Intelligent green landscape lighting adaptive control method and system based on deep learning

Abstract

This invention provides a deep learning-based adaptive control method and system for intelligent green landscape lighting. This system involves collecting parameters through environmental sensors, fusing features using a parallel residual network, constructing a layered adaptive neural network to generate lighting scenarios, and calculating the optimal control strategy using a grouped collaborative learning network to achieve lighting regulation. This method can automatically adjust lighting parameters based on environmental changes, improving visual comfort and reducing energy consumption, thereby achieving intelligent and green landscape lighting.

Description

Technical Field

[0001] The present invention relates to the field of lighting control technology, and in particular to a method and system for adaptive control of intelligent green landscape lighting based on deep learning. Background Art

[0002] With the acceleration of urbanization and rising awareness of environmental protection, smart green landscape lighting, as a crucial component of urban development, has become a key factor in enhancing a city's image and improving residents' quality of life. Traditional landscape lighting systems typically employ fixed patterns or simple time-controlled strategies, failing to intelligently adjust to environmental conditions and changes in traffic flow, resulting in energy waste and a compromised visual experience. In recent years, with the advancement of artificial intelligence and deep learning technologies, intelligent lighting control systems have begun to be applied to urban landscape lighting. By sensing environmental changes in real time and adaptively adjusting lighting parameters, they achieve a balance between energy conservation and emission reduction and visual comfort.

[0003] Existing landscape lighting systems have some problems in practical applications. Most systems lack the ability to effectively integrate multi-source environmental data and are unable to comprehensively consider the complex correlation between multi-dimensional environmental factors such as light intensity, crowd density and meteorological conditions, resulting in single lighting control decisions and insufficient adaptability; traditional lighting control algorithms are difficult to accurately evaluate visual comfort under different lighting scenarios, and fail to scientifically match human visual perception characteristics with luminous flux distribution. Some areas are too bright or too dark, affecting user experience; most existing lighting systems adopt independent control modes, lack a collaborative control mechanism between lighting units, fail to fully consider spatial distribution relationships and dynamic changes in energy consumption, and cannot achieve optimal energy utilization while ensuring visual effects. Summary of the Invention

[0004] The embodiments of the present invention provide a deep learning-based intelligent green landscape lighting adaptive control method and system, which can solve the problems in the prior art.

[0005] A first aspect of an embodiment of the present invention provides a method for adaptively controlling smart green landscape lighting based on deep learning, comprising:

[0006] Collect light intensity, crowd density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data;

[0007] The environmental parameter data is preprocessed and standardized. A parallel residual network is used to fuse spatial and temporal features to obtain multimodal features. Bidirectional mapping is used to perform transfer learning on the multimodal features. The number of factors is determined based on the dynamic association matrix, and the environmental feature vector is generated through variational inference and kernel density estimation.

[0008] A hierarchical adaptive neural network is constructed based on the environmental feature vector. Scenes are generated based on the regional illumination distribution law and the superposition relationship between the luminous flux of lighting units. Visual comfort is calculated by the matching degree between the luminous flux distribution and the human visual response characteristics. A dynamic programming algorithm is used to iteratively optimize and generate lighting scene parameters.

[0009] The lighting scene parameters are input into the group collaborative learning network, and the control groups are divided according to the spatial distribution relationship of the lighting units. The optimization mechanism of local competition within the group and global collaboration between groups is adopted, and the optimal control strategy is calculated in combination with the dynamic energy consumption prediction model.

[0010] Generate lighting equipment adjustment instructions based on the optimal control strategy and execute lighting adjustment.

[0011] In an optional embodiment, the environmental parameter data is subjected to standardized preprocessing, a parallel residual network is used to fuse spatial and temporal features to obtain multimodal features, bidirectional mapping is used to perform transfer learning on the multimodal features, the number of factors is determined based on the dynamic correlation matrix, and an environmental feature vector is generated through variational inference and kernel density estimation, including:

[0012] Preprocess and standardize environmental parameter data to obtain standardized environmental data;

[0013] The standardized environmental data is input into the spatial feature extraction branch of the parallel residual network, and the spatial features are obtained through the pyramid convolution structure. At the same time, the standardized environmental data is input into the temporal feature extraction branch of the parallel residual network, and the temporal features are obtained through the causal convolution structure. The spatial features and temporal features are fused according to the attention weight to generate multimodal nonlinear features.

[0014] The multimodal nonlinear features are input into the source domain feature space and the target domain feature space for bidirectional mapping, cycle consistency constraints are established, the uncertainty distribution of the mapped features is calculated, and the mapped features are reorganized according to the uncertainty distribution to obtain the migration features.

[0015] The migration features are input into the mutual information calculation unit to construct a dynamic correlation matrix, and the number of factors is determined from the dynamic correlation matrix using the spectral clustering method;

[0016] The number of factors is substituted into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings. The adaptive kernel density estimation is used to perform nonlinear rotation on the factor loadings and weight them based on the feature discrimination to output the environmental feature vector.

[0017] In an optional embodiment, the number of factors is substituted into the variational inference algorithm to optimize the factor loading to obtain the optimal factor loading, and the factor loading is nonlinearly rotated using adaptive kernel density estimation and weighted based on feature discrimination. The output environment feature vector includes:

[0018] The number of factors is input into the variational inference algorithm to construct a Gaussian prior distribution of the factor loads. A variational posterior approximation in the form of a normal distribution is set for the variational parameters. By minimizing the KL divergence criterion, the mean parameter and variance parameter of the variational posterior distribution are iteratively updated to obtain the optimal factor loads.

[0019] An adaptive Gaussian kernel function is constructed for the optimal factor loading. The kernel function bandwidth parameter is dynamically adjusted according to the sample distribution characteristics. The kernel density estimate of the factor loading is calculated. A rotation optimization objective function is constructed based on the kernel density estimate. The rotation matrix is ​​iteratively optimized through gradient ascent. A nonlinear rotation transformation is performed on the optimal factor loading to obtain the rotated factor loading.

[0020] The inter-class scatter matrix and intra-class scatter matrix of the rotated factor loadings are calculated, the feature discrimination is determined according to the trace ratio of the scatter matrix, the weight coefficients are set for the rotated factor loadings based on the feature discrimination, and the weighted factor loadings are input into the nonlinear feature mapping function to generate the environmental feature vector.

[0021] In an optional embodiment, a hierarchical adaptive neural network is constructed based on the environmental feature vector, and scene generation is performed according to the regional illumination distribution law and the luminous flux superposition relationship between lighting units. Visual comfort is calculated by the matching degree between the luminous flux distribution and the human visual response characteristics. The dynamic programming algorithm is used to iteratively optimize and generate lighting scene parameters, including:

[0022] Constructing a hierarchical adaptive neural network based on the environmental feature vector, dynamically adjusting the weight coefficients of the nodes in each layer of the neural network through the feature correlation matrix, and mapping and transforming the environmental feature vector to obtain a scene feature representation;

[0023] A luminous flux spatial distribution function is constructed using the scene feature representation. An initial luminous flux distribution of each lighting unit in the target area is calculated based on the luminous flux spatial distribution function. A luminous flux superposition coefficient between the lighting units is calculated based on the spatial three-dimensional coordinates, luminous direction, and shading coefficient of the lighting units. The initial luminous flux distribution and the luminous flux superposition coefficient are combined to generate a luminous flux distribution map of the target area.

