Photovoltaic area ecological management effect prediction method based on artificial intelligence

Abstract

The invention discloses a photovoltaic area ecological management effect prediction method based on artificial intelligence, and relates to the technical field of ecological management. The method comprises the following steps: dividing a two-dimensional grid of a photovoltaic area, arranging a sensor for the photovoltaic area to acquire soil data of a key area, and obtaining soil data of photovoltaic grid points through an inverse distance weighting method; red light and near-infrared band reflectivity is obtained through an unmanned aerial vehicle remote sensing image, and a vegetation index is calculated; obtaining a photovoltaic grid point comprehensive tensor by combining the photovoltaic grid point soil data and the vegetation index; inputting the photovoltaic grid point comprehensive tensor into a convolutional neural network model based on a time attention mechanism, and outputting to obtain an ecological characteristic evolution tensor; inputting the ecological characteristic evolution tensor into an ecological management effect prediction model, and outputting to obtain an ecological management prediction tensor; and calculating the difference between the ecological management real-time tensor and the ecological management prediction tensor to obtain an ecological management error index so as to optimize the parameters of the ecological management effect prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecological management, in particular to a photovoltaic area ecological management effect prediction method based on artificial intelligence. BACKGROUND

[0002] With the transformation of global energy structure to clean energy, the photovoltaic industry is entering a period of rapid development. The construction of large-scale photovoltaic areas not only promotes the greenization of energy, but also faces the challenges of ecological protection and management. The dynamic changes of ecological elements such as soil quality, vegetation coverage and microclimate in photovoltaic areas directly affect the regional ecological stability. Therefore, accurately predicting the ecological management effect of photovoltaic areas has become a key link to ensure the sustainable development of the photovoltaic industry and the coordinated promotion of ecological protection.

[0003] The ecological management effect of photovoltaic areas is mainly evaluated by traditional monitoring and analysis methods. Soil moisture, temperature and other data are collected by manually deploying sensors, and vegetation growth information is obtained by regular field sampling. Statistical models are used to analyze soil moisture, temperature and other data to infer the ecological management effect.

[0004] However, the existing methods lack the ability to model the time series of the dynamic evolution of soil, vegetation and other multi-dimensional parameters, and cannot effectively integrate the dynamic change rules of soil, vegetation and other ecological data in the time dimension, resulting in low accuracy of ecological management effect prediction and difficulty in meeting the needs of fine management of large-scale photovoltaic areas. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a photovoltaic area ecological management effect prediction method based on artificial intelligence to solve the problems in the background art.

[0006] To achieve the above purpose, the present application realizes the following technical scheme: a photovoltaic area ecological management effect prediction method based on artificial intelligence, comprising the following steps: Step S1: divide the photovoltaic area into a grid to obtain a photovoltaic area two-dimensional grid; deploy sensors in the photovoltaic area two-dimensional grid, and collect key area soil data through the sensors; based on the key area soil data, fill in each point in the photovoltaic area two-dimensional grid by inverse distance weighting method, and calculate the photovoltaic grid point soil data; Step S2: scan the photovoltaic area by a multi-spectral camera carried by a UAV to obtain remote sensing image data, and obtain red light band reflectance and near-infrared band reflectance by frequency band decomposition of the remote sensing image data; calculate the photovoltaic grid point vegetation index by combining the red light band reflectance and the near-infrared band reflectance; obtain the photovoltaic grid point comprehensive tensor by combining the photovoltaic grid point soil data and the photovoltaic grid point vegetation index; Step S3: constructing a convolutional neural network model based on a time attention mechanism, inputting the photovoltaic grid point comprehensive tensor into the convolutional neural network model, and outputting an ecological characteristic evolution tensor; Step S4: constructing an ecological management effect prediction model, inputting the ecological characteristic evolution tensor into the ecological management effect prediction model, and outputting an ecological management prediction tensor; Step S5: collecting ecological management data of the photovoltaic area in real time to obtain an ecological management real-time tensor; calculating an ecological management mean square error by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor; and calculating an ecological management channel similarity index by calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor; Step S6: calculating an ecological management error index by combining the ecological management mean square error and the ecological management channel similarity index; performing parameter optimization on the ecological management effect prediction model based on the ecological management error index, and finally obtaining an optimized ecological management effect prediction model, thereby realizing ecological management effect prediction.

[0007] Preferably, based on the key area soil data, the inverse distance weighted method is used to fill in each point in the two-dimensional grid of the photovoltaic area, and photovoltaic grid point soil data is calculated, including the following specific steps: The inverse distance weighted method is used to fill in each point in the two-dimensional grid of the photovoltaic area, and photovoltaic grid point soil data is calculated, based on the jth characteristic value in the key area soil data:

[0008] wherein, represents the jth characteristic value of the photovoltaic grid point soil data at the coordinate (h, w) in the two-dimensional grid of the photovoltaic area, is the distance from the coordinate (h, w) to the bth key area soil data, b represents the index of the bth key area soil data, and B represents the total number of key area soil data, represents the jth characteristic value of the bth key area soil data.

[0009] Preferably, the red light band reflectivity and the near-infrared band reflectivity are obtained by frequency band decomposition of the remote sensing image data, including the following steps: Each pixel in the unmanned aerial vehicle remote sensing image data records the DN value of each band, and the DN value is converted into the ground reflectivity:

[0010] wherein, is the ground reflectivity, For the radiance, converted from the DN value by the sensor gain and offset, For the distance between the sun and the earth, For the solar irradiance, For the solar zenith angle; Based on the ground reflectance data, extract the reflectance of the near-infrared band in the range of 700-1100nm in the unmanned aerial vehicle remote sensing image And the reflectance of the red light band in the range of 600-700nm .

