Solar panel control method and system based on photovoltaic monitoring platform

Abstract

The present application discloses a solar panel control method and system based on a photovoltaic monitoring platform, the method comprising: obtaining weather data, solar power generation data and solar panel images in real time through the photovoltaic monitoring platform; inputting indoor temperature, weather data, solar power generation data and solar panel images into a fusion model to obtain a solar panel optimization strategy; controlling the controller of the corresponding target solar panel according to the solar panel optimization strategy; constructing a geometric model of the target area in response to receiving a rain signal transmitted by a rain sensor; simulating the flow path of rainwater at different opening and closing angles of each target solar panel through the geometric model; determining the closing angle interval of each target solar panel according to the flow path; determining whether the opening and closing angle is within the closing angle interval; if so, controlling the controller corresponding to each target solar panel so that each target solar panel is closed according to the minimum closing angle in the closing angle interval.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of solar energy control technology, and in particular to a solar panel control method and system based on a photovoltaic monitoring platform. Background Art

[0002] Solar energy is a clean energy source. Currently, solar energy is mainly collected through solar panels and applied to daily life.

[0003] At present, solar panels installed on buildings are usually linked to windows through sliding rods. For example, for solar panels outside the computer room, when the weather is sunny and the indoor temperature is high, it is necessary to manually adjust the sliding rod so that the solar panels drive the windows to open, so that the temperature in the computer room will not be too high and the equipment is safe; when the weather is rainy, it is also necessary to manually adjust the sliding rod so that the solar panels drive the windows to close, so that the equipment in the computer room will not be wetted by rain. However, the above solution not only consumes manpower, but also poses a great challenge to the safety of the equipment in the computer room when the indoor temperature is high and it is rainy.

[0004] There is currently no better solution to the above problems. Summary of the invention

[0005] The embodiments of the present application provide a solar panel control method and system based on a photovoltaic monitoring platform, which are used to improve the safety of equipment in a computer room by controlling the solar panels.

[0006] To achieve the above objectives, the embodiments of the present application adopt the following technical solutions:

[0007] In a first aspect, a solar panel control method based on a photovoltaic monitoring platform is provided, which is applied to a service terminal, wherein the service terminal is equipped with a photovoltaic monitoring platform, wherein the photovoltaic monitoring platform is used to monitor a solar power generation system, wherein the solar power generation system includes a temperature sensor, a rain sensor, and a plurality of solar panels, wherein each of the solar panels is provided with a controller for adjusting the opening and closing angle of the solar panel, wherein the method includes:

[0008] Acquire weather data, solar power generation data within a preset time period, and solar panel images of a target area in real time through the photovoltaic monitoring platform;

[0009] In response to receiving the indoor temperature transmitted by the temperature sensor, the indoor temperature, the weather data, the solar power generation data and the solar panel image are input into a pre-trained fusion model to obtain a solar panel optimization strategy, wherein the solar panel optimization strategy is used to determine the opening and closing angle of the target solar panel;

[0010] Controlling a controller of a corresponding target solar panel according to the solar panel optimization strategy so that the controller adjusts the opening and closing angle;

[0011] In response to receiving a rain signal transmitted by the rain sensor, constructing a geometric model of the target area, wherein the geometric model includes the target solar panel;

[0012] By using the geometric model, the flow path of rainwater at different opening and closing angles of each target solar panel is simulated;

[0013] Determining a closed angle interval of each of the target solar panels according to the flow path;

[0014] Determining whether the opening angle is within the closing angle interval;

[0015] When the opening and closing angle is within the closing angle interval, controlling the controller corresponding to each of the target solar panels so that each of the target solar panels is closed according to the minimum closing angle in the closing angle interval;

[0016] The pre-trained fusion model is obtained by fusing the LSTM model with the CNN model.

[0017] In a possible implementation manner of the first aspect, the CNN model includes a first fully connected layer and multiple convolutional layers, the LSTM model includes an LSTM layer and a second fully connected layer, and the fusion step of the pre-trained fusion model includes:

[0018] Determining a target convolutional layer among all the convolutional layers of the CNN model;

[0019] Dynamically pruning the target convolutional layer of the CNN model to remove redundant convolution kernels;

[0020] Performing a pooling operation on the output of the pruned target convolutional layer to obtain a pooling layer output;

[0021] Flatten the output of the pooling layer and input it into the first fully connected layer, and flatten the output of the LSTM layer and input it into the second fully connected layer;

[0022] The output of the first fully connected layer is fused with the output of the second fully connected layer to achieve fusion of the pre-trained fusion model.

[0023] In a possible implementation manner of the first aspect, determining a target convolutional layer among all the convolutional layers of the CNN model includes:

[0024] Analyze all convolutional layers of the CNN model layer by layer, obtain the output feature map of each convolutional layer, determine the redundancy of each output feature map, and determine the importance score of each convolutional layer according to each output feature map, wherein the redundancy of each output feature map is obtained based on the comparison of the information entropy of each output feature map with an entropy threshold, and the importance score of each convolutional layer is obtained based on the activation value of the corresponding output feature map;

[0025] For any one of the convolutional layers of the CNN model, when the importance score is greater than a preset score and the redundancy of the corresponding output feature map is less than a preset redundancy, the convolutional layer is determined to be the target convolutional layer.

[0026] In a possible implementation manner of the first aspect, dynamically pruning the target convolution layer of the CNN model to remove redundant convolution kernels includes:

[0027] Adding a gating network before each of the target convolutional layers so that each of the target convolutional layers selects a convolution kernel according to the corresponding gating network;

[0028] The acquired training set data is used to simultaneously train the gated network and the CNN model, wherein each target convolution layer in the trained CNN model removes the redundant convolution kernel according to the corresponding gated network.

[0029] In a possible implementation of the first aspect, the gating network includes an input layer, a global pooling layer, a third fully connected layer, an attention mechanism layer, and a decision layer;

[0030] Among them, the input layer is used to receive the activation value of each channel of the corresponding target convolution layer, the global pooling layer is used to perform global maximum pooling on the activation value of each channel and generate global features, the third fully connected layer is used to perform nonlinear transformation on the global features and extract important features, the attention mechanism layer is used to calculate the attention score of each channel using self-attention, and the decision layer is used to generate a gating signal according to the attention score of each channel, and the gating signal is used to decide whether to retain the convolution kernel corresponding to each channel.

[0031] In a possible implementation manner of the first aspect, inputting the indoor temperature, the weather data, the solar power generation data, and the solar panel image into a pre-trained fusion model to obtain a solar panel optimization strategy includes:

[0032] Inputting the indoor temperature, the weather data, and the solar power generation data into the LSTM layer of the LSTM model respectively to obtain a first output;

[0033] Flattening the first output and inputting the flattened first output into the second fully connected layer to obtain a second output;

[0034] Inputting the solar panel image into the target convolution layer and the pooling layer of the CNN model to obtain a third output;

[0035] Flattening the third output and inputting the flattened third output into the first fully connected layer to obtain a fourth output;

[0036] Fusing the second output with the fourth output to obtain a fused output;

[0037] The fusion output is input into the decision layer of the pre-trained fusion model to generate a solar panel optimization strategy.