[0024] Extracting spectral energy distribution characteristics from the luminous flux distribution diagram of the target area, calculating dark adaptation thresholds and light adaptation thresholds according to a predetermined spectral response curve of human cone cells, establishing a visual comfort evaluation function, and performing matching calculation on the spectral energy distribution characteristics to obtain visual comfort parameters;

[0025] The visual comfort parameters are mapped into state transition probabilities, a state transition matrix of lighting scene parameters is constructed, and dynamic programming iterative calculations are performed on the lighting scene parameters using the state transition matrix to generate optimal lighting scene parameters.

[0026] In an optional embodiment, spectral energy distribution characteristics are extracted from the luminous flux distribution diagram of the target area, dark adaptation thresholds and light adaptation thresholds are calculated based on a predetermined spectral response curve of human cone cells, a visual comfort evaluation function is established, and a matching degree calculation is performed on the spectral energy distribution characteristics to obtain visual comfort parameters, including:

[0027] A two-dimensional discrete wavelet transform is performed on the luminous flux distribution map of the target area to obtain a multi-scale spectral coefficient matrix. A Gaussian mixture distribution is fitted to the multi-scale spectral coefficient matrix, and the spectral energy probability density function is calculated and generated through expectation maximization iteration. The spectral energy distribution characteristics are determined based on the spectral energy probability density function.

[0028] The spectral response curve of human cone cells is subjected to a two-way competitive operation with multiple sets of historical visual response data. A dynamic correction coefficient is obtained through the mutual game between the response parameter and the historical parameter. A convolution operation based on the dynamic correction coefficient is performed on the spectral response curve to generate a correction response curve. The dark adaptation threshold and the light adaptation threshold are calculated based on the correction response curve.

[0029] Multi-layer feature extraction is performed on the spectral energy distribution characteristics, dark adaptation threshold and light adaptation threshold to obtain the visual energy adaptation coefficient, visual brightness response coefficient and visual time adjustment coefficient, and the visual comfort evaluation function is constructed.

[0030] The spectral energy distribution characteristics and visual comfort evaluation function are dynamically weighted and iteratively optimized to calculate the visual comfort parameters.

[0031] In an optional embodiment, lighting scene parameters are input into a group collaborative learning network, control groups are divided according to the spatial distribution relationship of lighting units, and an optimization mechanism of local competition within groups and global collaboration between groups is adopted. The optimal control strategy is calculated in combination with a dynamic energy consumption prediction model, including:

[0032] Obtain the position coordinates, luminous flux parameters, and luminous angle parameters of each lighting unit in the lighting scene parameters, calculate and determine the spatial distance matrix and the lighting influence matrix between the lighting units, perform an exponential decay operation based on the spatial distance matrix and the lighting influence matrix to obtain the correlation strength between the lighting units, and divide the lighting units with correlation strength greater than a preset correlation threshold into the same control group;

[0033] The energy consumption parameters, lighting effect parameters, and group interference parameters of each lighting unit in the control group are combined to obtain the group optimization target value. A two-way competition mechanism is used to adjust the control parameters of the lighting units against each other and the parameter values ​​are corrected through the group energy balance constraint to generate the initial control parameters of the lighting units in the group.

[0034] The intra-group optimization target value of each control group and the cross-interference degree between the control groups are combined to calculate the inter-group collaborative target value, and the inter-group collaborative target value is optimized and solved by the distributed alternating direction multiplier method to output the inter-group collaborative control parameters;

[0035] The predicted energy consumption data is obtained by extracting the time series features of historical scene parameters, load characteristic parameters, and operating status parameters. The predicted energy consumption data is dynamically weighted iteratively calculated with the inter-group collaborative control parameters to generate the optimal control strategy for the lighting unit.

[0036] In an optional embodiment, the intra-group optimization target value of each control group and the cross-interference degree between the control groups are combined and calculated to obtain the inter-group coordination target value. The inter-group coordination target value is optimized and solved using the distributed alternating direction multiplier method. The output inter-group coordination control parameters include:

[0037] The luminous flux, angle, and distance parameters of the lighting units between different control groups are calculated to obtain the inter-group cross-interference matrix. The dynamic weight matrix is ​​generated through exponential normalization operation to calculate the inter-group coupling coefficient of each control group.

[0038] The intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient are compensated and calculated to obtain the inter-group synergy target value;

[0039] The time series features of ambient temperature parameters, crowd density parameters, and meteorological change parameters are extracted to obtain a time series sensitivity matrix. Local variables, global variables, and Lagrange multipliers are introduced into the inter-group collaborative target value to construct the first augmented Lagrangian function. The time series sensitivity matrix and the first augmented Lagrangian function are compensated for the response to obtain the second augmented Lagrangian function.

[0040] For the second augmented Lagrangian function, the global variables and Lagrange multipliers are fixed, and each control group independently performs gradient descent to obtain local variables; the local variables and Lagrange multipliers are fixed, and gradient descent is performed in combination with the global constraints to obtain global variables; a bidirectional mapping operation is performed to generate a variable mapping matrix and update the Lagrange multipliers;

[0041] Perform difference calculation on local variables and global variables to obtain the original residual, calculate the difference of global variables in adjacent iterations to obtain the dual residual, adjust the penalty parameter according to the ratio of the original residual to the dual residual, and input the second augmented Lagrangian function for iterative calculation;

[0042] When the original residual is less than a first preset threshold and the dual residual is less than a second preset threshold, the iteration is terminated, and the inter-group collaborative control parameters are calculated based on the local variables and the global variables.

[0043] A second aspect of an embodiment of the present invention provides a deep learning-based intelligent green landscape lighting adaptive control system, comprising:

[0044] The first unit is used to collect light intensity, crowd density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data;

[0045] The second unit is used to perform standardized preprocessing on environmental parameter data, fuse spatial and temporal features using a parallel residual network to obtain multimodal features, perform transfer learning on multimodal features using bidirectional mapping, determine the number of factors based on the dynamic correlation matrix, and generate environmental feature vectors through variational inference and kernel density estimation;

[0046] The third unit is used to build a hierarchical adaptive neural network based on the environmental feature vector, generate scenes according to the regional illumination distribution law and the superposition relationship between the luminous flux of lighting units, calculate visual comfort by matching the luminous flux distribution with the human visual response characteristics, and use a dynamic programming algorithm to iteratively optimize and generate lighting scene parameters;

[0047] The fourth unit is used to input lighting scene parameters into the group collaborative learning network, divide the control groups according to the spatial distribution relationship of the lighting units, adopt the optimization mechanism of local competition within the group and global collaboration between groups, and combine it with the dynamic energy consumption prediction model to calculate the optimal control strategy;

[0048] The fifth unit is used to generate an adjustment instruction for the lighting device based on the optimal control strategy and perform lighting adjustment.

