, reference phase shift signal and weighting coefficient of the clean photovoltaic panel surface;

[0076] Based on the empirical model of acoustic impedance physics, the relative changes of the attenuation signal and the reference attenuation signal, as well as the relative changes of the phase shift signal and the reference phase shift signal, are weighted and summed.

[0077] The weighted summation result was determined as the composite fouling adhesion index.

[0078] This embodiment specifically defines the determination of the composite fouling adhesion index in the quantitative evaluation module. This method aims to provide an accurate, online, non-destructive method for quantifying dirt adhesion.

[0079] In calculation At that time, the quantization evaluation module calls the attenuated signal from the original data stream. and phase shift signal ; meanwhile, the module calls preset reference attenuation signal of clean photovoltaic panel surface from system configuration , reference phase shift signal and weight coefficient from system configuration

[0080] reference attenuation signal and reference phase shift signal : refers to the acoustic impedance sensor reading obtained by calibration when the clean photovoltaic panel surface is free of contamination, which is derived from the reference value measured and stored in advance in the system initialization or laboratory calibration stage

[0081] weight coefficient : refers to a dimensionless parameter for balancing the contribution of attenuation signal and phase shift signal to total adhesion; it is determined through offline calibration experiment; specifically, a set of samples with different calibrated adhesion forces are prepared and denoted as , wherein , the contaminated samples are measured by standard pull-off test, and their corresponding calibrated attenuation signals and calibrated phase shift signals are measured respectively; based on the set of data, the weight coefficient that can make the calculated and have the best linear correlation is solved through multiple regression analysis or least square fitting, and satisfies and ;

[0082] The module performs calculation based on an empirical model combining the physical principle of acoustic impedance and the fitting of calibration data; the presence of the contaminated layer changes the propagation characteristics of acoustic waves at the interface, resulting in the shift of and relative to the reference values and ; this shift is quantified as the adhesion index by the following formula :

[0083]

[0084] wherein, : real-time attenuation / phase shift signal, derived from the data acquisition module

[0085] : reference attenuation / phase shift signal and weight coefficient, derived from preset calibration parameters

[0086] : relative change amount of attenuation signal and reference attenuation signal ​

[0087] : the relative change of the phase shift signal to the reference phase shift signal;

[0088] The two dimensionless relative changes are weighted and summed, and the module determines the composite fouling adhesion index as the result of the weighted sum ; both terms on the right side of the formula are the product of a dimensionless relative change and a dimensionless weight, so their sum is also a dimensionless parameter with the same dimension;

[0089] The embodiment provides an accurate method for online, non-destructive and quantitative evaluation of the adhesion degree of fouling; by fusing the relative changes of the attenuation and phase shift two acoustic signals and performing weighted sum, the two physical measurement indicators of different dimensions are fused into a single, robust composite fouling adhesion index ; this solves the problem that the traditional method cannot evaluate the physical fixation strength of fouling in real time, and provides a key and quantifiable input for the subsequent efficiency prediction module to accurately calculate the fouling resistance.

[0090] Embodiment 3

[0091] The quantitative evaluation module is configured to determine the chemical dissolution characteristic index, and includes:

[0092] The multi-channel reflectance vector in the original data stream is called, as well as a preset spectral channel index set representing an organic matter absorption peak, a spectral channel index set representing an inorganic matter absorption peak, and a sensitivity weight coefficient;

[0093] Based on an empirical proportion model, the sum of the weighted reflectances representing the organic matter in the multi-channel reflectance vector is calculated;

[0094] and the sum of the weighted reflectances representing the inorganic matter is calculated;

[0095] The chemical dissolution characteristic index is obtained by calculating the ratio of the sum of the weighted reflectances representing the organic matter to the sum of the weighted reflectances representing the inorganic matter.