[0038] In a possible implementation of the first aspect, simulating, by using the geometric model, a flow path of rainwater at different opening and closing angles of each target solar panel includes:

[0039] Dividing the geometric model into a computational grid, wherein each grid of the computational grid is used to represent a computational domain;

[0040] Setting boundary conditions for fluid entering each of the computational domains, the boundary conditions including inlet boundary conditions, outlet boundary conditions and wall boundary conditions;

[0041] According to the inlet boundary condition, the outlet boundary condition and the wall boundary condition, CFD software is used to solve the Navier-Stokes equations to obtain the flow path of rainwater at different opening and closing angles of each target solar panel.

[0042] In a possible implementation manner of the first aspect, the method further includes:

[0043] When the opening and closing angle is not within the closing angle interval, the controller corresponding to each of the target solar panels is controlled to make each of the target solar panels close according to the maximum closing angle in the closing angle interval.

[0044] In a possible implementation manner of the first aspect, the solar power generation system further includes an inverter and a combiner box, each of the target areas is provided with one of the inverter and the combiner box, and the method further includes:

[0045] Obtaining electricity demand data for each target area in a preset time period in the future and solar power generation data at the current moment;

[0046] Calculate the power difference of each target area according to the power demand data of each target area in a preset time period in the future and the solar power generation data at the current moment, wherein the power difference is the difference between the power demand data and the solar power generation data;

[0047] When the power consumption difference of one of the target areas is greater than 0, the inverter is used to call the combiner box of the target area where the power consumption difference is less than a preset value to achieve power balance.

[0048] In a second aspect, the present application provides a solar panel control system based on a photovoltaic monitoring platform, comprising a service terminal, a temperature sensor, a rain sensor and a plurality of solar panels, wherein the service terminal is deployed with a photovoltaic monitoring platform, each of the solar panels is provided with a controller for adjusting the opening and closing angle of the solar panel, and the temperature sensor, the rain sensor and the plurality of controllers are all connected to the service terminal;

[0049] The service terminal is used to obtain weather data, solar power generation data within a preset time period and solar panel images of the target area in real time through the photovoltaic monitoring platform; in response to receiving the indoor temperature transmitted by the temperature sensor, the indoor temperature, the weather data, the solar power generation data and the solar panel image are input into the pre-trained fusion model to obtain a solar panel optimization strategy, and the solar panel optimization strategy is used to determine the opening and closing angle of the target solar panel; the controller of the corresponding target solar panel is controlled according to the solar panel optimization strategy so that the controller adjusts the opening and closing angle; in response to receiving the rain signal transmitted by the rain sensor, a geometric model of the target area is constructed, and the geometric model includes the target solar panel; through the geometric model, the flow path of rainwater at different opening and closing angles of each target solar panel is simulated; according to the flow path, the closing angle interval of each target solar panel is determined; it is determined whether the opening and closing angle is within the closing angle interval; when the opening and closing angle is within the closing angle interval, the controller corresponding to each target solar panel is controlled so that each target solar panel is closed according to the minimum closing angle in the closing angle interval; wherein the pre-trained fusion model is obtained based on the fusion of the LSTM model and the CNN model.

[0050] Through the above technical solution, by deploying a photovoltaic monitoring platform, automatic monitoring and intelligent control of the solar power generation system are realized. Specifically, the photovoltaic monitoring platform can obtain weather data, solar power generation data and solar panel images in real time, and combine the data of temperature sensors and rain sensors to generate an optimization strategy for solar panels through a pre-trained fusion model (based on LSTM and CNN models). The optimization strategy can automatically adjust the opening and closing angles of solar panels to protect the safety of equipment. On sunny days with high indoor temperatures, the angle of the solar panel can be automatically adjusted according to the temperature and weather data to drive the window to open, thereby reducing the indoor temperature and ensuring the safety of the equipment. On rainy days, the geometric model can be used to simulate the flow path of rainwater at different opening and closing angles, determine the optimal closing angle range, and automatically adjust the angle of the solar panel to close the window to prevent rainwater from entering the room. The above-mentioned automatic control of solar panels not only reduces the need for human intervention and improves operating efficiency, but also can better protect the safety of equipment under complex weather conditions. Especially when the indoor temperature is high and it is rainy, the technical solution can comprehensively consider the temperature and rain factors, intelligently adjust the angle of the solar panel, ensure that the equipment is not damaged, and effectively overcome the shortcomings of traditional methods.

[0051] Other features and advantages of the embodiments of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of a flow chart of a solar panel control method based on a photovoltaic monitoring platform provided in an embodiment of the present application;

[0053] Figure 2 A schematic diagram of a three-dimensional model of a computer room provided in an embodiment of the present application;

[0054] Figure 3 A schematic diagram of the architecture of a fusion model provided in an embodiment of the present application;

[0055] Figure 4 A schematic diagram of a convolutional layer of a CNN model provided in an embodiment of the present application;

[0056] Figure 5 A schematic diagram of a connection between a solar panel and a window provided in an embodiment of the present application;

[0057] Figure 6 A structural block diagram of a solar panel control system based on a photovoltaic monitoring platform provided in an embodiment of the present application.

[0058] Description of Reference Numerals

[0059] 1 Solar panels 2 window 3 Screw 4 Electric Motor 5 Sliders DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the specific implementation methods described herein are only used to illustrate and explain the embodiments of the present application, and are not used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0061] It should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back...), the directional indications are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0062] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present application, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0063] Figure 1 The schematic diagram shows a flow chart of a solar panel control method based on a photovoltaic monitoring platform according to an embodiment of the present application. Figure 1 As shown, an embodiment of the present application provides a solar panel control method based on a photovoltaic monitoring platform, which is applied to a service terminal. The service terminal is deployed with a photovoltaic monitoring platform, and the photovoltaic monitoring platform is used to monitor a solar power generation system. The solar power generation system includes a temperature sensor, a rain sensor and multiple solar panels. Each solar panel is provided with a controller for adjusting the opening and closing angle of the solar panel. The method may include the following steps.

[0064] S110, obtaining weather data, solar power generation data within a preset time period, and solar panel images of a target area in real time through a photovoltaic monitoring platform;

[0065] S120, in response to receiving the indoor temperature transmitted by the temperature sensor, inputting the indoor temperature, weather data, solar power generation data and solar panel image into a pre-trained fusion model to obtain a solar panel optimization strategy, where the solar panel optimization strategy is used to determine the opening and closing angle of the target solar panel;

[0066] S130, controlling the controller of the corresponding target solar panel according to the solar panel optimization strategy, so that the controller adjusts the opening and closing angle;

[0067] S140, in response to receiving a rain signal transmitted by a rain sensor, constructing a geometric model of the target area, where the geometric model includes a target solar panel;

[0068] S150, simulating the flow path of rainwater at different opening and closing angles of each target solar panel through a geometric model;

[0069] S160, determining the closed angle interval of each target solar panel according to the flow path;

[0070] S170, determining whether the opening angle is within the closing angle interval;

[0071] S180, when the opening and closing angle is within the closing angle interval, controlling the controller corresponding to each target solar panel so that each target solar panel is closed according to the minimum closing angle in the closing angle interval;

[0072] Among them, the pre-trained fusion model is obtained by fusing the LSTM model and the CNN model.

[0073] Figure 2 A schematic diagram of a three-dimensional model of a computer room provided by an embodiment of the present application is shown, wherein the solar panel 1 is linked to the computer room window 2, and when the solar panel 1 is opened or closed, the computer room window 2 is opened or closed accordingly, so that the computer room window 2 can be opened or closed by controlling the opening or closing angle of the solar panel 1. It should be noted that this embodiment can be applied not only to computer room scenarios, but also to other scenarios, and this application does not limit this.