[0049] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0050] processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0053] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0054] In an embodiment of the present invention, through environmental sensor data acquisition and parallel residual network fusion technology, deep fusion and precise feature extraction of multi-source data are achieved, and transfer learning based on bidirectional mapping improves the accuracy and adaptability of environmental data processing, overcomes the problem of slow response of traditional lighting control methods to environmental changes, and significantly enhances perception ability and environmental adaptability; adopts a hierarchical adaptive neural network structure and a dynamic programming algorithm, realizes fine-grained regulation of lighting scenes according to human visual characteristics and luminous flux distribution relationship, realizes humanized design of light environment through visual comfort evaluation mechanism, makes lighting effect more in line with human visual needs, and improves user experience and environmental comfort; introduces group collaborative learning network and energy consumption dynamic prediction model, realizes minimization of energy consumption while ensuring lighting effect, optimizes control strategy through local competition within group and global collaboration mechanism between groups, intelligently adjusts lighting equipment parameters, reduces energy consumption compared with traditional control methods, and achieves a balance between energy saving and environmental protection and visual effect of landscape lighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a method for adaptively controlling smart green landscape lighting based on deep learning according to an embodiment of the present invention;

[0056] Figure 2 This is a bar chart comparing the performance of environmental feature extraction methods based on multimodal fusion and transfer learning according to an embodiment of the present invention;

[0057] Figure 3 The environmental feature vector generation process based on factor loading rotation optimization according to an embodiment of the present invention;

[0058] Figure 4 Schematic diagram of the comparison of visual comfort parameter response time according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0060] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1This is a flow chart of a method for adaptively controlling smart green landscape lighting based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] Collect light intensity, crowd density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data;

[0063] The environmental parameter data is preprocessed and standardized. A parallel residual network is used to fuse spatial and temporal features to obtain multimodal features. Bidirectional mapping is used to perform transfer learning on the multimodal features. The number of factors is determined based on the dynamic association matrix, and the environmental feature vector is generated through variational inference and kernel density estimation.

[0064] A hierarchical adaptive neural network is constructed based on the environmental feature vector. Scenes are generated based on the regional illumination distribution law and the superposition relationship between the luminous flux of lighting units. Visual comfort is calculated by the matching degree between the luminous flux distribution and the human visual response characteristics. A dynamic programming algorithm is used to iteratively optimize and generate lighting scene parameters.

[0065] The lighting scene parameters are input into the group collaborative learning network, and the control groups are divided according to the spatial distribution relationship of the lighting units. The optimization mechanism of local competition within the group and global collaboration between groups is adopted, and the optimal control strategy is calculated in combination with the dynamic energy consumption prediction model.

[0066] Generate lighting equipment adjustment instructions based on the optimal control strategy and execute lighting adjustment.

[0067] In an optional embodiment, the environmental parameter data is subjected to standardized preprocessing, a parallel residual network is used to fuse spatial and temporal features to obtain multimodal features, bidirectional mapping is used to perform transfer learning on the multimodal features, the number of factors is determined based on the dynamic correlation matrix, and the environmental feature vector is generated through variational inference and kernel density estimation, including:

[0068] Preprocess and standardize environmental parameter data to obtain standardized environmental data;

[0069] The standardized environmental data is input into the spatial feature extraction branch of the parallel residual network, and the spatial features are obtained through the pyramid convolution structure. At the same time, the standardized environmental data is input into the temporal feature extraction branch of the parallel residual network, and the temporal features are obtained through the causal convolution structure. The spatial features and temporal features are fused according to the attention weight to generate multimodal nonlinear features.

[0070] The multimodal nonlinear features are input into the source domain feature space and the target domain feature space for bidirectional mapping, cycle consistency constraints are established, the uncertainty distribution of the mapped features is calculated, and the mapped features are reorganized according to the uncertainty distribution to obtain the migration features.

[0071] The migration features are input into the mutual information calculation unit to construct a dynamic correlation matrix, and the number of factors is determined from the dynamic correlation matrix using the spectral clustering method;

[0072] The number of factors is substituted into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings. The adaptive kernel density estimation is used to perform nonlinear rotation on the factor loadings and weight them based on the feature discrimination to output the environmental feature vector.

[0073] In one embodiment, the collected environmental parameter data undergoes standardization preprocessing. Specifically, multidimensional environmental parameter data, including temperature, humidity, light, air pressure, and noise, is collected. Missing values ​​are filled in, outliers are processed using a moving average method, and the Z-score method is used to normalize the data in each dimension to a distribution with a mean of 0 and a standard deviation of 1. For example, for the original temperature data [25.6°C, 26.2°C, 24.9°C, 25.3°C], the normalization process yields [0.21, 1.37, -0.95, -0.63].

[0074] The standardized environmental data is fed into a parallel residual network structure, which consists of a spatial feature extraction branch and a temporal feature extraction branch. The spatial feature extraction branch uses a pyramidal convolutional structure consisting of three convolutional layers. The first layer uses 32 3×3 convolution kernels, the second layer uses 64 3×3 convolution kernels, and the third layer uses 128 3×3 convolution kernels. Each layer is followed by batch normalization and ReLU activation, while residual connections are used to preserve the original information. For five-dimensional data matrices such as temperature and humidity, spatial feature extraction results in a 128-dimensional spatial feature vector.

[0075] In the temporal feature extraction branch, a causal convolutional architecture is used to capture temporal dependencies. This architecture consists of three causal convolutional layers, each using convolution kernels with expansion rates of 1, 2, and 4, respectively, to ensure that the current output depends only on historical input. For example, for environmental data collected over seven consecutive days, causal convolution extracts temporal features to form a 96-dimensional temporal feature vector.

[0076] The spatial and temporal features are fused through an attention mechanism. The correlation score between the two features is calculated to generate an attention weight matrix, with weight values ​​ranging from 0 to 1. For example, the attention weight for the spatial feature of temperature and the temporal feature of humidity is 0.75, indicating a high correlation. Based on the attention weights, the spatial and temporal feature vectors are weighted and fused to generate a 160-dimensional multimodal nonlinear feature vector.

[0077] During the transfer learning phase, a bidirectional mapping network is constructed between the source domain feature space and the target domain feature space. The source domain can be well-annotated indoor environment data, while the target domain is new environment data to be processed. The bidirectional mapping network consists of two mapping functions: G, which maps from the source domain to the target domain, and F, which maps from the target domain to the source domain. Each mapping function is implemented by a three-layer fully connected network with 128, 64, and 32 hidden layer nodes, respectively.