[0011] Preferably, the photovoltaic grid point vegetation index is calculated by combining the reflectance of the red light band and the near-infrared band, including the following steps: The photovoltaic grid point vegetation index is calculated by combining the reflectance of the red light band and the near-infrared band:

[0012] Wherein, The photovoltaic grid point vegetation index at the coordinate (h, w) is represented, The near-infrared band reflectance at the coordinate (h, w) is represented, The red light band reflectance at the coordinate (h, w) is represented.

[0013] Preferably, the convolutional neural network model based on time attention mechanism is constructed, including the following steps: The convolutional neural network model based on time attention mechanism is constructed, including: spatial feature extraction branch, time feature extraction branch and feature fusion and compression module; The spatial feature extraction branch includes: convolution layer 1, convolution layer 2, spatial attention module, maximum pooling layer, convolution layer 3, convolution layer 4; Convolution layer 1: 2D convolution layer, convolution kernel 3 3, the output channel number is 64, the step is 1, the padding is 1, and the ReLU activation function is used; Convolution layer 2: 2D convolution layer, convolution kernel 3 3, the output channel number is 64, the step is 1, the padding is 1, and the ReLU activation function is used; Spatial attention module: channel attention: global average pooling-global maximum pooling-shared multi-layer perception MLP-Sigmoid activation to generate channel weight; Spatial attention: use channel average and maximum pooling to get 2 channel features-7x7 convolution-Sigmoid activation to generate spatial weight; Maximum pooling layer: pooling window 2 2, the step is 2; Convolution layer 3: 2D convolution layer, convolution kernel 3 3, output channel number 128, step 1, padding 1, ReLU activation function; Convolutional layer 4: 2D convolutional layer, convolution kernel 3 3, output channel number 128, step 1, padding 1, ReLU activation function; Spatial attention module: channel attention: global average pooling-global maximum pooling-shared multi-layer perception MLP-Sigmoid activation to generate channel weights; spatial attention: use channel average and maximum pooling to get 2 channel features-7*7 convolution-Sigmoid activation to generate spatial weights; Max pooling layer: pooling window 2 2, step 2; The time feature extraction branch: time attention layer, fully connected layer, feedforward neural network, global average pooling layer, max pooling layer; time attention layer: multi-head attention mechanism is adopted, Q, K and V are transformed in each head; fully connected layer for linear change; max pooling layer: 1D max pooling layer, pooling window 1*4, step 4; Feature fusion and compression module: time average pooling layer, splicing layer, convolutional layer 5; convolutional layer 5: convolution kernel 1*1, output channel 128.

[0014] Preferably, the photovoltaic grid point comprehensive tensor is input into the convolutional neural network model, and an ecological characteristic evolution tensor is output. The photovoltaic grid point comprehensive tensor X is input into the convolutional neural network model, and an ecological characteristic evolution tensor is output. , Wherein, H' and W' are the number of grid rows and the number of grid columns output by the convolutional neural network model, T is the time step, and D is the ecological parameter channel output by the convolutional neural network model.

[0015] Preferably, the ecological management effect prediction model is constructed, including the following specific steps: The ecological management effect prediction model is constructed, and the ecological management effect prediction model includes a 3D cavity convolution module and a hierarchical time sequence module, a multi-scale feature fusion layer, a 3D high-precision upsampling layer, and a prediction output layer. The 3D cavity convolution module includes a cavity convolution layer 1, a cavity convolution layer 2, and a spatial attention module. Cavity convolution layer 1: convolution kernel 3 3 3, cavity rate (1, 2, 2), output channel number 128; cavity convolution layer 2: convolution kernel 3 3 2, void ratio (1, 2, 2), output channel number 128; spatial attention module: channel attention: global average pooling, global maximum pooling, shared multi-layer perception MLP, sigmoid activation to generate channel weight; spatial attention: use channel average and maximum pooling to get 2 channel features, 7x7 convolution, sigmoid activation to generate spatial weight; The hierarchical temporal module comprises a short-term branch and a long-term branch, the short-term branch adopts a 3-layer convolution long short-term memory network 3D-ConvLSTM, and the long-term branch adopts a 4-head attention layer and a feedforward network; 3D high-precision up-sampling layer: transpose convolution: 3x3 kernel, step 2, 128 channels, output 100x100x128; twice transpose convolution: 3x3 kernel, step 2, 64 channels; Prediction output layer: ecological management prediction tensor , H is the number of grid rows, W is the number of grid columns, K is the ecological parameter channel, is the prediction time step.

[0016] Preferably, the ecological management mean square error is calculated by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor, comprising the following specific steps: The ecological management mean square error is calculated by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor:

[0017] Wherein, RMSE represents the ecological management mean square error, is the prediction time step, t is the time step index, H is the number of grid rows, W is the number of grid columns, h is the index of the number of grid rows, w is the index of the number of grid columns, represents the ecological management prediction tensor with h grid rows and w grid columns, represents the ecological management real-time tensor with h grid rows and w grid columns.

[0018] Preferably, the ecological management channel similarity index is calculated by calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor, comprising the following specific steps: The ecological management channel similarity index is calculated by calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor:

[0019] Wherein, SSIM is the ecological management channel similarity index, is the prediction time step, t is the time step index, K is the number of channels, k is the channel index, Predicted mean value of the kth channel at time step t, Observed mean value of the kth channel at time step t, Predicted variance of the kth channel at time step t, Observed variance of the kth channel at time step t, Covariance of the kth channel at time step t, Mean zero prevention constant, default is , Variance zero prevention constant, default is .