[0074] Specifically, this embodiment obtains weather data, solar power generation data within a preset time period, and solar panel images of the target area in real time through a photovoltaic monitoring platform. The photovoltaic monitoring platform is connected to various sensors and camera equipment, and can collect weather data such as temperature, humidity, wind speed, and precipitation in real time. At the same time, the photovoltaic monitoring platform can also record solar power generation data within a preset time period, including parameters such as voltage, current, and power. Solar power generation data can be obtained through sensors and power metering equipment installed around the solar panels. In addition, the images of the solar panels in the target area are transmitted to the platform in real time through high-definition cameras to ensure that the status and position of each solar panel can be accurately monitored.

[0075] In response to receiving the indoor temperature transmitted by the temperature sensor, the indoor temperature, weather data, solar power generation data and solar panel image are input into the pre-trained fusion model to obtain the solar panel optimization strategy. Specifically, the fusion model is based on the fusion of the LSTM (Long Short-Term Memory Network) model and the CNN (Convolutional Neural Network) model, and can process time series data and image data at the same time. The LSTM model is used to analyze time series data, such as weather changes and power generation data, to predict future trends. The CNN model is used to process solar panel images and identify the status and position of solar panels. By inputting the indoor temperature, weather data, power generation data and image into the fusion model, an optimization strategy can be generated. The optimization strategy includes the optimal opening and closing angle of each solar panel to maximize power generation efficiency and protect equipment safety. For example, on a sunny day and when the indoor temperature is high, the optimization strategy may be to adjust the solar panel to a larger angle to increase ventilation and heat dissipation.

[0076] According to the solar panel optimization strategy, the controller of the corresponding target solar panel is controlled so that the controller adjusts the opening and closing angle. Each solar panel is equipped with a controller that can adjust the opening and closing angle of the panel according to the received instructions. The controller drives the slide bar or hinge system through a motor to realize the opening and closing action of the solar panel. According to the optimization strategy, the controller will receive a specific opening and closing angle instruction and then perform the corresponding action. For example, if the optimization strategy is to adjust the solar panel to 45 degrees, the controller drives the motor to push the slide bar to the corresponding position so that the solar panel reaches the specified angle.

[0077] In response to receiving the rain signal transmitted by the rain sensor, a geometric model of the target area is constructed. The geometric model includes information such as the position, size and angle of the target solar panel. The rain sensor can detect the rainfall and the rain falling angle in real time. When the rainfall and the rain falling angle are detected, the rain sensor sends a signal to the photovoltaic monitoring platform. After receiving the signal, the photovoltaic monitoring platform can construct a three-dimensional geometric model based on the pre-stored solar panel geometry data. The three-dimensional geometric model is used to describe the relative position, size and current opening and closing angle of each solar panel. The three-dimensional geometric model of this embodiment can accurately simulate the flow path of rainwater at different opening and closing angles.

[0078] Specifically, after the three-dimensional geometric model is established, computational fluid dynamics (CFD) can be used to simulate the flow path of rainwater. CFD simulation can accurately calculate the flow of rainwater at different angles, including the flow rate, direction and location of water accumulation. During the simulation process, the material, surface characteristics and environmental factors of the solar panel, such as wind speed and direction, need to be considered to determine whether rainwater will flow into the room or accumulate in unfavorable locations at different opening and closing angles.

[0079] Through the above analysis of the rainwater flow path, it can be determined whether rainwater will enter the room or affect the equipment at different opening and closing angles. The closed angle interval refers to the interval within which the opening and closing angles of the solar panel can effectively prevent rainwater from entering the room. For example, through simulation, it was found that when the opening and closing angle of the solar panel is between 20 degrees and 30 degrees, rainwater will not enter the room, and this interval is the closed angle interval. The determination of the closed angle interval can provide a clear angle range for the subsequent control of the solar panel, ensuring that the solar panel can be effectively closed on rainy days to prevent rainwater from entering.

[0080] It is understandable that by comparing the current opening and closing angle with the closed angle interval, it can be determined whether the angle of the solar panel needs to be adjusted. If the current opening and closing angle is within the closed angle interval, no adjustment is required, and the solar panel is already at a safe angle. If the current opening and closing angle is not within the closed angle interval, the angle of the solar panel needs to be adjusted to enter the closed angle interval. For example, if the current opening and closing angle is 15 degrees, and the closed angle interval is 20 to 30 degrees, the angle of the solar panel needs to be adjusted to within this interval to ensure that rainwater does not enter the room.

[0081] When the opening and closing angle is within the closing angle interval, the controller corresponding to each target solar panel is controlled so that each target solar panel is closed according to the minimum closing angle in the closing angle interval. In order to ensure the maximum waterproof effect, the minimum closing angle in the closing angle interval is usually selected for adjustment. For example, if the closing angle interval is 20 to 30 degrees, 20 degrees is selected as the adjustment target. After receiving the instruction, the controller will drive the motor to adjust the solar panel to a position of 20 degrees to ensure that rainwater does not enter the room. In this way, the solar panel can automatically adjust to a safe angle on rainy days to prevent rainwater from entering and protect the safety of indoor equipment.

[0082] This embodiment realizes automatic monitoring and intelligent control of the solar power generation system by deploying a photovoltaic monitoring platform. Specifically, the photovoltaic monitoring platform can obtain weather data, solar power generation data and solar panel images in real time, and combine the data of temperature sensors and rain sensors to generate an optimization strategy for solar panels through a pre-trained fusion model (based on LSTM and CNN models). The optimization strategy can automatically adjust the opening and closing angles of solar panels to protect the safety of equipment. On sunny days and when the indoor temperature is high, the angle of the solar panel can be automatically adjusted according to the temperature and weather data to drive the window to open, thereby reducing the indoor temperature and ensuring the safety of the equipment. On rainy days, the geometric model can be used to simulate the flow path of rainwater at different opening and closing angles, determine the optimal closing angle range, and automatically adjust the angle of the solar panel to close the window to prevent rainwater from entering the room. The above-mentioned automatic control of solar panels not only reduces the need for human intervention and improves operating efficiency, but also can better protect the safety of equipment under complex weather conditions. Especially when the indoor temperature is high and it is rainy, the technical solution can comprehensively consider the temperature and rain factors, intelligently adjust the angle of the solar panel, ensure that the equipment is not damaged, and effectively overcome the shortcomings of traditional methods.

[0083] In one implementation of this embodiment, the CNN model includes a first fully connected layer and multiple convolutional layers, the LSTM model includes an LSTM layer and a second fully connected layer, and the fusion step of the pre-trained fusion model includes:

[0084] S210, determining a target convolution layer among all convolution layers of the CNN model;

[0085] S220, dynamically pruning the target convolution layer of the CNN model to remove redundant convolution kernels;

[0086] S230, performing a pooling operation on the output of the pruned target convolutional layer to obtain a pooling layer output;

[0087] S240, flattening the output of the pooling layer and inputting it into the first fully connected layer, and flattening the output of the LSTM layer and inputting it into the second fully connected layer;

[0088] S250, fusing the output of the first fully connected layer with the output of the second fully connected layer to achieve fusion of the pre-trained fusion model.