[0078] A cycle consistency constraint is established, requiring that for source domain feature x, after mapping through G and F, F(G(x))≈x; similarly, for target domain feature y, G(F(y))≈y. The bidirectional mapping network is trained by minimizing the cycle consistency error. For example, the source domain feature [0.5, 0.3, 0.8] is mapped to the target domain to [0.7, 0.2, 0.6], and then mapped back to the source domain to [0.48, 0.32, 0.79], with a cycle consistency error of 0.032.

[0079] To calculate the uncertainty distribution of the mapped features, Monte Carlo sampling was used, randomly sampling each feature point 50 times and calculating the variance as the uncertainty measure. For feature points in high uncertainty regions (variance greater than 0.2), the uncertainty was reduced through neighborhood smoothing. Based on the uncertainty distribution, the mapped features were reorganized, increasing the proportion of low-uncertainty features and decreasing the proportion of high-uncertainty features, ultimately obtaining a 160-dimensional migration feature.

[0080] The migration features are input into the mutual information calculation unit, and the mutual information coefficients between the features are calculated to construct a 160×160-dimensional dynamic correlation matrix. For example, the mutual information coefficient between the temperature and humidity features is 0.72, indicating a high correlation between them. Spectral clustering is used to process the correlation matrix. By calculating the eigenvalue distribution of the Laplacian matrix, the optimal number of factors is determined. In practice, the optimal number of factors is 8.

[0081] Based on the determined number of factors, a variational inference algorithm was used to optimize factor loadings. The factor loading matrix was initialized to a 160×8 random matrix. The variational inference algorithm was then iteratively optimized to maximize the marginal likelihood of the observed data. The iterations were terminated when the likelihood function growth rate fell below 0.001 or when the maximum number of iterations, 100, was reached. The optimized factor loadings explained approximately 85% of the variance in the environmental data.

[0082] To further improve the interpretability of the factor loadings, we used adaptive kernel density estimation to perform nonlinear rotations on the factor loadings. We selected a Gaussian kernel function with a bandwidth parameter adaptively adjusted based on the data distribution, ranging from 0.1 to 0.5. For example, for the temperature-related factor loadings, we used a Gaussian kernel with a bandwidth of 0.3 for density estimation.

[0083] The rotated factor loadings are weighted based on the feature's discriminability, which is calculated by calculating the ratio of between-class variance to within-class variance. For example, the temperature feature has a discriminability of 0.85, while the humidity feature has a discriminability of 0.62, giving the former a higher weight. The weighted factor loadings form the final environmental feature vector, with dimensions of 8 × 1, where each dimension represents a key characteristic of the environment.

[0084] In this embodiment, by preprocessing and standardizing the environmental parameter data, the influence of dimensional differences and outliers is eliminated, the consistency and reliability of the data are ensured, and the stability and prediction accuracy of the model are improved; the pyramid convolution structure and causal convolution structure of the parallel residual network are adopted to effectively capture the spatial and temporal characteristics of the environmental data, and enhance the richness of feature expression and the generalization ability of the model; through bidirectional mapping and cycle consistency constraints, the feature migration between the source domain and the target domain is optimized, the distribution differences between the domains are reduced, and the adaptability of the model in different environments is improved; the spectral clustering method is used to automatically determine the number of factors, and the factor loadings are optimized through variational inference, combined with adaptive kernel density estimation, to achieve nonlinear rotation and weighting of the factor loadings, thereby improving the accuracy and interpretability of environmental feature extraction; through the output of the environmental feature vector, the model's perception of environmental changes is enhanced, providing more reliable and intuitive support for environmental analysis and decision-making.

[0085] like Figure 2 The figure shows a performance comparison of different methods across five evaluation metrics, with the horizontal axis representing the five different evaluation metrics and the vertical axis representing percentages. The figure compares the performance of three different approaches: traditional methods, basic deep learning, and our proposed method. Specifically, in terms of feature extraction accuracy, our proposed method achieved 93.2%, surpassing the 85.6% of basic deep learning and the 78.4% of traditional methods. In terms of temporal association recognition rate, our proposed method achieved 91.5%, also outperforming the 79.3% of basic deep learning and the 65.7% of traditional methods. In terms of inter-domain transfer efficiency, our proposed method achieved 88.7%, significantly exceeding the 73.2% of basic deep learning and the 59.3% of traditional methods. In terms of feature discrimination, our proposed method achieved 90.3%, still outperforming the 80.5% of basic deep learning and the 72.1% of traditional methods. Finally, in terms of computational efficiency, our proposed method achieved 79.6%, slightly lower than the 83.5% of traditional methods but higher than the 71.8% of basic deep learning. Overall, the method proposed in this paper shows obvious advantages in most evaluation indicators, especially in key indicators such as feature extraction, temporal association and inter-domain migration.

[0086] In an optional embodiment, the number of factors is substituted into the variational inference algorithm to optimize the factor loading to obtain the optimal factor loading, and the factor loading is nonlinearly rotated using adaptive kernel density estimation and weighted based on feature discrimination. The output environment feature vector includes:

[0087] The number of factors is input into the variational inference algorithm to construct a Gaussian prior distribution of the factor loads. A variational posterior approximation in the form of a normal distribution is set for the variational parameters. By minimizing the KL divergence criterion, the mean parameter and variance parameter of the variational posterior distribution are iteratively updated to obtain the optimal factor loads.

[0088] An adaptive Gaussian kernel function is constructed for the optimal factor loading. The kernel function bandwidth parameter is dynamically adjusted according to the sample distribution characteristics. The kernel density estimate of the factor loading is calculated. A rotation optimization objective function is constructed based on the kernel density estimate. The rotation matrix is ​​iteratively optimized through gradient ascent. A nonlinear rotation transformation is performed on the optimal factor loading to obtain the rotated factor loading.

[0089] The inter-class scatter matrix and intra-class scatter matrix of the rotated factor loadings are calculated, the feature discrimination is determined according to the trace ratio of the scatter matrix, the weight coefficients are set for the rotated factor loadings based on the feature discrimination, and the weighted factor loadings are input into the nonlinear feature mapping function to generate the environmental feature vector.

[0090] like Figure 3 As shown, the method includes:

[0091] In one specific embodiment, variational inference is performed to optimize factor loadings. In practical applications, assume that there is an environmental data sample matrix X with dimensions n×p, where n is the number of samples and p is the original feature dimension. Based on domain knowledge, an appropriate number of factors k (e.g., k=5) is determined to construct a factor analysis model. During the variational inference optimization process, a zero-mean Gaussian prior distribution is set for the factor loading matrix Λ, with the mean vector being the zero vector and the covariance matrix being the identity matrix multiplied by the prior precision parameter 0.1. To achieve variational approximation, the variational posterior distribution of the factor loadings is set to a multivariate Gaussian distribution, and the initial mean matrix is ​​randomly initialized to a small random number from a standard normal distribution, e.g., with a mean in the range [-0.01, 0.01], and the initial variance parameter is set to 0.5.