[0020] Preferably, the ecological management error index is calculated by combining the ecological management mean square error and the ecological management channel similarity index, comprising the following specific steps: The ecological management error index is calculated by combining the ecological management mean square error and the ecological management channel similarity index:

[0021] Wherein, EGDV is the ecological management error index, is the normalized root mean square error, = , MAXSE is the maximum root mean square error, H is the number of grid rows, W is the number of grid columns, h is the index of the number of grid rows, w is the index of the number of grid columns, is the mean square error weight coefficient, is the channel similarity index weight coefficient, + =1.

[0022] The application provides a photovoltaic area ecological management effect prediction method based on artificial intelligence, relates to machine learning and deep learning technology, and has the following beneficial effects: (1) By combining the soil data and vegetation index of the photovoltaic grid point, a photovoltaic grid point comprehensive tensor is formed, and the soil properties and vegetation state of each grid point are integrated into a unified data structure, and the local characteristics of each grid point are not retained.

[0023] (2) The ecological management error index calculated by combining the ecological management root mean square error and the ecological management channel similarity index provides a multi-dimensional model performance evaluation index. The ecological management error index not only focuses on the absolute accuracy of the predicted value, but also emphasizes the performance of the predicted result in spatial structure consistency and multi-factor collaborative change trend matching degree, can more truly reflect the comprehensive performance of the model in simulating complex ecological system dynamics, avoids the spatial prediction distortion problem that may be covered by the single RMSE index, and provides a more reliable basis for precise optimization of the model.

[0024] (3) Based on the calculated ecological management error index, the parameters of the ecological management effect prediction model are optimized, and finally the optimized ecological management effect prediction model is obtained, realizing the continuous self-evolution of the model prediction ability and the self-adaptation to the environmental dynamics. The ecological management error index is used as a feedback signal to automatically adjust the model parameters such as neural network weights, learning rate, regularization coefficient, etc. The system deviation or deficiency found in the model prediction can be corrected immediately according to the latest monitoring data feedback. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Fig. 1 A step flow chart of a photovoltaic area ecological management effect prediction method based on artificial intelligence is provided for the present application. Fig. 2 A step level diagram for obtaining a photovoltaic grid point tensor in a photovoltaic area ecological management effect prediction method based on artificial intelligence is provided for the present application. Fig. 3 A step level diagram for obtaining an optimized ecological management effect prediction model in a photovoltaic area ecological management effect prediction method based on artificial intelligence is provided for the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] Please refer to Figs. 1-3 The present application provides a technical solution: a photovoltaic area ecological management effect prediction method based on artificial intelligence.

[0028] Step S1 divides the photovoltaic area into grids to obtain a photovoltaic area two-dimensional grid; sensors are arranged in the photovoltaic area two-dimensional grid, and key area soil data is collected by the sensors; based on the key area soil data, each point in the photovoltaic area two-dimensional grid is filled by inverse distance weighting method, and photovoltaic grid point soil data is calculated.

[0029] In the selected photovoltaic area, the photovoltaic area is meshed to obtain a photovoltaic area two-dimensional grid, and the mesh resolution can be determined by the terrain. The mesh resolution of a flat area is 10 m by default, and the mesh resolution of a slope or a gully area can be fine-tuned to 1-5 m. Sensors are arranged at key areas of the photovoltaic area two-dimensional grid, including bare areas under the photovoltaic area two-dimensional grid, peripheral vegetation restoration areas, and soil erosion-prone zones. Soil moisture sensors, ground temperature sensors, and wind speed sensors are arranged at the key areas, respectively. The soil moisture sensors are buried in different depths of soil layers to collect soil moisture data at different depths. The ground temperature sensors are fixed near the ground surface to record real-time ground temperature changes. The wind speed sensors are installed at a height of about 1.5 meters above the ground to continuously monitor the wind speed in the area, and finally obtain soil data in the key areas.

[0030] The photovoltaic grid point soil data is calculated by filling each point in the photovoltaic area two-dimensional grid using the inverse distance weighting method based on the jth feature value of the key area soil data.

[0031] represents the jth feature value of the photovoltaic grid point soil data at the coordinate (h, w) in the photovoltaic area two-dimensional grid, is the distance from the coordinate (h, w) to the bth key area soil data, b represents the index of the bth key area soil data, and B represents the total number of key area soil data, represents the jth feature value of the bth key area soil data.

[0032] It should be noted that the soil data of each grid point in the entire photovoltaic area two-dimensional grid is calculated based on the key area soil data collected by the sensors using the inverse distance weighting method, which gives higher weights to the grid points closer to the sensor location (inverse distance squared), and fills in the data for the areas where sensors are not arranged, solving the practical limitation that sensors cannot fully cover the photovoltaic area, and ensuring the continuity and smooth transition of spatial data.

[0033] Through the inverse distance weighting method, the photovoltaic grid point soil data of each grid point in the photovoltaic area two-dimensional grid is obtained.

[0034] Step S2: The photovoltaic area is scanned by a multi-spectral camera carried by a drone to obtain remote sensing image data. The red light band reflectance and near-infrared band reflectance are obtained by frequency band decomposition of the remote sensing image data. The photovoltaic grid point vegetation index is calculated by combining the red light band reflectance and near-infrared band reflectance. The photovoltaic grid point comprehensive tensor is obtained by combining the photovoltaic grid point soil data and the photovoltaic grid point vegetation index.​

[0035] Before scanning the photovoltaic area, the flight range is first determined according to the area topography, photovoltaic panel distribution and ecological monitoring focus, the flight height is usually set to 50-100 meters, the image resolution and coverage efficiency are balanced, and the period with stable light such as 10 am to 2 pm is selected to avoid spectral interference caused by strong light direct radiation or overcast days; then the unmanned aerial vehicle battery endurance, GPS positioning accuracy and multispectral camera state are checked, the camera is calibrated to ensure the accuracy of different waveband reflectivity data and geometric calibration to reduce the influence of lens distortion; during flight, the unmanned aerial vehicle automatically cruises according to the preset route, the multispectral camera synchronously collects images including visible light, near-infrared and other wavebands covering the vegetation sensitive waveband, which is convenient for subsequent identification of vegetation growth state, real-time monitoring of flight trajectory and data storage, and key areas such as photovoltaic panel gap bare land and vegetation restoration area are scanned; after flight, the equipment is recovered and the remote sensing image data is exported.