[0096] The embodiment further limits the way in which the quantitative evaluation module determines the chemical dissolution characteristic index ; this method aims to provide an online method for distinguishing the organic / inorganic chemical composition of fouling;

[0097] In the calculation , the quantitative evaluation module calls the multi-channel reflectance vector in the original data stream ; at the same time, the module calls a preset spectral channel index set representing an organic matter absorption peak , a spectral channel index set representing an inorganic matter absorption peak , and a sensitivity weight coefficient ;

[0098] Spectral Channel Index Set : refers to the numbering of a specific wavelength channel pre-selected based on the spectral absorption characteristics of common pollutants, such as CH bonds representing organic matter, Si-O bonds or OH bonds representing inorganic matter or water. The source is a pre-set standard pollutant spectral library.

[0099] Sensitivity weighting coefficient This refers to the calibration coefficient for the corresponding spectral channel, used to correct for sensitivity differences between different channels; it is obtained through multivariate calibration experiments; specifically, the preparation... The group has a known ratio of organic to inorganic components, denoted as . ,in Standard soiled samples were used, and their corresponding multi-channel reflectance vectors were measured. Based on this Group The data, after being calibrated using multivariate methods such as partial least squares regression, is used to solve for the following formula: and Sensitivity weighting coefficients with optimal relevance and ;

[0100] The module is based on an empirical proportional model, and its technical motivation stems from the Beer-Lambert law in spectral analysis, which states that the content of a specific component is related to the absorbance / reflectance at a specific wavelength. The calculation is performed using the following formula:

[0101]

[0102] in, Chemical solubility index, dimensionless; This usually indicates that the soiling is mainly composed of organic matter. This indicates that the fraction is primarily composed of inorganic substances; to ensure robustness of calculations, a constant is added to the denominator. It is a preset non-negative minimum value used to prevent division by zero errors, for example... This is to avoid computational overflow caused by extremely weak signals from inorganic materials.

[0103] : The first vector or The reflectivity value of the channel comes from the data acquisition module;

[0104] The spectral channel index set and sensitivity weighting coefficients are derived from preset calibration parameters;

[0105] The sum of the weighted reflectance representing the organic matter in the multi-channel reflectance vector, i.e., the numerator of the formula ;

[0106] and the sum of the weighted reflectance representing the inorganic matter, i.e., the denominator of the formula ;

[0107] The chemical dissolution characteristic index is obtained by calculating the ratio of the sum of the weighted reflectance representing the organic matter to the sum of the weighted reflectance representing the inorganic matter ;

[0108] In the formula, are dimensionless reflectance, is also a dimensionless calibration coefficient, so the numerator and the denominator are dimensionless quantities, and the ratio is also a dimensionless parameter, and has the same dimension;

[0109] The embodiment provides a method for online and rapid differentiation of chemical components of a stain; through the empirical proportion model, complex multi-spectral vector data is reduced to a single and intuitive chemical dissolution characteristic index ; the index clearly reflects whether the stain is inclined to water-soluble inorganic matter or solvent organic matter, and provides a core decision basis for subsequent efficiency prediction module calculation of a cleaning rate and strategy generation module intelligent adjustment of a concentration ratio of cleaning agent A organic and cleaning agent B inorganic.

[0110] The efficiency prediction module is configured to determine a predicted cleaning success rate, and includes:

[0111] The stain resistance index is determined according to the adhesion index, the real-time temperature, the real-time humidity and a preset model coefficient;

[0112] The cleaning rate index is determined according to the chemical dissolution characteristic index, the cleaning strategy control variable and a preset rate coefficient;

[0113] The predicted cleaning success rate is determined based on a logistic growth model, in combination of the stain resistance index, the cleaning rate index, an action time in the cleaning strategy control variable and a preset gain coefficient.

[0114] The embodiment further limits the detailed three-step logic of the efficiency prediction module for determining the predicted cleaning success rate ; the prediction model is a core self-defined model for connecting stain characteristics and a cleaning strategy to predict a cleaning result;

[0115] The efficiency prediction module determines the adhesion index , the real-time temperature , the real-time humidity and the preset model coefficient determining the fouling resistance index ;

[0116] fouling resistance index refers to a dimensionless parameter that integrates the physical adhesion strength of the fouling and the environmental factors of temperature and humidity, and is used to quantify the inherent difficulty of removing the current fouling;

[0117] the chemical dissolution characteristic index calculated by the quantification and evaluation module , the cleaning strategy control variable to be evaluated , and the preset rate coefficient , to determine the cleaning rate index ;