[0092] During the variational inference iterations, stochastic gradient ascent is used to update the variational parameters. Each iteration uses a mini-batch of 32 samples, with an initial learning rate of 0.01. The Adam optimizer is used for parameter updates. The optimization objective is the evidence lower bound (ELBO), which is achieved by calculating the KL divergence between the variational distribution and the true posterior distribution. The number of iterations is set to 1000, and the iteration is terminated early when the ELBO value changes by less than a threshold of 1e-6 for 10 consecutive iterations. After the iterations are completed, the mean of the variational posterior distribution is taken as the optimal factor loading matrix Λ*, with dimensions p × k.

[0093] After obtaining the optimal factor loading, adaptive kernel density estimation nonlinear rotation is performed. The standard deviation of the sample set in each dimension is calculated, and the Scott rule is selected to determine the initial kernel function bandwidth parameter h_0 = 1.06 × σ × n -1 / 5 , where σ is the standard deviation and n is the number of samples. An adaptive bandwidth adjustment mechanism is designed to address the data distribution characteristics of different feature dimensions: for dimensions with skewness greater than 2 or kurtosis greater than 4, the bandwidth parameter is multiplied by 0.8; for data-dense areas (where the spacing between data points is less than 0.1 standard deviations), the bandwidth parameter is multiplied by 0.9; and for data-sparse areas (where the spacing between data points is greater than 2 standard deviations), the bandwidth parameter is multiplied by 1.2.

[0094] A Gaussian kernel function was constructed based on the adjusted bandwidth parameter, and kernel density estimates were calculated for each element of the factor loading matrix. For example, if the first element of the optimal factor loading was 0.732, its density estimate calculated using the adaptive kernel function was 0.253. After calculating density estimates for all elements, a rotation optimization objective function was constructed, which consisted of a weighted simple structure criterion and a density distribution consistency criterion with a weight ratio of 7:3. The rotation matrix R was optimized using the gradient ascent method with an iteration step size of 0.05, a maximum number of iterations of 200, and a convergence threshold of 1e-4. Iterations were terminated when the objective function changed less than the threshold for five consecutive times. The resulting rotation matrix R was used to transform the optimal factor loadings, resulting in the rotated factor loading matrix Λ_r = Λ* × R.

[0095] The rotated factor loadings are weighted based on feature discrimination. Assuming that the environmental samples come from different categories (e.g., indoor and outdoor environments), the inter-class scatter matrix S_b and the intra-class scatter matrix S_w corresponding to the rotated factor loadings are calculated. The inter-class scatter is calculated as the sum of the squares of the differences between the mean of each category and the overall mean, while the intra-class scatter is calculated as the sum of the squares of the differences between each sample and the mean of its category. Feature discrimination is calculated using the Fisher criterion: d_i = tr(S_b_i) / tr(S_w_i), where tr represents the matrix trace, S_b_i and S_w_i are the inter-class and intra-class scatters of the i-th feature, respectively. The calculated discrimination is normalized to obtain the normalized discrimination d'_i.

[0096] Based on the normalized discrimination, we design weight coefficients w_i = exp(α × d'_i), where α is a tuning parameter and is set to 2.0 by default. We apply the weight coefficients to the rotated factor loadings to obtain weighted factor loadings Λ_w, where Λ_w[i, j] = Λ_r[i, j] × w_i.

[0097] The weighted factor loadings are passed through a nonlinear feature mapping function to generate an environmental feature vector. The ReLU activation function is selected as the nonlinear component of the mapping function, and a residual connection is added to preserve the original information. The specific mapping formula is: y = ReLU(Λ_w × z) + 0.2 × z, where z is the input factor and y is the output environmental feature vector. To prevent eigenvalue explosion, the final feature vector is L2-normalized.

[0098] In the existing technology, landscape lighting control mostly adopts fixed schedules or simple light sensor triggering mechanisms, which lack the ability to comprehensively analyze complex environmental factors. Traditional feature extraction methods such as principal component analysis (PCA) or linear discriminant analysis (LDA) often use linear transformations, which make it difficult to capture nonlinear relationships in the data and are sensitive to noise. The method of this embodiment replaces the traditional maximum likelihood estimation with a variational inference algorithm, thereby improving the stability and generalization ability of factor analysis; introduces an adaptive Gaussian kernel function and nonlinear rotation transformation to enhance the ability to extract nonlinear environmental features; and dynamically adjusts feature weights through inter-class divergence and intra-class divergence analysis to improve the identification accuracy of key environmental factors.

[0099] In an optional embodiment, a hierarchical adaptive neural network is constructed based on the environmental feature vector, and scene generation is performed according to the regional illumination distribution law and the luminous flux superposition relationship between lighting units. Visual comfort is calculated by the matching degree between the luminous flux distribution and the human visual response characteristics. The dynamic programming algorithm is used to iteratively optimize and generate lighting scene parameters, including:

[0100] Constructing a hierarchical adaptive neural network based on the environmental feature vector, dynamically adjusting the weight coefficients of the nodes in each layer of the neural network through the feature correlation matrix, and mapping and transforming the environmental feature vector to obtain a scene feature representation;

[0101] A luminous flux spatial distribution function is constructed using the scene feature representation. An initial luminous flux distribution of each lighting unit in the target area is calculated based on the luminous flux spatial distribution function. A luminous flux superposition coefficient between the lighting units is calculated based on the spatial three-dimensional coordinates, luminous direction, and shading coefficient of the lighting units. The initial luminous flux distribution and the luminous flux superposition coefficient are combined to generate a luminous flux distribution map of the target area.

[0102] Extracting spectral energy distribution characteristics from the luminous flux distribution diagram of the target area, calculating dark adaptation thresholds and light adaptation thresholds according to a predetermined spectral response curve of human cone cells, establishing a visual comfort evaluation function, and performing matching calculation on the spectral energy distribution characteristics to obtain visual comfort parameters;

[0103] The visual comfort parameters are mapped into state transition probabilities, a state transition matrix of lighting scene parameters is constructed, and dynamic programming iterative calculations are performed on the lighting scene parameters using the state transition matrix to generate optimal lighting scene parameters.

[0104] In one specific embodiment, the lighting scene generation system based on environmental feature vectors begins by collecting environmental feature data. The environmental feature vectors contain key parameters such as indoor temperature, humidity, light intensity, occupancy density, and time of day. These parameters are collected in real time via a sensor network, pre-processed to eliminate outliers, and standardized to form a feature vector with a dimension of 15.

[0105] A four-layer adaptive neural network was constructed, consisting of an input layer, two hidden layers, and an output layer. The input layer had 15 nodes, matching the dimensions of the environmental feature vector. The first hidden layer had 32 nodes, the second hidden layer had 16 nodes, and the output layer had 8 nodes corresponding to the scene feature representation. The network weights were dynamically adjusted by calculating the feature correlation matrix. When the correlation coefficient between temperature and light intensity reached 0.85, the connection weight between these two nodes was automatically increased by a factor of 1.2. For feature pairs with a correlation below 0.2, such as humidity and occupancy density, the connection weight was reduced to 0.8, thereby optimizing the network structure. Initial weights were randomly generated in the interval [-0.5, 0.5]. ReLU was used as the activation function, and batch normalization techniques were applied to improve training stability.