[0089] Convolutional neural network (CNN) models are usually composed of multiple convolutional layers, pooling layers, and fully connected layers, each of which is used to extract different features of the input data. In order to optimize the performance of the model, it is necessary to determine the target convolutional layer, which can be selected based on the effect of feature extraction, computational cost, and overall performance of the model. For example, the middle layer can be selected as the target convolutional layer because the middle layer is usually able to extract higher-level features while having a relatively low computational cost. After determining the target convolutional layer, the next steps will focus on optimizing and processing the target convolutional layer to optimize the performance of the fusion model.

[0090] Dynamically prune the target convolution layer of the CNN model to remove redundant convolution kernels. The convolution kernels in the convolution layer are used to extract features, but not all convolution kernels contribute significantly to the final model output. There are some redundant convolution kernels. Removing redundant convolution kernels can reduce the amount of model computation and improve operating efficiency. Therefore, this embodiment analyzes the output of each convolution kernel, determines which convolution kernels contribute less, and then removes these convolution kernels to achieve dynamic pruning. A specific implementation method may include calculating the L1 norm or L2 norm of each convolution kernel. A smaller norm indicates that the output of the convolution kernel contributes less to the final result. After removing these redundant convolution kernels, the number of model parameters is reduced, the calculation speed is improved, and the accuracy of the model is maintained.

[0091] The output of the pruned target convolutional layer is pooled to obtain the pooling layer output. Pooling operations include maximum pooling and average pooling. Pooling operations are used to reduce the size of feature maps by downsampling while retaining important feature information. For example, maximum pooling selects the maximum value in each local area of ​​the feature map, while average pooling calculates the average value of the local area. Pooling operations not only reduce the amount of calculation, but also increase the translation invariance of features. After the pooling operation, the size of the feature map is greatly reduced, but important feature information is still retained.

[0092] Next, the output of the pooling layer is flattened and input to the first fully connected layer, and the output of the LSTM layer is flattened and input to the second fully connected layer. The feature map output by the pooling layer is usually multi-dimensional. In order to be input to the first fully connected layer, it needs to be flattened into a one-dimensional vector. The flattening operation refers to expanding the multi-dimensional feature map into a one-dimensional vector by row or column, retaining the order of the features. The first fully connected layer receives the flattened feature vector, performs linear transformation and non-linear activation, and further extracts high-level features. Similarly, the output of the LSTM layer is also multi-dimensional time series data, which needs to be flattened into a one-dimensional vector before being input to the second fully connected layer. The LSTM layer captures the temporal relationship in the data by analyzing the time series data, and then flattens it and inputs it to the second fully connected layer for further processing.

[0093] The output of the first fully connected layer is fused with the output of the second fully connected layer to achieve the fusion of the pre-trained fusion model. The fusion step combines the image features of the CNN model and the time series features of the LSTM model to form a comprehensive feature vector. In a specific implementation, the output vectors of the first fully connected layer and the second fully connected layer can be spliced ​​together to form a new feature vector, which is then input into the subsequent decision layer for classification or regression. Through fusion, the fusion model can use both image information and time series information to improve the accuracy and robustness of the prediction.

[0094] Figure 3 A schematic diagram of the architecture of a fusion model provided in an embodiment of the present application is shown, referring to Figure 3 , the fusion model includes a first fully connected layer, multiple convolutional layers, an LSTM layer, and a second fully connected layer, wherein the output of the first fully connected layer is connected to the output of the second fully connected layer to achieve the fusion of the CNN model and the LSTM model. In other words, the fusion model combines the advantages of the CNN model and the LSTM model, and can effectively improve the accuracy of prediction compared to a single model.

[0095] The fusion model of this embodiment effectively combines the advantages of CNN and LSTM to process complex multimodal data. Among them, dynamic pruning and pooling operations can effectively reduce the computational complexity of the fusion model and improve the operating efficiency. Flattening and fully connected layer processing further extract high-level features, and the fusion step can combine the features of different modalities to form a comprehensive feature vector. Ultimately, the fusion model can utilize both image information and time series information, effectively improving the accuracy and robustness of the prediction, while reducing computational costs and improving the performance of the fusion model.

[0096] In one implementation of this embodiment, determining a target convolution layer among all convolution layers of a CNN model includes the following steps:

[0097] S310, analyzing all convolutional layers of the CNN model layer by layer, obtaining an output feature map of each convolutional layer, determining the redundancy of each output feature map, and determining an importance score of each convolutional layer according to each output feature map, wherein the redundancy of each output feature map is obtained based on a comparison between the information entropy of each output feature map and an entropy threshold, and the importance score of each convolutional layer is obtained based on an activation value of the corresponding output feature map;

[0098] S320. For any convolutional layer of the CNN model, when the importance score is greater than a preset score and the redundancy of the corresponding output feature map is less than a preset redundancy, determine that the convolutional layer is a target convolutional layer.

[0099] In this embodiment, all convolutional layers of the CNN model are first analyzed layer by layer, the output feature map of each convolutional layer is obtained, the redundancy of each output feature map is determined, and the importance score of each convolutional layer is determined according to each output feature map. The output feature map generated by each convolutional layer in the convolutional neural network (CNN) contains a lot of information, but not all information is useful. Layer-by-layer analysis is used to evaluate the contribution and redundancy of each convolutional layer. First, the output feature map of each convolutional layer is obtained, and the output feature map is the result after the convolution kernel is processed. Then, the information entropy of each feature map is calculated. The information entropy is used to measure the amount of information. The higher the entropy value, the greater the amount of information. The calculated information entropy is compared with the preset entropy threshold to determine the redundancy of the feature map. If the information entropy is lower than the entropy threshold, the feature map contains more redundant information. Next, the importance score of each convolutional layer is evaluated, which can be achieved by calculating the activation value of the feature map. The activation value reflects the degree of response of the convolutional layer to the input data. The higher the activation value, the greater the contribution of the convolutional layer to feature extraction. In summary, each convolutional layer can be assigned a redundancy and an importance score to provide a basis for the subsequent selection of the target convolutional layer.

[0100] For any convolutional layer of the CNN model, if the importance score is greater than the preset score and the redundancy of the corresponding output feature map is less than the preset redundancy, the convolutional layer is determined to be the target convolutional layer.

[0101] Among them, the selection of the target convolutional layer is based on two key indicators: importance score and redundancy. Specifically, the importance score reflects the contribution of the convolutional layer in feature extraction, while the redundancy measures the amount of redundant information in the feature map. In a specific implementation, a preset score threshold and a preset redundancy threshold can be set as a criterion for selecting the target convolutional layer. For each convolutional layer, if its importance score is higher than the preset score threshold, it means that the convolutional layer has a greater contribution to feature extraction; at the same time, if the redundancy of its feature map is lower than the preset redundancy threshold, it means that there is less redundant information in the feature map. Only convolutional layers that meet both conditions can be determined as target convolutional layers. This embodiment can ensure that the selected target convolutional layer has both a higher feature extraction capability and contains less redundant information.

[0102] This implementation can effectively determine the target convolution layer in the CNN model, providing a basis for subsequent dynamic pruning and optimization. The method of layer-by-layer analysis and calculation of redundancy and importance scores can effectively ensure the scientificity and rationality of the selection process. By setting a preset score and a preset redundancy threshold, the optimal convolution layer can be screened out, which not only retains the feature extraction ability of the model, but also reduces redundant information, which not only improves the computational efficiency of the fusion model, but also can achieve the improvement of model performance and the reduction of computational costs.