[0036] It should be noted that if the pixel resolution of the unmanned aerial vehicle image and the photovoltaic grid resolution are inconsistent, the bilinear interpolation (pixel < grid) or pixel aggregation (pixel > grid) method is adopted to ensure that one photovoltaic grid point corresponds to one vegetation index value.

[0037] Each pixel in the unmanned aerial vehicle remote sensing image data records the DN value of each waveband, which is converted into ground reflectivity to eliminate sensor noise and atmospheric influence:

[0038] Among them, is the ground reflectivity, is the radiance, which is converted from the DN value through the sensor gain and offset, is the distance from the earth to the sun, is the solar irradiance, is the solar zenith angle, which is calculated by imaging time and geographic location.

[0039] It should be noted that the radiance L, L=G*DN+B, DN is the original digital value, G is the gain coefficient, which is obtained through laboratory calibration, and B is the offset.

[0040] For the ground reflectivity data after radiometric correction, the reflectivity values of the preset near-infrared waveband (usually corresponding to the range of 700-1100 nm) and the red waveband (usually corresponding to the range of 600-700 nm) in the unmanned aerial vehicle remote sensing image are extracted, respectively, and denoted as and , represents the near-infrared waveband reflectivity at coordinate (h, w), represents the red waveband reflectivity at coordinate (h, w).

[0041] The photovoltaic grid point vegetation index is calculated by combining the red light band reflectivity and the near-infrared band reflectivity.

[0042] wherein, represents the photovoltaic grid point vegetation index at the coordinate (h, w), represents the near-infrared band reflectivity at the coordinate (h, w), represents the red light band reflectivity at the coordinate (h, w).

[0043] It should be noted that the photovoltaic grid point vegetation index is calculated by combining the red light band reflectivity and the near-infrared band reflectivity, the red light band (600nm-700nm or so) reflects the chlorophyll absorption characteristics, and the near-infrared band (700-1100nm) is highly sensitive to leaf cell structure, and the normalized ratio of the reflectivity of the two can effectively eliminate the changes of light, soil background interference, highlight the physiological characteristics of vegetation, and thus convert the complex spectral information into vegetation index, The value range of the vegetation index is [-1, 1], and the greater the positive value represents the higher the vegetation coverage.

[0044] After the photovoltaic grid point soil data and the photovoltaic grid point vegetation index are normalized by maximum and minimum values, the photovoltaic grid point comprehensive tensor O is formed by combination, O , H is the number of grid rows, W is the number of grid columns, K is the ecological parameter channel, and T is the time step, the ecological parameter channel includes the photovoltaic grid point soil data and the photovoltaic grid point vegetation index.

[0045] Step S3: constructing a convolutional neural network model based on a time attention mechanism, inputting the photovoltaic grid point comprehensive tensor into the convolutional neural network model, and outputting a ecological feature evolution tensor.

[0046] The convolutional neural network model based on the time attention mechanism includes a spatial feature extraction branch, a time feature extraction branch, and a feature fusion and compression module.

[0047] The spatial feature extraction branch includes a convolutional layer 1, a convolutional layer 2, a spatial attention module, a maximum pooling layer, a convolutional layer 3, and a convolutional layer 4.

[0048] Convolutional layer 1: 2D convolutional layer, convolution kernel 3 3, the number of output channels is 64, the step is 1, the padding is 1, and the ReLU activation function.

[0049] Convolutional layer 2: 2D convolutional layer, convolution kernel 3 3, the number of output channels is 64, the step is 1, the padding is 1, and the ReLU activation function.

[0050] Spatial attention module: Channel attention: global average pooling-global max pooling-shared multi-layer perceptron (MLP) (2 layers, neuron number 64 and 32)-Sigmoid activation to generate channel weights; Spatial attention: 2-channel features using channel average and max pooling-7x7 convolution-Sigmoid activation to generate spatial weights.

[0051] Max pooling layer: pooling window 2 2, stride 2.

[0052] Convolution layer 3: 2D convolution layer, convolution kernel 3 3, output channel number 128, stride 1, padding 1, ReLU activation function.

[0053] Convolution layer 4: 2D convolution layer, convolution kernel 3 3, output channel number 128, stride 1, padding 1, ReLU activation function.

[0054] Spatial attention module: Channel attention: global average pooling-global max pooling-shared multi-layer perceptron (MLP) (2 layers, neuron number 64 and 32)-Sigmoid activation to generate channel weights; Spatial attention: 2-channel features using channel average and max pooling-7x7 convolution-Sigmoid activation to generate spatial weights.

[0055] Max pooling layer: pooling window 2 2, stride 2.

[0056] Spatial global average pooling is performed on the photovoltaic grid point synthesis tensor, and is flattened into K The time sequence of T is input into the time feature extraction branch. The time feature extraction branch includes a time attention layer, a fully connected layer, a feedforward neural network, a global average pooling layer, and a max pooling layer. The time attention layer (number of heads 4, total dimension of Q and K 0.1): adopts a multi-head attention mechanism, each head performs Q, K, and V transformation; the fully connected layer performs linear transformation; the max pooling layer is a 1D max pooling layer, the pooling window is 1*4, the stride is 4, and the scope is the T dimension of the time sequence, which compresses the time step to .