[0106] After completing the feature mapping, the obtained 8-dimensional scene feature representation is used to construct the luminous flux spatial distribution function. This function uses an improved Gaussian distribution model to describe the scattering law of light in space. For an office area of ​​50 square meters, the system divides it into 10×10 grid cells and calculates the initial luminous flux value of each grid cell. For the lighting unit located at coordinates (3, 5), the initial luminous flux is set to 850 lumens; the lighting unit located at (7, 2) is set to 750 lumens. The spatial position information of the lighting unit is further considered, for example, the height of the ceiling-mounted lighting equipment is 2.8 meters, and the height of the wall-mounted lighting equipment is 1.5 meters, and its light direction angle is recorded.

[0107] A luminous flux superposition model was established to calculate the impact between lighting units. When an obstruction exists between two lighting units, the obstruction coefficient is set to 0.6; when there is no obstruction, the coefficient is set to 1.0. Through iterative calculations, the system generates a luminous flux distribution map for the target area, showing a luminous flux of 950 lumens in the center and 650 lumens at the edges, creating a smooth transition lighting effect.

[0108] After the luminous flux distribution diagram is generated, the system extracts the spectral energy distribution characteristics, including the energy distribution curve within the visible spectrum. The system has built-in spectral response curve data for three types of cone cells in the human body (S type, M type, and L type). Under dark conditions, the response threshold of the S type cell is set to 10cd / m 2 The thresholds of M-type and L-type cells are 15 cd / m 2 and 12cd / m 2 In bright environments, the adaptation thresholds of these three cells are increased to 100 cd / m 2 , 120cd / m 2 and 110cd / m 2 .

[0109] The established visual comfort evaluation function comprehensively considers three key indicators: the degree of match between spectral energy distribution and human visual response, illumination uniformity, and glare index. For a typical office environment, the weighting for spectral matching is set to 0.5, illumination uniformity to 0.3, and glare index to 0.2. The calculated visual comfort parameter for the current luminous flux distribution is 0.78, which is close to but not yet at the ideal value of 0.85.

[0110] To optimize lighting parameters, visual comfort parameters are mapped to state transition probabilities. The constructed state transition matrix includes three dimensions: brightness level, color temperature, and illumination angle. Brightness levels are divided into 10 levels, and the color temperature range is from 2700K to 6500K, divided into 8 levels. The illumination angle ranges from 0° to 180° and is divided into 6 intervals. The system uses a value iteration method for dynamic programming calculations, setting a discount factor of 0.95 and a maximum number of iterations of 100.

[0111] After 57 iterations, the optimal lighting scenario parameters were generated: the central lighting unit was adjusted to 920 lumens, the color temperature was set to 4200K, and the illumination angle was 75°; the edge lighting units were adjusted to 680 lumens, the color temperature was set to 3800K, and the illumination angle was 60°. Under these parameters, the calculated visual comfort index increased to 0.87, exceeding the target value of 0.85, while also reducing energy consumption by 12%.

[0112] In this embodiment, the weights of nodes in each layer are dynamically adjusted through a hierarchical adaptive neural network and a feature correlation matrix, making the scene features mapped by the environmental feature vector more discriminative and robust, thereby improving the basic quality of subsequent luminous flux calculations. A luminous flux spatial distribution function based on scene characteristics is constructed, and the superposition coefficient is calculated by combining three-dimensional coordinates, luminous direction, and occlusion coefficient, thereby achieving high-precision simulation of the interaction between multiple light sources in the target area. Spectral energy characteristics are extracted from the luminous flux distribution diagram, and the dark / light adaptation threshold is calculated with reference to the spectral response curve of human cone cells. Visual comfort is quantified through a matching function, making the evaluation more closely aligned with human perception. Visual comfort parameters are mapped to state transition probabilities, and a state transition matrix of lighting scene parameters is constructed. Dynamic programming iterative search is used to automatically obtain the global optimal lighting solution for a given scene. A closed-loop adaptive process is formed from environmental characteristics to scene characteristics, comfort assessment, and parameter optimization. Each step has clearly defined physical or physiological indicators, which improves the interpretability of the method and the deployment trust.

[0113] In an optional embodiment, spectral energy distribution characteristics are extracted from the luminous flux distribution diagram of the target area, dark adaptation thresholds and light adaptation thresholds are calculated based on a predetermined spectral response curve of human cone cells, a visual comfort evaluation function is established, and a matching degree calculation is performed on the spectral energy distribution characteristics to obtain visual comfort parameters, including:

[0114] A two-dimensional discrete wavelet transform is performed on the luminous flux distribution map of the target area to obtain a multi-scale spectral coefficient matrix. A Gaussian mixture distribution is fitted to the multi-scale spectral coefficient matrix, and the spectral energy probability density function is calculated and generated through expectation maximization iteration. The spectral energy distribution characteristics are determined based on the spectral energy probability density function.

[0115] The spectral response curve of human cone cells is subjected to a two-way competitive operation with multiple sets of historical visual response data. A dynamic correction coefficient is obtained through the mutual game between the response parameter and the historical parameter. A convolution operation based on the dynamic correction coefficient is performed on the spectral response curve to generate a correction response curve. The dark adaptation threshold and the light adaptation threshold are calculated based on the correction response curve.

[0116] Multi-layer feature extraction is performed on the spectral energy distribution characteristics, dark adaptation threshold and light adaptation threshold to obtain the visual energy adaptation coefficient, visual brightness response coefficient and visual time adjustment coefficient, and the visual comfort evaluation function is constructed.

[0117] The spectral energy distribution characteristics and visual comfort evaluation function are dynamically weighted and iteratively optimized to calculate the visual comfort parameters.

[0118] In a specific embodiment, the luminous flux distribution map of the target area is processed, a two-dimensional discrete wavelet transform operation is performed, and the luminous flux distribution map is decomposed into multiple layers using the Haar wavelet basis function to obtain detail coefficients in the horizontal, vertical and diagonal directions respectively. In actual operation, a three-layer wavelet decomposition is selected, and the high-frequency component matrix obtained after each layer of decomposition can be expressed as three directions: H (horizontal), V (vertical) and D (diagonal), and the low-frequency component is represented by A. After three layers of decomposition, ten sub-bands are finally obtained: A3, H3, V3, D3, H2, V2, D2, H1, V1, D1. These sub-bands together constitute a multi-scale spectral coefficient matrix, which represents the energy characteristics of the luminous flux distribution at different scales and directions. For example, for a luminous flux distribution map of 1024×1024 pixels, the dimension of the third-layer low-frequency coefficient matrix A3 is 128×128.