[0103] In one implementation of this embodiment, dynamically pruning the target convolution layer of the CNN model to remove redundant convolution kernels includes the following steps:

[0104] S410, adding a gating network before each target convolution layer, so that each target convolution layer selects a convolution kernel according to the corresponding gating network;

[0105] S420, using the acquired training set data to simultaneously train the gated network and the CNN model, wherein each target convolution layer in the trained CNN model removes redundant convolution kernels according to the corresponding gated network.

[0106] In this embodiment, a gating network is first added before each target convolution layer, so that each target convolution layer selects a convolution kernel according to the corresponding gating network. Specifically, Figure 4 A convolutional layer diagram of a CNN model provided in an embodiment of the present application is shown, Figure 4 As shown in the figure, there is a gating network before each convolutional layer in the CNN model.

[0107] The gating network is an additional neural network structure used to dynamically control the use of convolution kernels. The gating network can be a fully connected layer or a subnetwork, and its output is used to decide which convolution kernels will be activated and which will be disabled. The input of the gating network can be the input features of the target convolution layer. By processing these input features, the gating network can output a set of gating signals, which correspond to the convolution kernels one by one. Each gating signal is a value between 0 and 1, indicating the enable probability of the corresponding convolution kernel. For example, if a gating signal is close to 1, the corresponding convolution kernel will be fully enabled; if it is close to 0, the convolution kernel will be disabled or partially enabled. In this way, the gating network can dynamically adjust the use of convolution kernels in each target convolution layer to ensure that only convolution kernels that contribute significantly to feature extraction are retained.

[0108] If the input feature map has C channels, the gated network in this embodiment has C input nodes and C output nodes, and the input is nonlinearly transformed through the weight matrix and activation function to extract important features. For example, the input vector is [x1, x2, ..., xC], and the fully connected layer output is [y1, y2, ..., yC].

[0109] Secondly, the acquired training set data is used to simultaneously train the gated network and the CNN model, wherein each target convolution layer in the trained CNN model removes redundant convolution kernels according to the corresponding gated network. During the training process, the gated network and the CNN model are jointly optimized to achieve the best pruning effect. A specific implementation method may adopt a joint loss function, which includes the standard loss of the CNN model and the regularization loss of the gated network. The regularization loss is used to encourage the gated network to output sparse gated signals, thereby reducing the number of redundant convolution kernels.

[0110] For example, L1 regularization can be used for gating signals so that most signals are close to 0, thereby disabling the corresponding convolution kernels. During the training process, the standard back-propagation algorithm is used to update the parameters of the CNN model and the gating network. The training set data passes through the gating network and the CNN model through forward propagation, and the parameters are updated through back-propagation after the loss is calculated. After multiple rounds of iterative training, the gating network can gradually learn how to select the most useful convolution kernels, and the CNN model can gradually adapt to this selection mechanism. In the end, the trained gating network can effectively remove redundant convolution kernels and only retain the convolution kernels that contribute significantly to feature extraction.

[0111] The introduction of the gated network in this embodiment enables each target convolution layer to dynamically adjust the use of convolution kernels according to actual needs, avoiding the limitations of fixed pruning strategies. During the joint training process, the gated network and the CNN model achieve the best pruning effect by adapting to each other. In the trained CNN model, each target convolution layer only retains the convolution kernels that contribute significantly to feature extraction, and removes redundant convolution kernels, thereby reducing the computational complexity of the model and improving operational efficiency. Ultimately, dynamic pruning not only optimizes the structure of the model, but also improves the performance and computational efficiency of the model.

[0112] In one implementation of this embodiment, the gated network includes an input layer, a global pooling layer, a third fully connected layer, an attention mechanism layer, and a decision layer;

[0113] Among them, the input layer is used to receive the activation value of each channel of the corresponding target convolution layer, the global pooling layer is used to perform global maximum pooling on the activation value of each channel and generate global features, the third fully connected layer is used to perform nonlinear transformation on the global features and extract important features, the attention mechanism layer is used to calculate the attention score of each channel using self-attention, and the decision layer is used to generate a gating signal based on the attention score of each channel, and the gating signal is used to decide whether to retain the convolution kernel corresponding to each channel.

[0114] In this embodiment, the gated network includes an input layer, a global pooling layer, a third fully connected layer, an attention mechanism layer and a decision layer.

[0115] First, the input layer is used to receive the activation value of each channel of the corresponding target convolution layer. The output feature map of each convolution layer contains multiple channels, each channel represents the output of a different feature detector. The input layer takes the activation value of each channel as input, so that the input layer can capture the overall feature information of the target convolution layer.

[0116] The global pooling layer is used to perform global maximum pooling on the activation values ​​of each channel and generate global features. Global maximum pooling is a dimensionality reduction operation that represents the features of each channel by selecting the maximum value on each channel. This method can effectively reduce the data dimension while retaining the most significant feature information in each channel. Specifically, for the activation value matrix of each channel, the maximum value is taken as the global feature of the channel. The global feature constitutes a feature vector that contains important information of all channels. The use of the global pooling layer can significantly reduce the computational complexity while ensuring that key features are not lost, providing concise and effective input for feature extraction and decision-making in subsequent steps.

[0117] The third fully connected layer is used to perform nonlinear transformation on global features and extract important features. The fully connected layer processes the input global features through weight matrices and activation functions to extract higher-level feature information. The specific implementation method can use nonlinear activation functions such as ReLU (Rectified Linear Unit) to make the output features more expressive. Through the processing of the fully connected layer, the global features are mapped to a new feature space. In the new feature space, important features are further highlighted and redundant information is weakened to transform global features into more compact and meaningful feature representations, providing high-quality input for the calculation of the attention mechanism layer.

[0118] The attention mechanism layer is used to calculate the attention score of each channel using self-attention. The self-attention mechanism assigns importance weights to different features by calculating the similarity between features. Specifically, the attention mechanism layer first calculates the similarity score of each channel feature with other channel features, and then generates an attention score through a normalization operation. The attention score represents the importance of each channel in the overall feature representation. The self-attention mechanism can capture the dependencies and interactions between different channels, so as to more accurately evaluate the contribution of each channel. At the same time, the attention mechanism layer can identify the channels that contribute most to feature extraction and provide a basis for the decision layer.

[0119] The decision layer is used to generate a gating signal based on the attention score of each channel. The gating signal is used to decide whether to retain the convolution kernel corresponding to each channel. A specific implementation method may be to compare the attention score with a preset threshold. If the attention score of a channel is higher than the threshold, a high-value gating signal is generated, indicating that the convolution kernel corresponding to the channel is retained; otherwise, a low-value gating signal is generated, indicating that the convolution kernel is disabled or partially disabled. In this way, the decision layer can dynamically adjust the use of convolution kernels, retain the most useful convolution kernels according to actual needs, and remove redundant convolution kernels. The generation of gating signals is the core step of the entire gating network. Through precise decision-making, the structure and performance of the CNN model can be significantly optimized.

[0120] The gated network of this embodiment realizes dynamic pruning of the target convolution layer through the collaborative work of the input layer, global pooling layer, third fully connected layer, attention mechanism layer and decision layer. Each step processes and optimizes the feature information at different levels to ensure that the final generated gated signal can accurately reflect the actual contribution of each convolution kernel. In this way, the gated network not only improves the computational efficiency of the CNN model, but also effectively enhances the feature extraction capability of the model.