[0057] Feature fusion and compression module: the output of the time feature extraction branch is 1 1 128, which is expanded to by broadcasting operation, and is spliced with the output of the spatial feature extraction branch, and then compressed by convolution layer 5. Time average pooling layer, splicing layer, convolution layer 5. Convolution layer 5: convolution kernel 1x1, output channel 128.

[0058] The photovoltaic grid point comprehensive tensor X is input into the convolutional neural network model, and the ecological characteristic evolution tensor is output. , , where H' and W' are the number of grid rows and grid columns output by the convolutional neural network model, H'= ,W'= (After two times of convolutional neural network model 2 pooling), T is the time step, and D is the ecological parameter channel (128 channels) output by the convolutional neural network model.

[0059] It should be noted that a convolutional neural network model based on the temporal attention mechanism (including spatial feature extraction branches and temporal feature extraction branches) is constructed to perform deep spatiotemporal feature fusion on the input photovoltaic grid point comprehensive tensor (including soil data and vegetation index of each grid point under time step T), and its output is the ecological feature evolution tensor , where each grid point (h', w') at each time step t∈T generates a 128-dimensional abstract feature vector (such as vegetation response pattern or water diffusion trend), providing a high-dimensional dynamic representation for subsequent predictions.

[0060] Step S4: construct an ecological governance effect prediction model, input the ecological characteristic evolution tensor into the ecological governance effect prediction model, and output an ecological governance prediction tensor.

[0061] An ecological governance effect prediction model is constructed, which includes a 3D void convolution module and a layered time series module, a multi-scale feature fusion layer, a 3D high-precision upsampling layer, and a prediction output layer.

[0062] The 3D dilated convolution module includes: dilated convolution layer 1, dilated convolution layer 2 and a spatial attention module.

[0063] Atrous convolution layer 1: convolution kernel 3 3 3, dilation rate (1,2,2), number of output channels 128; dilated convolution layer 2: convolution kernel 3 3 3, void ratio (1,2,2), number of output channels 128; spatial attention module: channel attention: global average pooling, global maximum pooling, shared multi-layer perceptron MLP, sigmoid activation to generate channel weights; spatial attention: use channel average and maximum pooling to obtain 2-channel features, 7×7 convolution, sigmoid activation to generate spatial weights.

[0064] The ecological characteristic evolution tensor After inputting into the spatial attention module, the output is the spatial output tensor.

[0065] The hierarchical temporal module includes a short-term branch and a long-term branch. The short-term branch: adopts a 3-layer convolutional long short-term memory network 3D-ConvLSTM, and outputs a short-term output tensor.

[0066] The long-term branch: a 4-head attention layer and a feedforward network, and outputs a long-term output tensor.

[0067] The multi-scale feature fusion layer:

[0068] wherein, is a fusion tensor, S is a spatial output tensor, is a spatial feature weight, denotes a channel concatenation operation, is a short-term output tensor, is a long-term output tensor.

[0069] The 3D high-precision up-sampling layer: a transpose convolution: 3x3 3 kernels, a stride of (1, 2, 2), 128 channels, and an output of 100x100x128; a second transpose convolution: 3x3 3 kernels, a stride of 2, and 64 channels.

[0070] The prediction output layer: an ecological management prediction tensor , , H is the number of grid rows, W is the number of grid columns, and K is the ecological parameter channel, is a prediction time step.

[0071] It should be noted that by constructing an ecological management effect prediction model of a multi-level structure (including a 3D hollow convolution module to capture spatial multi-scale information and a hierarchical temporal module combining a short-term 3D-ConvLSTM and a long-term multi-head attention mechanism to model complex time dependence), an ecological management prediction tensor is output, so as to accurately predict the evolution trend of key ecological parameters such as vegetation coverage and soil moisture of the photovoltaic area at a future time step, and provide a dynamic and quantitative scientific basis for ecological management decision-making.

[0072] Step S5: Real-time collection of ecological management data of the photovoltaic area to obtain an ecological management real-time tensor; calculation of an ecological management mean square error by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor; and calculation of an ecological management channel similarity index by calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor.

[0073] Real-time collection of ecological management data of the photovoltaic area to obtain an ecological management real-time tensor , H is the number of grid rows, W is the number of grid columns, and K is the number of ecological parameter channels. is the prediction time step.

[0074] It should be noted that the ecological management real-time tensor is the actual data of each grid point in the HxW grid at the prediction time step of the ecological management prediction tensor. The data of each grid point includes soil moisture data after maximum and minimum value standardization, land surface temperature, wind speed, and vegetation index.

[0075] By calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor, the ecological management mean square error is calculated:

[0076] where RMSE represents the ecological management mean square error, is the prediction time step, t is the time step index, H is the number of grid rows, W is the number of grid columns, h is the index of the number of grid rows, and w is the index of the number of grid columns. represents the ecological management prediction tensor with h grid rows and w grid columns, represents the ecological management real-time tensor with h grid rows and w grid columns.

[0077] By calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor, the ecological management channel similarity index is calculated:

[0078] where SSIM is the ecological management channel similarity index, is the prediction time step, t is the time step index, K is the number of channels, and k is the channel index. is the mean value of the prediction of the kth channel at time step t, is the mean value of the measured value of the kth channel at time step t, is the variance of the prediction of the kth channel at time step t, is the variance of the measured value of the kth channel at time step t, is the covariance of the kth channel at time step t, is the mean zero prevention constant, and the default is , is the variance zero prevention constant, and the default is .