[0118] cleaning rate index refers to a parameter for quantifying the effectiveness of a specific cleaning strategy per unit time, and its physical dimension is set as i.e. the proportion of cleanliness completed per second;

[0119] the fouling resistance index obtained in the first step , the cleaning rate index obtained in the second step , the action time in the cleaning strategy control variable , and the preset gain coefficient , based on the logistic growth model, to determine the predicted cleaning success rate ;

[0120] gain coefficient refers to a dimensionless parameter used to control the steepness of the logistic function curve, which is obtained by historical cleaning data regression training together with other model coefficients such as ; during system operation, this coefficient will be dynamically updated by the model correction module;

[0121] The technical motivation of this logistic / Sigmoid model is that the cleaning effect is not simply linear accumulation, but there is a resistance threshold and a cleaning saturation effect; the cleaning effect must overcome representing the resistance, and the cleaning effect will eventually saturate around 100%; the logistic function can well describe this S-shaped growth curve, and the model formula used in this embodiment is as follows:

[0122]

[0123] wherein, : predicted cleaning success rate, dimensionless (0-1);

[0124] : cleaning rate index, dimension ;

[0125] : Action time, dimensionless ;

[0126] : Fouling resistance index, a dimensionless parameter;

[0127] Gain coefficient, a preset parameter, is a dimensionless parameter;

[0128] The core of this model lies in calculating net cleaning power, i.e. ;where Fc\tau This represents the total cleaning effect and is dimensionless. Representing fouling resistance, it is also dimensionless; therefore, the exponential term... The whole is dimensionless. It is also dimensionless, yet has consistent dimensions; when the total action is greater than the fouling resistance, the net efficacy is positive, and the cleaning success rate is high. The net effectiveness approaches 1; conversely, when the amount of action is insufficient to overcome the resistance, the net effectiveness is negative. Approaching 0;

[0129] This embodiment constructs a system capable of connecting to fouling characteristics. With cleaning strategy The core prediction engine abstracts the complex cleaning process into an adversarial relationship between resistance and rate, and uses a logistic growth model for nonlinear fitting, enabling the system to adapt to any strategy. Any soiling and environment Cleaning effect Performing quantitative predictions is the prerequisite and foundation for the subsequent policy generation module to perform constraint optimization.

[0130] The performance prediction module is used to determine the fouling resistance index, including:

[0131] Call the adhesion index, real-time humidity, real-time temperature, and preset humidity normalization baseline value, temperature normalization baseline value and model coefficients;

[0132] Based on an empirical linear regression model, the adhesion index, the ratio of real-time humidity to the humidity normalized baseline value, and the ratio of real-time temperature to the temperature normalized baseline value are linearly combined.

[0133] The result of the linear combination was determined as the fouling resistance index.

[0134] This embodiment specifically defines the determination of the fouling resistance index in the performance prediction module. The way;

[0135] In calculating , the performance prediction module calls the adhesion index , real-time humidity , real-time temperature , and preset humidity normalization reference value , temperature normalization reference value and model coefficients from the previous steps;

[0136] Normalization reference value : refers to the reference value for the dimensionless of environmental parameters, for example , 25°C, which is preset;

[0137] Model coefficients : refers to the dimensionless fitting coefficients, which are obtained by training multiple linear regression on historical cleaning data sets; the data set contains groups of historical data, each group of data including historical adhesion index , historical humidity , historical temperature and corresponding, experimentally calibrated historical fouling resistance , where ; these coefficients are fitted by regression analysis to minimize the residual of calculated by the following formula ;

[0138] The calculation of uses an empirical linear regression model, which is technically motivated to comprehensively evaluate the combined effects of physical adhesion and environmental factors on cleaning difficulty, and the model formula is as follows:

[0139]

[0140] where, : fouling resistance index, a dimensionless parameter;

[0141] : adhesion index, derived from the quantitative evaluation module, a dimensionless parameter;

[0142] : ratio of real-time humidity to humidity normalization reference value, a dimensionless parameter;

[0143] : ratio of real-time temperature to temperature normalization reference value, a dimensionless parameter;

[0144] : model coefficient, derived from preset parameters, dimensionless parameter;

[0145] The model linearly combines the adhesion index , the humidity normalized ratio , and the temperature normalized ratio ; all terms on the right side of the formula are dimensionless, so is dimensionless and has the same dimension; based on physical cognition, the coefficient adhesion and humidity are expected to be positive, that is, the stronger the adhesion and the greater the humidity, the more difficult the fouling is to remove, and the greater the resistance, while temperature is expected to be negative, that is, the higher the temperature, the more likely the fouling is to soften, and the smaller the resistance; the model determines the fouling resistance index as the result of the linear combination.