[0121] In one implementation of this embodiment, indoor temperature, weather data, solar power generation data, and solar panel images are input into a pre-trained fusion model to obtain a solar panel optimization strategy, including the following steps:

[0122] S510, inputting the indoor temperature, weather data, and solar power generation data into the LSTM layer of the LSTM model respectively to obtain a first output;

[0123] S520, flatten the first output and input the flattened output into a second fully connected layer to obtain a second output;

[0124] S530, inputting the solar panel image into the target convolution layer and pooling layer of the CNN model to obtain a third output;

[0125] S540, flatten the third output and input the flattened output into the first fully connected layer to obtain a fourth output;

[0126] S550, fusing the second output with the fourth output to obtain a fused output;

[0127] S560: Input the fusion output into the decision layer of the pre-trained fusion model to generate a solar panel optimization strategy.

[0128] The indoor temperature, weather data and solar power generation data are input into the LSTM layer of the LSTM model respectively. Among them, the indoor temperature, weather data and solar power generation data are all time series data with time correlation. The LSTM layer receives the input data and can selectively remember or forget the historical information through the operation of the input gate, forget gate and output gate, thereby generating the first output. The first output contains the time series characteristics of the input data.

[0129] The first output is flattened and fed into the second fully connected layer. The flattening operation converts the multidimensional tensor output by the LSTM layer into a one-dimensional vector so that it can be fed into the fully connected layer. The second fully connected layer uses the weight matrix and activation function to perform linear and nonlinear transformations on the flattened vector to extract higher-level features. Through the processing of the fully connected layer, the first output is converted into the second output, which contains the comprehensive features of indoor temperature, weather data, and solar power generation data.

[0130] At the same time, the solar panel image is input into the target convolution layer and pooling layer of the CNN model. CNN (convolutional neural network) is good at processing image data and can automatically extract the spatial features of the image. The target convolution layer extracts the local features of the solar panel image through convolution operations and performs nonlinear transformations through activation functions. The pooling layer further reduces the dimension and compresses the feature map through maximum pooling or average pooling to retain important feature information. Specifically, the convolution layer performs multiple convolution operations on the input image to generate a feature map; the pooling layer downsamples the feature map to generate a third output. The third output contains the spatial features of the solar panel image.

[0131] Then, the third output is flattened and input to the first fully connected layer. The flattening operation converts the multi-dimensional feature map after convolution and pooling into a one-dimensional vector for input to the first fully connected layer. The first fully connected layer performs linear and nonlinear transformations on the flattened vector through the weight matrix and activation function to extract higher-level image features. Through the processing of the first fully connected layer, the third output is converted into the fourth output, which contains the comprehensive features of the solar panel image.

[0132] In this embodiment, the second output is fused with the fourth output. The fusion operation refers to merging feature vectors from two different sources into a comprehensive feature vector. A specific implementation method may be to splice the second output and the fourth output by dimension to generate a new feature vector. The fused feature vector contains comprehensive information of indoor temperature, weather data, solar power generation data, and solar panel images, providing a comprehensive feature representation for the final decision.

[0133] Finally, the fusion output is input into the decision layer of the pre-trained fusion model to generate the solar panel optimization strategy. The decision layer processes the fused feature vector through a fully connected layer or other forms of classifiers to generate the final output. A specific implementation method can be to generate the probability distribution of each optimization strategy through a fully connected layer and a Softmax activation function, and select the strategy with the highest probability as the final output. The output of the decision layer is a specific solar panel optimization strategy that can be used to guide how to adjust the angle, position or other parameters of the solar panel to maximize the efficiency of solar power generation.

[0134] This implementation method can effectively combine information from multiple data sources to generate an optimized solar panel strategy by inputting indoor temperature, weather data, solar power generation data, and solar panel images into a pre-trained fusion model. By processing and optimizing the input data at different levels, it is ensured that the optimization strategy generated in the end can accurately reflect the combined impact of different data sources, so that the model not only improves the efficiency of solar power generation, but also enhances its adaptability to different environments and conditions, ultimately achieving more efficient and intelligent solar panel optimization.

[0135] In one implementation of this embodiment, the flow path of rainwater at different opening and closing angles of each target solar panel is simulated by a geometric model, including the following steps:

[0136] S610, dividing the geometric model into a computational grid, wherein each grid of the computational grid is used to represent a computational domain;

[0137] S620, setting boundary conditions for fluid entering each calculation domain, the boundary conditions including inlet boundary conditions, outlet boundary conditions and wall boundary conditions;

[0138] S630. According to the inlet boundary conditions, the outlet boundary conditions and the wall boundary conditions, the CFD software is used to solve the Navier-Stokes equations to obtain the flow path of rainwater at different opening and closing angles of each target solar panel.

[0139] The geometric model is a three-dimensional representation of the solar panel, including its shape and surface features. In order to perform fluid dynamics calculations, the continuous geometric model needs to be discretized and divided into multiple small calculation units, called grids or grid cells. Each grid cell represents a calculation domain, which is the solution area of ​​the fluid mechanics equations. The fineness of the grid division will affect the accuracy and speed of the calculation. Since complex geometric shapes and fluid behaviors require finer grids to capture details and improve calculation accuracy, in order to ensure calculation accuracy and capture details as much as possible, the calculation grid in this embodiment is as fine as possible. Grid division uses structured grids. Through grid division, the geometric model is discretized into multiple calculation domains.

[0140] Next, set the boundary conditions for the fluid to enter each computational domain. Boundary conditions refer to the state and behavior of the fluid when it enters, leaves, and contacts a solid surface. For rainwater flow simulation, boundary conditions include inlet boundary conditions, outlet boundary conditions, and wall boundary conditions. Inlet boundary conditions define parameters such as the speed, pressure, and flow rate of rainwater entering the computational domain, and are usually set according to actual rainfall conditions. Outlet boundary conditions define the state of rainwater when it leaves the computational domain, and are usually set to free outflow or fixed pressure. Wall boundary conditions define the behavior of rainwater on the surface of the solar panel, such as no-slip conditions (the velocity of the fluid on the solid surface is zero) or slip conditions (the fluid has a certain velocity on the solid surface). By setting boundary conditions, the basic rules for the fluid to enter, leave, and flow within the computational domain are determined, providing initial and boundary information for fluid dynamics solutions.

[0141] Then, according to the inlet boundary conditions, outlet boundary conditions and wall boundary conditions, CFD (computational fluid dynamics) software is used to solve the Navier-Stokes equations. The Navier-Stokes equations are the basic equations that describe fluid motion, including the mass conservation equation, momentum conservation equation and energy conservation equation. CFD software solves the Navier-Stokes equations through numerical methods to simulate the flow behavior of the fluid in the computational domain. In a specific implementation, the continuous Navier-Stokes equations are discretized into algebraic equations, and the field variables such as velocity, pressure and temperature of the fluid are obtained by iterative solution. CFD software usually includes three stages: pre-processing, solution and post-processing. The pre-processing stage includes geometric modeling, meshing and boundary condition setting; the solution stage includes equation discretization and numerical solution; the post-processing stage includes result analysis and visualization. Through the solution of CFD software, the flow path of rainwater at different opening and closing angles of each target solar panel can be obtained, which is specifically manifested as fluid velocity field, pressure field and streamline diagram.