[0119] The obtained multi-scale spectral coefficient matrix was fitted with a Gaussian mixture distribution using three sets of Gaussian distribution models, each with a different mean and covariance parameter. Iterative optimization was performed using the expectation-maximization algorithm, with an iteration threshold of 0.001 and a maximum number of 100 iterations. In a typical case, after 72 iterations, convergence was achieved, resulting in weights of 0.45, 0.35, and 0.2 for the three Gaussian components, respectively. The mean vectors were [0.78, 0.62], [0.41, 0.58], and [0.25, 0.31]. The covariance matrices represent the distribution characteristics of each component. Based on these parameters, a spectral energy probability density function was constructed. The probability density value of each point was calculated at 100 equally spaced sampling points within the sample spectral energy space. After normalization, a spectral energy distribution feature vector with a dimension of 100 was obtained, which characterizes the statistical distribution characteristics of the spectral energy in the target area.

[0120] In terms of processing the spectral response curves of human cone cells, the standard response curve data of the three known types of cone cells, L-type, M-type, and S-type, for the visible spectrum (380nm-780nm) are used, with a sampling interval of 5nm, to form three 81-dimensional vectors. Combined with multiple sets of collected historical visual response data (including 50 sets of human eye adaptation parameters under different lighting environments), a two-way competitive operation is performed. Specifically, a similarity score is calculated for each set of historical data and the current response curve, and the top 10 historical data with the highest similarity are selected for weighted fusion, with the weight coefficients allocated proportionally to the similarity. Through the mutual adjustment between the response parameters and the historical parameters, a dynamic correction coefficient is generated. This coefficient shows a nonlinear distribution in the range of 380-780nm. The correction coefficient is larger in the middle band (500-600nm) (average value 1.15) and smaller in the two end bands (average value 0.92). A convolution operation is performed on the original spectral response curve and the dynamic correction coefficient to generate a corrected response curve. In actual cases, the response peak of the corrected M-type cone cells at 545nm is increased from the original 0.95 to 1.08, which is more in line with the actual visual response characteristics under the target light environment.

[0121] According to the correction response curve, the dark adaptation threshold and light adaptation threshold are calculated. The dark adaptation threshold is calculated by the incremental brightness detection method, starting from the brightness value close to zero (0.01cd / m 2 ) and press 0.01cd / m 2 The step size is increased, and the activation degree of the three types of cone cells is calculated for each brightness value. When the activation degree exceeds the set threshold (0.15) for the first time, the corresponding brightness value is the dark adaptation threshold. The dark adaptation threshold obtained in this case is 0.13cd / m 2 The light adaptation threshold is similar to the method, but it is calculated from the high brightness (10000cd / m 2 ) is calculated in descending order. When the supersaturation response of the three types of cone cells is lower than the set threshold (0.85), the corresponding brightness value is the light adaptation threshold. The light adaptation threshold obtained in this case is 2450 cd / m 2 .

[0122] The construction of the visual comfort evaluation function involves multiple layers of feature extraction. First, the visual energy adaptation coefficient is calculated based on the spectral energy distribution characteristics and the adaptation threshold. The specific method is to perform a dot product operation on the spectral energy distribution value at each wavelength and the response curve value of the three types of cone cells at that wavelength. The total stimulus is then accumulated and compared with the difference between the dark adaptation threshold and the light adaptation threshold to generate a normalized visual energy adaptation coefficient. The value range is [0, 1]. The adaptation coefficient calculated in this case is 0.78. The visual brightness response coefficient is calculated by calculating the ratio of the weighted mean of the spectral energy to the light adaptation threshold. In this case, the coefficient is 0.62. The visual temporal accommodation coefficient takes into account the temporal characteristics of the light environment. It is obtained by calculating the standard deviation of the spectral data sampled for 10 seconds and then performing nonlinear mapping. In this case, the coefficient is 0.85.

[0123] When constructing the visual comfort evaluation function, the three coefficients mentioned above are combined using a weighted combination, with weights of 0.4, 0.35, and 0.25, respectively, to form a complete evaluation function. This function exhibits nonlinear characteristics and exhibits varying sensitivities to changes in the three parameters. Dynamic weight iterative optimization of the spectral energy distribution characteristics and the visual comfort evaluation function is performed using a gradient descent method with a learning rate of 0.05 and 50 iterations. The weights are adjusted with each iteration to ensure that the evaluation results are more consistent with the actual visual comfort experience. In this case, the final visual comfort parameter obtained is 0.726, indicating good visual comfort in the current light environment.

[0124] Existing visual comfort evaluation technologies primarily employ simple illuminance threshold methods or brightness uniformity evaluation methods, which consider only the physical quantities of the lighting system and ignore the physiological characteristics of the human visual system and its dynamic adaptation process. For example, traditional methods typically use 300-500 lux as a fixed standard for indoor lighting comfort evaluation, or use the ratio of maximum to minimum brightness to assess uniformity. These methods lack consideration of spectral energy distribution and an understanding of the complexity of the human visual system. Existing spectral analysis techniques often employ Fourier transforms, making it difficult to capture local features and multi-scale information in the light environment. Visual response models often employ the fixed-parameter CIE standard observer function, which is unable to adapt to differences in visual response across individuals and environmental conditions. The method of this embodiment replaces traditional Fourier analysis with a two-dimensional discrete wavelet transform to more accurately capture the multi-scale and directional characteristics of the luminous flux distribution; introduces a Gaussian mixture distribution model to probabilistically model the spectral coefficients to more comprehensively characterize the statistical characteristics of the spectral energy; innovatively proposes a bidirectional competitive operation mechanism to achieve dynamic correction of the cone cell response curve, so that the visual evaluation model can adapt to different individuals and environmental conditions; constructs a multi-level visual comfort evaluation function that comprehensively considers factors such as energy adaptation, brightness response, and time adjustment, which is more in line with the physiological mechanism of the human eye's visual comfort.

[0125] like Figure 4 As shown in the figure, the calculation response time performance of four different algorithms is shown when the ambient light intensity changes from 0lx to 5000lx. Different symbols are used in the figure to represent different algorithm implementations: the circular markers represent the original HMC Monte Carlo algorithm, which is the most basic probabilistic sampling method and has the longest computational response time, which gradually decreases from 67.8ms at 0lx to 52.4ms at 5000lx. The triangular markers represent the Markov chain sampling algorithm, which improves computational efficiency by optimizing the sampling strategy, and the response time is reduced from 53.4ms at 0lx to 35.2ms at 5000lx. The diamond markers represent the optimization algorithm based on Bayesian inference. This method introduces prior knowledge to guide the sampling process, which improves the computational response time from 42.7ms at 0lx to 23.9ms at 5000lx. The star markers represent the adaptive Gaussian kernel density estimation algorithm proposed in this paper. This method achieves optimal computational efficiency by dynamically adjusting the kernel function parameters and sampling weights, and the response time is significantly reduced from 28.5ms at 0lx to 12.8ms at 5000lx. From the overall trend, the response time of all algorithms decreases with the increase of ambient light intensity, which shows that the improvement of lighting conditions helps to improve computational efficiency, and the adaptive algorithm proposed in this paper maintains the best performance under various lighting conditions.