[0146] The definition of the fouling resistance in this embodiment is more comprehensive and closer to the actual working condition; it not only considers the physical adhesion of the fouling itself , but also innovatively incorporates environmental factors into the resistance model through normalization and linear regression; this enables the system to automatically identify and quantify, for example, the case where the cleaning resistance significantly increases in a low-temperature and high-humidity environment, thereby greatly improving the accuracy of the index and the robustness of the prediction model.

[0147] The performance prediction module is configured to determine a cleaning rate index, and includes:

[0148] The chemical dissolution characteristic index and the cleaning strategy control variables, including water pressure, cleaning liquid temperature, cleaning agent A concentration, and cleaning agent B concentration, are called;

[0149] and the preset pressure normalization reference value, temperature normalization reference value, physical cleaning rate coefficient, and chemical enhancement coefficient are called;

[0150] Based on the physical cleaning rate coefficient, the ratio of the water pressure to the pressure normalization reference value and the ratio of the cleaning liquid temperature to the temperature normalization reference value are weighted and summed to obtain a physical cleaning rate term;

[0151] Based on the chemical dissolution characteristic index, cleaning agent A concentration, cleaning agent B concentration, and chemical enhancement coefficient, a chemical enhancement factor is determined, wherein the chemical enhancement factor is used to dynamically adjust the effectiveness of the cleaning agent A concentration and the cleaning agent B concentration;

[0152] The cleaning rate index is determined according to the product of the physical cleaning rate term and the chemical enhancement factor.

[0153] The embodiment specifically defines the determination of the cleaning rate index in the performance prediction module The model employs an empirical multiplicative interaction model, aiming to capture the synergistic effect of physical and chemical cleaning;

[0154] When calculating , the performance prediction module calls the chemical solubility characteristic index from the previous step, as well as the cleaning strategy control variables to be evaluated; in this embodiment, the cleaning strategy control variables are specifically defined to include: water pressure , cleaning liquid temperature , cleaning agent A concentration , and cleaning agent B concentration ;

[0155] Cleaning agent A and cleaning agent B refer to special cleaning agents for organic matter and inorganic matter, respectively;

[0156] At the same time, the module calls preset pressure normalization reference values , temperature normalization reference values , physical cleaning rate coefficients , and chemical enhancement coefficients ;

[0157] Normalization reference values : reference values for pressure and temperature dimensionless, preset;

[0158] Rate / enhancement coefficients : obtained by regression training on a historical cleaning data set; the data set contains groups of historical data, each group of data including historical fouling characteristics , historical cleaning strategies used , and corresponding, experimentally calibrated historical cleaning rates , where ; these coefficients are fitted by nonlinear regression to minimize the residual of calculated by the following formula ;

[0159] To ensure dimensional consistency, the dimension of the physical cleaning rate coefficient is set to ; the dimension of the chemical enhancement coefficient is set to the inverse of the concentration, for example , assuming that the dimension of is , The calculation formula is as follows:

[0160]

[0161] The physical cleaning rate term is obtained:

[0162] The first part of the calculation formula is defined as the physical cleaning rate term;

[0163] This term is based on a weighted sum of the ratio of water pressure to a pressure normalized reference value , the ratio of cleaning fluid temperature to a temperature normalized reference value ,

[0164] Dimension check: and are dimensionless, the dimension of is ,

[0165] The chemical enhancement factor is determined:

[0166] The second part of the calculation formula is defined as the chemical enhancement factor;

[0167] This factor is based on the chemical dissolution characteristic index , the concentration of cleaning agent A , the concentration of cleaning agent B