[0142] This implementation method can simulate the flow path of rainwater at different opening and closing angles through geometric models, meshing, boundary condition settings and CFD solutions, so as to analyze and optimize the design and installation angles of solar panels, improve their performance and stability under rainy conditions, and help optimize and improve solar panels at different opening and closing angles.

[0143] In one implementation of this embodiment, the following steps are also included:

[0144] S710: When the opening and closing angle is not within the closing angle interval, control the controller corresponding to each target solar panel to make each target solar panel close according to the maximum closing angle in the closing angle interval.

[0145] When the opening and closing angle is not within the closing angle interval, the controller corresponding to each target solar panel is controlled so that each target solar panel is closed according to the maximum closing angle in the closing angle interval. When the opening and closing angle of the solar panel exceeds this range, adjustment is required to ensure its normal operation and protection. A specific implementation method can compare the current angle of the solar panel with the closing angle interval. If the current angle is out of range, the controller issues an instruction to drive the actuator to adjust the solar panel to the maximum closing angle. The above process can be achieved by an electric motor, a hydraulic cylinder or other mechanical devices.

[0146] Specifically, in this embodiment, the actuator includes a screw rod 3 and a motor 4. Figure 5 The motor 4 can drive the screw rod 3 to rotate, and the screw rod 3 drives the solar panel 1 to open and close, and the solar panel 1 is connected to the window 2 through the slide bar 5, so that the window 2 opens and closes when the solar panel 1 opens and closes.

[0147] This embodiment realizes efficient operation and intelligent management of solar panels under different environmental conditions through intelligent control of the opening and closing angles, which helps to protect the equipment in the computer room.

[0148] In one implementation of this embodiment, the solar power generation system further includes an inverter and a combiner box, each target area is provided with an inverter and a combiner box, and further includes the following steps:

[0149] S810, obtaining electricity demand data of each target area in a preset time period in the future and solar power generation data at the current moment;

[0150] S820, calculating the electricity difference of each target area according to the electricity demand data of each target area in a preset time period in the future and the solar power generation data at the current moment, wherein the electricity difference is the difference between the electricity demand data and the solar power generation data;

[0151] S830: When the power consumption difference of one of the target areas is greater than 0, the combiner box of the target area where the power consumption difference is less than a preset value is called up through the inverter to achieve power consumption balance.

[0152] Solar power generation systems include not only solar panels, but also inverters and combiner boxes. Each target area is equipped with an inverter and combiner box to manage and optimize the generation and distribution of electricity. First, it is necessary to obtain the power demand data for each target area in the future preset time period and the solar power generation data at the current moment. Power demand data is usually obtained through smart meters or energy consumption monitoring systems, which can record and predict users' power consumption in real time. The preset time period can be one day, one week or longer, depending on system requirements and user habits. Solar power generation data is obtained through sensors and data acquisition systems on solar panels to record the current power generation in real time. By obtaining these data, the power demand and power generation capacity of each target area can be fully understood, providing a basis for subsequent power allocation and management.

[0153] Next, the electricity difference of each target area is calculated based on the electricity demand data of each target area in the future preset time period and the solar power generation data at the current moment. The electricity difference refers to the difference between the electricity demand and the solar power generation, which is expressed by the formula:

[0154] ΔE=E demand-E power generation;

[0155] Among them, E demand is the electricity demand data, and E generation is the solar power generation data. If the electricity difference is a positive value, it means that the electricity demand in the area exceeds the solar power generation, and electricity needs to be supplemented from other sources; if the electricity difference is a negative value, it means that the solar power generation exceeds the electricity demand, and the excess electricity can be stored or transmitted to other areas. By calculating the electricity difference, the power profit and loss situation of each target area can be clarified, providing a basis for power allocation.

[0156] When the power difference in one of the target areas is greater than 0, the inverter is used to call the combiner box of the target area where the power difference is less than the preset value to achieve power balance. The inverter is a device that converts direct current into alternating current, and also has power allocation and management functions. When it is detected that the power difference in a certain target area is greater than 0, it means that the area is short of power and needs to allocate power from other areas. At this time, through the collaborative work of the inverter and the combiner box, cross-regional power allocation can be achieved. The specific implementation method is: the inverter first identifies the target area where the power difference is less than the preset value, that is, the area with surplus power. Then, by controlling the combiner box, the excess power in the area with surplus power is transmitted to the area with insufficient power. This process can be achieved through power transmission lines and intelligent control systems to ensure efficient transmission and balanced distribution of power between different areas. In this way, solar power generation resources can be maximized, power waste can be reduced, and the overall efficiency and stability of the system can be improved.

[0157] This implementation method can achieve intelligent management and optimization of the solar power generation system by acquiring power demand and power generation data, calculating the power difference, and implementing power allocation through inverters and combiner boxes, ensuring that the final power supply can meet the needs of different regions. In this way, the utilization rate and stability of the solar power generation system can be improved, and the dependence on traditional power resources can be reduced. At the same time, through intelligent control and power allocation, the flexibility and adaptability of the system are further enhanced, and efficient utilization and management of power resources are achieved.

[0158] Reference Figure 6 , the embodiment of the present application also provides a solar panel control system based on a photovoltaic monitoring platform, including a service terminal, a temperature sensor, a rain sensor and a plurality of solar panels, the service terminal is deployed with a photovoltaic monitoring platform, each solar panel is provided with a controller for adjusting the opening and closing angle of the solar panel, and the temperature sensor, the rain sensor and the plurality of controllers are connected to the service terminal;

[0159] The service terminal is used to obtain weather data, solar power generation data within a preset time period and solar panel images of the target area in real time through the photovoltaic monitoring platform; in response to receiving the indoor temperature transmitted by the temperature sensor, the indoor temperature, weather data, solar power generation data and solar panel images are input into the pre-trained fusion model to obtain a solar panel optimization strategy, and the solar panel optimization strategy is used to determine the opening and closing angle of the target solar panel; the controller of the corresponding target solar panel is controlled according to the solar panel optimization strategy so that the controller adjusts the opening and closing angle; in response to receiving the rain signal transmitted by the rain sensor, a geometric model of the target area is constructed, and the geometric model includes the target solar panel; through the geometric model, the flow path of rainwater at different opening and closing angles of each target solar panel is simulated; according to the flow path, the closing angle interval of each target solar panel is determined; it is determined whether the opening and closing angle is within the closing angle interval; when the opening and closing angle is within the closing angle interval, the controller corresponding to each target solar panel is controlled so that each target solar panel is closed according to the minimum closing angle in the closing angle interval; wherein the pre-trained fusion model is obtained based on the fusion of the LSTM model and the CNN model.

[0160] An embodiment of the present application also provides a machine-readable storage medium, on which instructions are stored, and the instructions are used to enable a machine to execute the above-mentioned solar panel control method based on the photovoltaic monitoring platform.