[0079] It should be noted that the mean zero prevention constant and the variance zero prevention constant is to avoid the denominator tending to zero due to the mean or variance of the predicted value or the measured value being too small when calculating the ecological management channel similarity index SSIM, thereby ensuring the stability and rationality of the calculation. Their values are usually related to the dynamic range of the ecological parameters, for example, a small constant (such as the square of the maximum value of the parameter multiplied by a very small coefficient) can be set.

[0080] Step S6: Calculate the ecological management error index by combining the ecological management mean square error and the ecological management channel similarity index; optimize the parameters of the ecological management effect prediction model based on the ecological management error index, and finally obtain the optimized ecological management effect prediction model, thereby realizing ecological management effect prediction.

[0081] By combining the ecological management mean square error and the ecological management channel similarity index, the ecological management error index is calculated:

[0082] wherein, EGDV is the ecological management error index, is the normalized root mean square error, = , MAXSE is the maximum value of the root mean square error, H is the number of grid rows, W is the number of grid columns, h is the index of the number of grid rows, w is the index of the number of grid columns, is the mean square error weight coefficient, is the channel similarity index weight coefficient, + = 1.

[0083] It should be noted that the maximum value of the root mean square error is: t (t The prediction time step ) corresponds to the maximum value of the ecological management mean square error.

[0084] It should be noted that the mean square error weight coefficient and the channel similarity index weight coefficient , if the ecological management needs to be quantitatively controlled, the numerical accuracy needs to be prioritized, then 0.7, 0.3 can be defaulted. In the training process, it can be automatically adjusted once every period, for example, when the SSIM improvement rate < RMSE drop rate, the space structure learning needs to be increased by increasing the beta weight; when the SSIM improvement rate ≥ RMSE drop rate, the physical quantity precision needs to be optimized by increasing the alpha weight.

[0085] It should be noted that the ecological governance error index (EGDV) is calculated by combining the root mean square error (RMSE) of the ecological governance and the ecological governance channel similarity index (SSIM). RMSE reflects the overall degree of numerical deviation between the predicted value and the real-time value, reflecting the absolute magnitude of the error; SSIM measures the similarity of each ecological parameter channel and reflects the matching degree at the feature level. The combination of the two can comprehensively quantify the error of the prediction model, avoiding the limitation of a single numerical indicator that ignores feature similarity and compensating for the inability of a single similarity indicator to reflect absolute deviation. This provides a more comprehensive basis for parameter optimization of the ecological governance effect prediction model, ensuring that the optimized model can more accurately predict the ecological governance effect of the photovoltaic zone.

[0086] Based on the ecological governance error index, the parameters of the ecological governance effect prediction model are optimized, and finally an optimized ecological governance effect prediction model is obtained, thereby realizing ecological governance effect prediction.

[0087] It should be noted that when optimizing the parameters of the ecological governance effect prediction model based on the ecological governance error index EGDV, the goal is to minimize EGDV, the back propagation algorithm is used, and the error index is used as the core indicator of the loss function. By calculating the gradient of model parameters such as the convolution kernel weight of the void convolution layer, the hidden layer parameters of ConvLSTM in the layered time series module, the weight matrix of the Transformer attention head, the transposed convolution parameters of the upsampling layer, etc. with respect to EGDV, the parameters are iteratively updated according to the preset learning rate; at the same time, the iteration termination condition is set (such as the EGDV change for multiple consecutive rounds is less than the threshold or the maximum number of iterations is reached), and the changes of EGDV during the optimization process are monitored in real time through the validation set to avoid overfitting. Finally, the model parameters are adjusted to stably output the prediction results with the smallest difference from the real-time data, and the optimized ecological governance effect prediction model is obtained.

[0088] It should be noted that the ecological management error index EGDV is also used to optimize the parameters of the convolutional neural network model in step S3. Through the end-to-end backpropagation mechanism, the gradient of the EGDV loss function is propagated from the output layer of the ecological management effect prediction model in step S4 to the convolutional neural network model in step S3: First, calculate the gradient of EGDV on the ecological feature evolution tensor F, then update all trainable parameters in the convolutional neural network model based on the chain rule, including the convolution kernel weights and bias terms of the spatial convolution layers (Conv1-4), and the query (Q), key (K), value (V) projection matrices and multi-head attention weights in the temporal attention module; In the optimization process, an adaptive learning rate strategy (such as the Adam optimizer) is used to dynamically adjust the parameter update step, and gradient clipping (threshold set to 1.0) is added to the temporal attention layer to prevent gradient explosion in the temporal feature learning process, finally the convolutional neural network model adaptively optimizes its spatio-temporal feature extraction capability, significantly improving the capture accuracy of dynamic patterns such as vegetation index mutation and seasonal fluctuations in soil moisture.

[0089] An artificial intelligence-based intelligent prediction method for ecological management effect in photovoltaic area is proposed in this paper, aiming to solve the problems of relying on single data source, static model, lack of consideration of climate-soil-vegetation multi-factor coupling and insufficient adaptive ability in existing technology. This method integrates multi-source heterogeneous data through the construction of a multi-dimensional environmental information fusion framework, and designs an automatic extraction mechanism based on spatio-temporal key features, a multi-level spatio-temporal deep learning prediction model, and a model adaptive online updating and feedback mechanism, realizing high-precision and dynamic prediction of ecological management effects such as future vegetation coverage, soil moisture retention and ground temperature change in photovoltaic area, significantly improving the scientificity, timeliness and adaptability to complex environmental changes of the prediction.

[0090] By combining the soil data and vegetation index of the photovoltaic grid point, a comprehensive tensor of the photovoltaic grid point is formed, which integrates the soil properties and vegetation state of each grid point into a unified data structure, without preserving the local characteristics of each grid point.