[0168] Dimension check: ( ) ( ) dimensionless dimensionless; ( ) ( ) dimensionless dimensionless; the constant 1 is also dimensionless; therefore, the chemical enhancement factor as a whole is dimensionless;

[0169] The core innovation of this design is that the chemical enhancement factor is used to dynamically adjust the effectiveness of the concentrations of cleaning agent A and cleaning agent B; specifically:

[0170] When the contamination is biased towards organic matter, a high value means that the term has a high weight, the term has a low weight, and at this time the contribution of organic solvents is amplified;

[0171] When the contamination is biased towards inorganic matter, the value is low, the item weight is high, the item weight is low, and the contribution of inorganic solvents is amplified;

[0172] The cleaning rate index is determined:

[0173] According to the product of the physical cleaning rate term and the chemical enhancement factor, and the dimensionless product, the total cleaning rate index is determined; its final dimension is ;

[0174] This embodiment realizes the intelligent fusion of physical strategies and chemical strategies through a clever, dimension-consistent multiplication interaction model; in particular, the design of the chemical enhancement factor, which uses the chemical dissolution characteristic index as a dynamic weight, automatically amplifies the effect of the symptomatic cleaning agent or , while suppressing the contribution of mismatched cleaning agents; this makes the index accurately reflect the true cleaning rate of a specific strategy for a specific contamination , avoiding the waste of resources and potential chemical residues caused by the blind use of two cleaning agents.

[0175] The strategy generation module is used to solve the optimal cleaning strategy, including:

[0176] The cleaning strategy control variable, real-time temperature, and preset cost coefficients are called;

[0177] Based on the linear cost accounting model, a resource cost function is constructed, wherein the resource cost function includes energy consumption cost and cleaning agent cost;

[0178] A constraint optimization objective is constructed, wherein the constraint optimization objective aims to minimize the resource cost function and is constrained by the condition that the predicted cleaning success rate is not lower than the cleaning efficiency threshold;

[0179] The constraint optimization objective is solved to determine the optimal cleaning strategy.

[0180] This embodiment specifically limits the way in which the strategy generation module solves the optimal cleaning strategy corresponding to the minimization of the resource cost function;

[0181] To perform the solution, the strategy generation module constructs the resource cost function in the following steps: the module calls the cleaning strategy control variable to be optimized​ Real-time temperature From the data acquisition module, and the preset cost coefficient of each item ;

[0182] Cost coefficient : refers to the unit cost of each resource, which is set in advance according to the actual electricity price and the cleaning agent purchase price; in order to ensure dimensional consistency, its dimension is set as: For example: yuan Pa s , For example: yuan K s , For example: yuan s , assuming The dimension is ;

[0183] Based on the linear cost accounting model, the module constructs a resource cost function ;

[0184] Resource cost function Refers to the total economic cost consumed by the execution strategy Dimension: yuan; this function includes energy cost and cleaning agent cost in this embodiment, and its specific form is as follows:

[0185]

[0186] Dimension check:

[0187] Energy cost item 1: (Yuan Pa s ) (Pa) (s) Yuan;

[0188] Energy cost item 2: (Yuan K s ) (K) (s) Yuan;

[0189] Cleaning agent cost item: (Yuan s ) (%)+ (Yuan s​ ) (%) (s) Yuan;

[0190] All terms in the formula are in units of yuan, can be added together, and have consistent dimensions.

[0191] First item The total energy cost includes the cost of water pump pressurization and the cost of water heating;

[0192] Second item Total cleaning agent cost;

[0193] After constructing the cost function, the module constructs a constraint optimization objective;

[0194] The goal of this optimization problem is to find a strategy that minimizes cost and achieves the desired results. ;

[0195] In this embodiment, the constrained optimization objective is defined as:

[0196] Objective function: That is, the objective is to minimize the resource cost function;

[0197] Constraints: That is, the constraint is that the predicted cleaning success rate is not lower than the cleaning efficiency threshold.