[0161] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0162] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0163] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0164] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0165] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0166] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0167] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0168] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0169] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A solar panel control method based on a photovoltaic monitoring platform, characterized in that: Applied to a service terminal, the service terminal is deployed with a photovoltaic monitoring platform, the photovoltaic monitoring platform is used to monitor a solar power generation system, the solar power generation system includes a temperature sensor, a rain sensor and a plurality of solar panels, each of the solar panels is provided with a controller for adjusting the opening and closing angle of the solar panel, each of the solar panels is connected to a window, so as to drive the window to open and close when the solar panel is opened and closed, the method includes: Acquire weather data, solar power generation data within a preset time period, and solar panel images of a target area in real time through the photovoltaic monitoring platform; In response to receiving the indoor temperature transmitted by the temperature sensor, the indoor temperature, the weather data, the solar power generation data and the solar panel image are input into a pre-trained fusion model to obtain a solar panel optimization strategy, wherein the solar panel optimization strategy is used to determine the opening and closing angle of the target solar panel; Controlling a controller of a corresponding target solar panel according to the solar panel optimization strategy so that the controller adjusts the opening and closing angle; In response to receiving a rain signal transmitted by the rain sensor, constructing a geometric model of the target area, wherein the geometric model includes the target solar panel; By using the geometric model, the flow path of rainwater at different opening and closing angles of each target solar panel is simulated; Determining a closed angle interval of each of the target solar panels according to the flow path; Determining whether the opening angle is within the closing angle interval; When the opening and closing angle is within the closing angle interval, controlling the controller corresponding to each of the target solar panels so that each of the target solar panels drives the window to close according to the minimum closing angle in the closing angle interval; The pre-trained fusion model is obtained by fusing the LSTM model with the CNN model.

2. The method according to claim 1, characterized in that The CNN model includes a first fully connected layer and a plurality of convolutional layers, the LSTM model includes an LSTM layer and a second fully connected layer, and the fusion step of the pre-trained fusion model includes: Determining a target convolutional layer among all the convolutional layers of the CNN model; Dynamically pruning the target convolutional layer of the CNN model to remove redundant convolution kernels; Performing a pooling operation on the output of the pruned target convolutional layer to obtain a pooling layer output; Flatten the output of the pooling layer and input it into the first fully connected layer, and flatten the output of the LSTM layer and input it into the second fully connected layer; The output of the first fully connected layer is fused with the output of the second fully connected layer to achieve fusion of the pre-trained fusion model.

3. The method according to claim 2, characterized in that Determining a target convolutional layer among all the convolutional layers of the CNN model comprises: Analyze all convolutional layers of the CNN model layer by layer, obtain the output feature map of each convolutional layer, determine the redundancy of each output feature map, and determine the importance score of each convolutional layer according to each output feature map, wherein the redundancy of each output feature map is obtained based on the comparison of the information entropy of each output feature map with an entropy threshold, and the importance score of each convolutional layer is obtained based on the activation value of the corresponding output feature map; For any one of the convolutional layers of the CNN model, when the importance score is greater than a preset score and the redundancy of the corresponding output feature map is less than a preset redundancy, the convolutional layer is determined to be the target convolutional layer.

4. The method according to claim 2, characterized in that: The dynamically pruning the target convolution layer of the CNN model to remove redundant convolution kernels includes: Adding a gating network before each of the target convolutional layers so that each of the target convolutional layers selects a convolution kernel according to the corresponding gating network; The acquired training set data is used to simultaneously train the gated network and the CNN model, wherein each target convolution layer in the trained CNN model removes the redundant convolution kernel according to the corresponding gated network.

5. The method according to claim 4, characterized in that The gated network includes an input layer, a global pooling layer, a third fully connected layer, an attention mechanism layer and a decision layer; Among them, the input layer is used to receive the activation value of each channel of the corresponding target convolution layer, the global pooling layer is used to perform global maximum pooling on the activation value of each channel and generate global features, the third fully connected layer is used to perform nonlinear transformation on the global features and extract important features, the attention mechanism layer is used to calculate the attention score of each channel using self-attention, and the decision layer is used to generate a gating signal according to the attention score of each channel, and the gating signal is used to decide whether to retain the convolution kernel corresponding to each channel.

6. The method according to claim 5, characterized in that The step of inputting the indoor temperature, the weather data, the solar power generation data and the solar panel image into a pre-trained fusion model to obtain a solar panel optimization strategy includes: Inputting the indoor temperature, the weather data, and the solar power generation data into the LSTM layer of the LSTM model respectively to obtain a first output; Flattening the first output and inputting the flattened first output into the second fully connected layer to obtain a second output; Inputting the solar panel image into the target convolution layer and the pooling layer of the CNN model to obtain a third output; Flattening the third output and inputting the flattened third output into the first fully connected layer to obtain a fourth output; Fusing the second output with the fourth output to obtain a fused output; The fusion output is input into the decision layer of the pre-trained fusion model to generate a solar panel optimization strategy.

7. The method according to claim 1, characterized in that The method of simulating the flow path of rainwater at different opening and closing angles of each target solar panel through the geometric model includes: Dividing the geometric model into a computational grid, wherein each grid of the computational grid is used to represent a computational domain; Setting boundary conditions for fluid entering each of the computational domains, the boundary conditions including inlet boundary conditions, outlet boundary conditions and wall boundary conditions; According to the inlet boundary condition, the outlet boundary condition and the wall boundary condition, CFD software is used to solve the Navier-Stokes equations to obtain the flow path of rainwater at different opening and closing angles of each target solar panel.

8. The method according to claim 1, characterized in that The method further comprises: When the opening and closing angle is not within the closing angle interval, the controller corresponding to each of the target solar panels is controlled to make each of the target solar panels close according to the maximum closing angle in the closing angle interval.

9. The method according to claim 1, characterized in that: The solar power generation system further includes an inverter and a combiner box, each target area is provided with one of the inverter and the combiner box, and the method further includes: Obtaining electricity demand data for each target area in a preset time period in the future and solar power generation data at the current moment; Calculate the power difference of each target area according to the power demand data of each target area in a preset time period in the future and the solar power generation data at the current moment, wherein the power difference is the difference between the power demand data and the solar power generation data; When the power consumption difference of one of the target areas is greater than 0, the inverter is used to call the combiner box of the target area where the power consumption difference is less than a preset value to achieve power balance.

10. A solar panel control system based on a photovoltaic monitoring platform, characterized in that: It includes a service terminal, a temperature sensor, a rain sensor and a plurality of solar panels. The service terminal is equipped with a photovoltaic monitoring platform. Each solar panel is provided with a controller for adjusting the opening and closing angle of the solar panel. Each solar panel is connected to a window to drive the window to open and close when the solar panel is opened or closed. The temperature sensor, the rain sensor and the plurality of controllers are connected to the service terminal. The service terminal is used to obtain weather data, solar power generation data within a preset time period and solar panel images of a target area in real time through the photovoltaic monitoring platform; in response to receiving the indoor temperature transmitted by the temperature sensor, the indoor temperature, the weather data, the solar power generation data and the solar panel image are input into a pre-trained fusion model to obtain a solar panel optimization strategy, and the solar panel optimization strategy is used to determine the opening and closing angle of the target solar panel; the controller of the corresponding target solar panel is controlled according to the solar panel optimization strategy so that the controller adjusts the opening and closing angle; in response to receiving the rain signal transmitted by the rain sensor, the target A geometric model of an area, wherein the geometric model includes the target solar panels; through the geometric model, simulating the flow path of rainwater at different opening and closing angles of each target solar panel; determining the closing angle interval of each target solar panel according to the flow path; determining whether the opening and closing angle is within the closing angle interval; when the opening and closing angle is within the closing angle interval, controlling the controller corresponding to each target solar panel so that each target solar panel drives the window to close according to the minimum closing angle in the closing angle interval; wherein the pre-trained fusion model is obtained based on the fusion of the LSTM model and the CNN model.

Images