[0091] By combining the ecological management root mean square error and the ecological management channel similarity index, the ecological management error index is calculated, providing a multi-dimensional model performance evaluation index. This ecological management error index not only focuses on the absolute accuracy of the predicted value, but also emphasizes the performance of the predicted result in terms of spatial structure consistency and multi-factor collaborative change trend matching degree, which can more truly reflect the comprehensive performance of the model in simulating complex ecological system dynamics, avoiding the spatial prediction distortion problem (such as average correct but spatial distribution error) that may be hidden by the single RMSE index, and providing a more reliable basis for precise optimization of the model.

[0092] Based on the calculated ecological governance error index, the parameters of the ecological governance effect prediction model are optimized, and the optimized ecological governance effect prediction model is finally obtained, so as to realize the continuous self-evolution of the model prediction ability and the self-adaptation to the environmental dynamics. This process uses the error index as a feedback signal to drive the automatic adjustment of model parameters (such as neural network weights, learning rate, regularization coefficient, etc.). It can correct the systematic bias or deficiency found in model prediction in real time according to the latest monitoring data feedback; through continuous optimization, it can enhance the generalization ability and long-term prediction stability of the model in complex and non-stationary environments; it reduces the need for frequent manual adjustment of the model, and realizes the intelligent and autonomous operation and performance maintenance of the prediction system without or with less human intervention.

[0093] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the statement "including a limited element" does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0094] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the ecological management effect of a photovoltaic area based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1: grid division is performed on the photovoltaic area to obtain a photovoltaic area two-dimensional grid; a sensor is arranged on the photovoltaic area two-dimensional grid, and key area soil data is collected through the sensor; based on the key area soil data, each point in the photovoltaic area two-dimensional grid is filled through an inverse distance weighted method, and photovoltaic grid point soil data is calculated; Step S2: the photovoltaic area is scanned by a multi-spectral camera carried by a UAV to obtain remote sensing image data, and red light band reflectivity and near-infrared band reflectivity are obtained by frequency band decomposition of the remote sensing image data; the photovoltaic grid point vegetation index is calculated by combining the red light band reflectivity and the near-infrared band reflectivity; and the photovoltaic grid point comprehensive tensor is obtained by combining the photovoltaic grid point soil data and the photovoltaic grid point vegetation index; Step S3: a convolutional neural network model based on a time attention mechanism is constructed, the photovoltaic grid point comprehensive tensor is input into the convolutional neural network model, and an ecological feature evolution tensor is output; Step S4: an ecological management effect prediction model is constructed, the ecological feature evolution tensor is input into the ecological management effect prediction model, and an ecological management prediction tensor is output; Step S5: ecological management data of the photovoltaic area is collected in real time to obtain an ecological management real-time tensor; an ecological management mean square error is calculated by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor; and an ecological management channel similarity index is calculated by calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor; Step S6: an ecological management error index is calculated by combining the ecological management mean square error and the ecological management channel similarity index; the ecological management effect prediction model is parameter-optimized based on the ecological management error index, and an optimized ecological management effect prediction model is finally obtained, so that the ecological management effect prediction is realized.

2. The method according to claim 1, characterized in that: The photovoltaic grid point soil data is calculated by filling each point in the photovoltaic area two-dimensional grid through an inverse distance weighted method based on the key area soil data, and comprises the following specific steps: The photovoltaic grid point soil data is calculated by filling each point in the photovoltaic area two-dimensional grid through an inverse distance weighted method based on the jth characteristic value in the key area soil data; ； wherein, represents the jth eigenvalue of the soil data at the photovoltaic grid point of the coordinate (h, w) in the two-dimensional grid of the photovoltaic region, is the distance from the coordinate (h, w) to the bth key region soil data, b represents the index of the bth key region soil data, and B represents the total number of key region soil data, represents the jth eigenvalue of the bth key region soil data.

3. The method of claim 2, wherein the method is based on artificial intelligence. The red light band reflectivity and the near-infrared band reflectivity are obtained by frequency band decomposition of the remote sensing image data, and comprise the following steps: Each pixel in the UAV remote sensing image data records the DN value of each band, and the DN value is converted into the ground reflectivity; ； wherein, is the surface reflectance, is the radiance, converted from the DN value by the sensor gain and offset, is the distance from the earth, is the solar irradiance, is the solar zenith angle; extracting reflectance in the near-infrared band of 700-1100 nm range and reflectance in the red band of 600-700 nm range from the UAV remote sensing image based on the surface reflectance data and reflectance in the red band of 600-700 nm range .

4. The method according to claim 3, characterized in that: The photovoltaic grid point vegetation index is calculated by combining the red light band reflectivity and the near-infrared band reflectivity, and comprises the following steps: The photovoltaic grid point vegetation index is calculated by combining the red light band reflectivity and the near-infrared band reflectivity; ； wherein, represents the vegetation index at the photovoltaic grid point of coordinates (h, w), represents the reflectance in the near-infrared band at the coordinates (h, w), represents the reflectance in the red band at the coordinates (h, w).