[0198] The module solves for the constrained optimization objective;

[0199] The system employs numerical optimization algorithms, such as Sequential Least Squares Programming (SQP), or intelligent algorithms like Genetic Algorithms and Particle Swarm Optimization, to solve the above problems. The final solution... This refers to the determined optimal cleaning strategy; the choice of specific algorithm can be based on the efficiency requirements, the continuity and discreteness of the decision variables, and the cost function. and constraints The nonlinearity and nonconvexity intensity are adapted to achieve a balance between computational resources and global optimization.

[0200] This embodiment transforms the cleaning decision from a fuzzy, experience-based operational problem into a precise, solvable constrained optimization problem; it does so by constructing a quantifiable cost function that ensures dimensional consistency. and performance constraints This system can ensure that the cleaning effect is no less than that of other systems. Under the rigid premise, automatically find the strategy combination that minimizes the total cost of resources such as water, electricity, and cleaning agents. This enables precise cost control and optimal resource allocation in the cleaning process.

[0201] The energy consumption cost in the resource cost function includes a unit temperature difference energy consumption cost,

[0202] When the cleaning liquid temperature is greater than the real-time temperature, the unit temperature difference energy consumption cost is calculated based on the difference between the cleaning liquid temperature and the real-time temperature.

[0203] When the cleaning liquid temperature is not greater than the real-time temperature, the unit temperature difference energy consumption cost is zero.

[0204] The embodiment specifically defines a key detail of the energy consumption cost in the resource cost function , that is, the calculation method of the heating cost.

[0205] The energy consumption cost in the resource cost function specifically includes a unit temperature difference energy consumption cost; the mathematical expression of this cost item in is .

[0206] The function refers to taking the larger value between 0 and .

[0207] The technical motivation of the model is to correct the logical defect that the cost may be negative in the simple linear model , so that the cost accounting model is more consistent with the physical reality; the specific logic is as follows:

[0208] When the cleaning liquid temperature is greater than the real-time temperature : At this time , ; the unit temperature difference energy consumption cost is calculated based on the difference between the cleaning liquid temperature and the real-time temperature, and the total heating cost is ; this represents the actual cost of the system consuming energy to heat the water from the environmental temperature to ;

[0209] When the cleaning liquid temperature is not greater than the real-time temperature : that is ; at this time , ; the unit temperature difference energy consumption cost is zero; this represents that the system uses water at the environmental temperature or colder water for cleaning, for example, at night or in winter, and does not need to consume additional energy for heating, so the heating cost is 0.

[0210] The embodiment accurately defines the triggering condition and calculation logic of the heating cost by introducing this nonlinear function; this non-negativity correction ensures that the cost item always conforms to the physical meaning that the cost cannot be negative, so that the cost function is more accurate, thereby ensuring that the optimal strategy obtained by the strategy generation module Reliability and economy.

[0211] Example 9:

[0212] The model correction module, used to correct the performance prediction module, includes:

[0213] The actual cleaning efficiency and the predicted cleaning success rate are compared.

[0214] Calculate the deviation between actual cleaning efficiency and predicted cleaning success rate, and determine the loss function based on the deviation;

[0215] Based on the gradient descent update rule, the gradient of the loss function with respect to the preset model coefficients, rate coefficients, and gain coefficients in the performance prediction module is calculated.

[0216] Update the model coefficients, rate coefficients, and gain coefficients based on the gradient and the preset learning rate.

[0217] This embodiment specifically defines the closed-loop feedback mechanism of the model correction module for correcting the performance prediction module; this mechanism adopts the standard gradient descent update rule of online machine learning to achieve model adaptation and self-evolution;

[0218] In the optimal strategy After the process is completed and the actual cleaning effect is obtained, the model correction module performs the following correction process:

[0219] Module calls include, for example, the actual cleaning efficiency measured by a light transmittance sensor. The predicted cleaning success rate used in this forecast The data originates from the performance prediction module, and the deviation between the two is calculated. ;

[0220] Based on this bias, the module determines a loss function. In this embodiment, the mean squared error (MSE) is used as the loss function:

[0221]

[0222] loss function Used to quantify the accuracy of model predictions. The smaller the value, the more accurate the prediction; because and All are dimensionless. It is also dimensionless;

[0223] To achieve model correction, the module calculates the loss function based on the gradient descent update rule. The set of all preset coefficients to be trained in the performance prediction module is denoted as . gradient ;

[0224] coefficient set In this embodiment, the following are included:

[0225] model coefficients in the model ;

[0226] rate coefficients in the model ;

[0227] gain coefficients in the model ;

[0228] the gradient is calculated by the chain rule from via the formula logistic model back propagation, which is in the form of ;

[0229] After the gradient is calculated, the module updates the above-mentioned model coefficients, rate coefficients and gain coefficients according to the gradient and a preset learning rate ;

[0230] The learning rate is a preset, dimensionless tuning parameter, for example 0.01, used to control the step size of each correction, and the formula of the update rule is as follows:

[0231]

[0232] wherein, is the current coefficient, is the updated coefficient; to ensure dimensional consistency, the update rule requires that the dimension of the correction term must match the dimension of the coefficient ; for non-dimensionless coefficients, such as rate coefficients, the corresponding gradient term needs to be processed for necessary dimensional adaptation before updating, to ensure that the update operation is physically valid; for example, the gradient term can be multiplied by a reference value with the same dimension as the coefficient to be updated, or divided by the typical scale value of the coefficient, to realize the dimensionless of the correction term;

[0233] This embodiment constructs a complete online machine learning closed loop based on actual effect feedback; through the gradient descent update rule, the system can use the data of each actual cleaning to automatically and iteratively fine-tune the internal coefficients of the core prediction model ; This makes the system have strong adaptive ability, can continuously learn the characteristics of new type of pollution or the influence brought by environmental change, ensure that the performance prediction module evolves over time, always keep its high accuracy of prediction and effectiveness of optimal strategy; This closed-loop correction mechanism based on actual data feedback also effectively compensates for the loss of physical fidelity in the experience model of the performance prediction module for simplified calculation, ensuring that the final prediction effect of the model can continuously approach the real physical process through data-driven way.