5. The method of claim 4, wherein the method is based on artificial intelligence. The convolutional neural network model based on the time attention mechanism is constructed, and comprises the following steps: A convolutional neural network model based on time attention mechanism is constructed, which comprises a spatial feature extraction branch, a time feature extraction branch and a feature fusion and compression module; The spatial feature extraction branch comprises a convolution layer 1, a convolution layer 2, a spatial attention module, a maximum pooling layer, a convolution layer 3 and a convolution layer 4; Convolutional layer 1: 2D convolutional layer, convolution kernel 3 3, output channel number 64, step 1, padding 1, ReLU activation function; Convolutional layer 2: 2D convolutional layer, convolution kernel 3 3, output channel number 64, step 1, padding 1, ReLU activation function; The spatial attention module comprises channel attention: global average pooling-global maximum pooling-shared multi-layer perceptron MLP-Sigmoid activation to generate channel weights; and spatial attention: using channel average and maximum pooling to obtain 2-channel features-7*7 convolution-Sigmoid activation to generate spatial weights. Max pooling layer: pooling window 2 2, stride 2; Convolutional layer 3: 2D convolutional layer, convolution kernel 3 3, output channel number 128, step 1, padding 1, ReLU activation function Convolutional layer 4: 2D convolutional layer, convolution kernel 3 3, output channel number 128, step 1, padding 1, ReLU activation function The spatial attention module comprises channel attention: global average pooling-global maximum pooling-shared multi-layer perceptron MLP-Sigmoid activation to generate channel weights; and spatial attention: using channel average and maximum pooling to obtain 2-channel features-7*7 convolution-Sigmoid activation to generate spatial weights. Max pooling layer: pooling window 2 2, stride 2; The time feature extraction branch comprises a time attention layer, a fully connected layer, a feedforward neural network, a global average pooling layer and a maximum pooling layer; the time attention layer adopts a multi-head attention mechanism, each head performs Q, K and V transformations; the fully connected layer performs linear transformation; and the maximum pooling layer is a 1D maximum pooling layer with a pooling window of 1*4 and a step of 4. The feature fusion and compression module comprises a time average pooling layer, a concatenation layer and a convolution layer 5; the convolution layer 5 has a convolution kernel of 1*1 and outputs 128 channels.

6. The method according to claim 5, wherein the method is characterized by: The photovoltaic grid point comprehensive tensor is input into the convolutional neural network model to output an ecological feature evolution tensor, comprising the following specific steps: inputting the photovoltaic grid point synthesis tensor X into the convolutional neural network model, and outputting an ecological characteristic evolution tensor , , wherein H' and W' are the number of grid rows and the number of grid columns output by the convolutional neural network model, T is a time step, and D is an ecological parameter channel output by the convolutional neural network model.

7. The method of claim 6, wherein the method is based on artificial intelligence. The ecological management effect prediction model is constructed, comprising the following specific steps: The ecological management effect prediction model is constructed, comprising a 3D cavity convolution module and a hierarchical time sequence module, a multi-scale feature fusion layer, a 3D high-precision upsampling layer and a prediction output layer. The 3D cavity convolution module comprises a cavity convolution layer 1, a cavity convolution layer 2 and a spatial attention module. Dilated convolution layer 1: kernel 3 3 3, dilation (1, 2, 2), output channel number 128; dilated convolution layer 2: kernel 3 3 3, dilation (1, 2, 2), output channel number 128; spatial attention module: channel attention: global average pooling, global max pooling, shared multi-layer perception (MLP), sigmoid activation to generate channel weights; spatial attention: use channel average and max pooling to get 2 channel features, 7x7 convolution, sigmoid activation to generate spatial weights; The hierarchical time sequence module comprises a short-term branch and a long-term branch, the short-term branch adopts a 3-layer convolution long short-term memory network 3D-ConvLSTM, and the long-term branch adopts a 4-head attention layer and a feedforward network. The 3D high-precision upsampling layer comprises a transpose convolution with a kernel of 3*3*3, a step of (1, 2, 2), 128 channels and an output of 100*100*128; and a second transpose convolution with a kernel of 3*3*3, a step of 2 and 64 channels. Predicted output layer: ecological governance prediction tensor , , H is the number of grid rows, W is the number of grid columns, K is the ecological parameter channel, is the prediction time step.

8. The artificial intelligence-based photovoltaic region ecological governance effect prediction method according to claim 7, characterized in that: The ecological management mean square error is calculated by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor, comprising the following specific steps: The ecological management mean square error is calculated by calculating the parameter difference between the ecological management real-time tensor and the ecological management prediction tensor: ； wherein RMSE represents the ecological management mean square error, for a prediction time step, t is a time step index, H is a number of grid rows, W is a number of grid columns, h is an index of the number of grid rows, w is an index of the number of grid columns, represents an ecological management prediction tensor for a grid row h and a grid column w, represents an ecological management real-time tensor for a grid row h and a grid column w.

9. The method according to claim 8, wherein the method is characterized by: The ecological management channel similarity index is calculated by calculating the channel difference between the ecological management real-time tensor and the ecological management prediction tensor, comprising the following specific steps: The ecological management channel similarity index is calculated by calculating the channel similarity between the ecological management real-time tensor and the ecological management prediction tensor: ； where SSIM is the ecological governance channel similarity index, is the predicted value mean for the kth channel at time step t, is the predicted value mean for the kth channel at time step t, is the measured value mean for the kth channel at time step t, is the predicted value variance for the kth channel at time step t, is the measured value variance for the kth channel at time step t, is the covariance for the kth channel at time step t, is the mean anti-zero constant, default is , is the variance anti-zero constant, default is .

10. The method of claim 9, wherein the method is based on artificial intelligence. The ecological management error index is calculated by combining the ecological management mean square error and the ecological management channel similarity index, and includes the following specific steps: The ecological management error index is calculated by combining the ecological management mean square error and the ecological management channel similarity index. ； wherein EGDV is an ecological governance deviation index, is a normalized root mean square error, , MAXSE is a root mean square error maximum value, H is a grid row number, W is a grid column number, h is an index of the grid row number, w is an index of the grid column number, is a mean square error weight coefficient, is a channel similarity index weight coefficient, = 1.​​

Images