[0234] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An Internet of Things based photovoltaic intelligent monitoring and management system, characterized in that, The method comprises the following steps: a data acquisition module is configured to acquire a raw data stream of a sensor array deployed on a photovoltaic panel, the raw data stream comprising: an attenuation signal and a phase shift signal of an acoustic impedance sensor, a multi-channel reflectance vector of a multi-spectral sensor, and real-time temperature and real-time humidity of an environmental sensor; a quantitative evaluation module is configured to determine a composite contamination adhesion index and a chemical dissolution characteristic index based on the raw data stream and preset calibration parameters; an empirical model based on the physical principle of acoustic impedance is configured to perform weighted summation on the relative change amount of the attenuation signal and a reference attenuation signal of a clean photovoltaic panel surface, and the relative change amount of the phase shift signal and a reference phase shift signal of the clean photovoltaic panel surface; and determine the weighted summation result as the composite contamination adhesion index; an empirical proportional model is configured to calculate the sum of weighted reflectances representing organic matter in the multi-channel reflectance vector, and the sum of weighted reflectances representing inorganic matter in the multi-channel reflectance vector, respectively; a chemical dissolution characteristic index is obtained by calculating the ratio of the sum of weighted reflectances representing organic matter to the sum of weighted reflectances representing inorganic matter; an efficiency prediction module is configured to determine a predicted cleaning success rate based on the composite contamination adhesion index, the chemical dissolution characteristic index, the real-time temperature, the real-time humidity, and a preset cleaning strategy control variable; the efficiency prediction module is configured to determine the predicted cleaning success rate by: determining a contamination resistance index according to the composite contamination adhesion index, the real-time temperature, the real-time humidity, and a preset model coefficient; determining a cleaning rate index according to the chemical dissolution characteristic index, the cleaning strategy control variable, and a preset rate coefficient; combining the contamination resistance index, the cleaning rate index, the action time in the cleaning strategy control variable, and a preset gain coefficient, and determining the predicted cleaning success rate based on a logistic growth model; a strategy generation module is configured to solve an optimal cleaning strategy corresponding to a minimum resource cost function based on the predicted cleaning success rate and a preset cleaning efficiency threshold value; a model correction module is configured to obtain an actual cleaning efficiency after executing the optimal cleaning strategy, and correct the efficiency prediction module according to the deviation between the actual cleaning efficiency and the predicted cleaning success rate. The quantitative evaluation module is configured to call the attenuation signal and the phase shift signal in the raw data stream, and the reference attenuation signal, the reference phase shift signal, and the weight coefficient. 2.The photovoltaic intelligent monitoring and management system based on Internet of Things according to claim 1, characterized in that, The quantitative evaluation module is configured to call the multi-channel reflectance vector in the raw data stream, and a spectral channel index set representing an organic matter absorption peak, a spectral channel index set representing an inorganic matter absorption peak, and a sensitivity weight coefficient. 3.The photovoltaic intelligent monitoring and management system based on Internet of Things according to claim 2, characterized in that, The efficiency prediction module is configured to determine the contamination resistance index by: 4.The photovoltaic intelligent monitoring and management system based on Internet of Things according to claim 3, characterized in that, calling the composite contamination adhesion index, the real-time humidity, the real-time temperature, a humidity normalization reference value, a temperature normalization reference value, and a model coefficient; performing linear combination on the ratio of the composite contamination adhesion index to the humidity normalization reference value, and the ratio of the real-time temperature to the temperature normalization reference value, based on an empirical linear regression model; determining the linear combination result as the contamination resistance index. The efficiency prediction module is configured to determine the cleaning rate index by: 5.The photovoltaic intelligent monitoring and management system based on Internet of Things according to claim 4, characterized in that, ​ The chemical dissolution characteristic index is called, as well as cleaning strategy control variables, including: water pressure, cleaning liquid temperature, cleaning agent A concentration and cleaning agent B concentration; And preset pressure normalization reference value, temperature normalization reference value, physical cleaning rate coefficient and chemical enhancement coefficient are called; Based on the physical cleaning rate coefficient, the ratio of water pressure to pressure normalization reference value and the ratio of cleaning liquid temperature to temperature normalization reference value are weighted and summed to obtain a physical cleaning rate term; Based on the chemical dissolution characteristic index, cleaning agent A concentration, cleaning agent B concentration and chemical enhancement coefficient, a chemical enhancement factor is determined, wherein the chemical enhancement factor is used to dynamically adjust the effectiveness of cleaning agent A concentration and cleaning agent B concentration; According to the product of the physical cleaning rate term and the chemical enhancement factor, a cleaning rate index is determined. 6.The photovoltaic intelligent monitoring and management system based on Internet of Things according to claim 5, characterized in that, The strategy generation module is used to solve the optimal cleaning strategy, including: Calling cleaning strategy control variables, real-time temperature, and preset cost coefficients; Based on a linear cost accounting model, a resource cost function is constructed, wherein the resource cost function includes energy consumption cost and cleaning agent cost; A constraint optimization target is constructed, wherein the constraint optimization target aims to minimize the resource cost function, and the constraint optimization target is constrained by the predicted cleaning success rate being not lower than the cleaning efficiency threshold; The constraint optimization target is solved to determine the optimal cleaning strategy. 7.The photovoltaic intelligent monitoring and management system based on Internet of Things according to claim 6, characterized in that, The energy consumption cost in the resource cost function includes a unit temperature difference energy consumption cost, Wherein, when the cleaning liquid temperature is greater than the real-time temperature, the unit temperature difference energy consumption cost is calculated based on the difference between the cleaning liquid temperature and the real-time temperature; When the cleaning liquid temperature is not greater than the real-time temperature, the unit temperature difference energy consumption cost is zero. 8.The photovoltaic intelligent monitoring and management system based on the Internet of Things according to claim 7, characterized in that, The model correction module is used to correct the efficiency prediction module, including: Calling actual cleaning efficiency and predicted cleaning success rate; The deviation of the actual cleaning efficiency and the predicted cleaning success rate is calculated, and a loss function is determined based on the deviation; Based on the gradient descent update rule, the gradient of the loss function to the preset model coefficient, rate coefficient and gain coefficient in the efficiency prediction module is calculated; According to the gradient and the preset learning rate, the model coefficient, the rate coefficient and the gain coefficient are updated.

Images