[0126] In an optional embodiment, lighting scene parameters are input into a group collaborative learning network, control groups are divided according to the spatial distribution relationship of lighting units, and an optimization mechanism of local competition within groups and global collaboration between groups is adopted. The optimal control strategy is calculated in combination with a dynamic energy consumption prediction model, including:

[0127] Obtain the position coordinates, luminous flux parameters, and luminous angle parameters of each lighting unit in the lighting scene parameters, calculate and determine the spatial distance matrix and the lighting influence matrix between the lighting units, perform an exponential decay operation based on the spatial distance matrix and the lighting influence matrix to obtain the correlation strength between the lighting units, and divide the lighting units with correlation strength greater than a preset correlation threshold into the same control group;

[0128] The energy consumption parameters, lighting effect parameters, and group interference parameters of each lighting unit in the control group are combined to obtain the group optimization target value. A two-way competition mechanism is used to adjust the control parameters of the lighting units against each other and the parameter values ​​are corrected through the group energy balance constraint to generate the initial control parameters of the lighting units in the group.

[0129] The intra-group optimization target value of each control group and the cross-interference degree between the control groups are combined to calculate the inter-group collaborative target value, and the inter-group collaborative target value is optimized and solved by the distributed alternating direction multiplier method to output the inter-group collaborative control parameters;

[0130] The predicted energy consumption data is obtained by extracting the time series features of historical scene parameters, load characteristic parameters, and operating status parameters. The predicted energy consumption data is dynamically weighted iteratively calculated with the inter-group collaborative control parameters to generate the optimal control strategy for the lighting unit.

[0131] In a specific embodiment, the collection of lighting scene parameters includes information such as the position coordinates, luminous flux parameters, and luminous angle parameters of each lighting unit. The system obtains the real-time position data of the lighting unit through a sensor array and records it as three-dimensional coordinates (x, y, z); the luminous flux value of each lighting unit is measured by a photometer, with a typical value range of 300-1500 lumens; the luminous angle parameter is obtained by measuring the main optical axis direction and divergence angle of the lighting unit, and the divergence angle is generally between 30°-120°. Taking office area A as an example, 20 LED lighting units are deployed in the area, numbered L1-L20, of which L1 is located at coordinates (2.5, 3.0, 2.8), with a luminous flux of 800 lumens and a main luminous angle of 45°.

[0132] The spatial distance matrix calculation uses the Euclidean distance formula to calculate the physical distance between any two lighting units. For lighting units L1 (2.5, 3.0, 2.8) and L2 (5.0, 3.5, 2.8) in this example, the calculated spatial distance is 2.55 meters. This calculation is performed for all 20 lighting units, forming a 20×20 spatial distance matrix D.

[0133] The lighting impact matrix calculation takes into account the lighting unit's luminous flux, beam angle, and occlusion factors. Based on the inverse square attenuation principle of light, combined with a cosine correction for the beam angle, the system calculates the lighting impact intensity between any two lighting units. For example, considering the impact of L1 on L2, with L1's luminous flux of 800 lumens, a distance of 2.55 meters, and a beam angle of 15°, the calculated lighting impact intensity is 75.2 lumens per square meter. Similar calculations are performed for all lighting units, forming a 20×20 lighting impact matrix I.

[0134] The correlation strength calculation uses an exponential decay function. Each element of the distance matrix D and the illumination influence matrix I is exponentially transformed using decay coefficients α = 0.3 and β = 0.5, respectively. The two transformed matrix elements are then weighted and summed, with a weight ratio of 0.4 for distance and 0.6 for illumination influence, to obtain the correlation strength matrix R between lighting units. Correlation strength values ​​range from 0 to 1, with larger values ​​indicating stronger correlations.

[0135] The control group division set the correlation threshold to 0.65, and the lighting units with correlation strength greater than this threshold were classified into the same control group. The correlation strength matrix R was analyzed by clustering algorithm, and the 20 lighting units were finally divided into four control groups: G1{L1, L3, L5, L7}, G2{L2, L4, L6, L8, L10}, G3{L9, L11, L13, L15, L17}, and G4{L12, L14, L16, L18, L19, L20}.

[0136] The calculation of the intra-group optimization target value comprehensively considers energy consumption parameters, lighting effect parameters, and intra-group interference parameters. Energy consumption parameters include the product of each lighting unit's rated power and its current brightness level; lighting effect parameters include the target illumination achievement rate and illumination uniformity; and the intra-group interference parameter characterizes the degree of mutual influence between lighting units within the group. Taking control group G1 as an example, the energy consumption parameters of the four lighting units are 10W, 8W, 9W, and 7W, respectively; the lighting effect parameters are 0.95, 0.92, 0.90, and 0.93; and the intra-group interference parameter is 0.15. Through combined calculations, the energy consumption parameter is assigned a weight of 0.3, the lighting effect parameter a weight of 0.5, and the intra-group interference parameter a weight of 0.2, resulting in a intra-group optimization target value of 0.82 for G1.

[0137] The two-way competition mechanism is a process in which the control parameters of lighting units within a group are adjusted against each other. The system employs an iterative game algorithm, allowing each lighting unit to adjust its brightness level and angle parameters based on its own status and the influence of other units. In each iteration, lighting units compete with each other based on the needs of the target area and energy consumption, dynamically adjusting parameter values. For example, after 15 iterations, the brightness level of L1 was adjusted from the initial 85% to 78%, and the beam angle was fine-tuned from 45° to 42°. The brightness level of L3 was adjusted from 80% to 76%, while the angle remained unchanged. Similar adjustments were made to other units, ultimately achieving energy balance within the group.

[0138] The inter-group synergy objective is calculated by combining the optimization objective within each control group with the cross-interference between the control groups. The cross-interference is calculated by analyzing the combined illumination effects between the lighting units of different control groups. For the four control groups in this example, the inter-group cross-interference matrix is ​​a 4×4 matrix with element values ​​ranging from 0 to 0.5. The cross-interference between G1 and G2 is 0.25, indicating a certain degree of illumination influence between the two groups. A weighted combination of the optimization objective and the cross-interference for each group is used to derive the inter-group synergy objective function.

[0139] The distributed alternating direction multiplier method is used to solve the inter-group collaborative optimization problem. This method decomposes the global optimization problem into multiple local subproblems. Each control group solves the subproblems locally and exchanges information. Global convergence is achieved by introducing Lagrange multipliers and augmentation terms. During the iterative process, each control group updates its control strategy based on its current parameters and information from other groups, and the process continues until convergence. After 30 iterations, the system obtained the globally approximately optimal inter-group collaborative control parameters.

[0140] The energy consumption prediction model is built based on historical scenario parameters, load characteristics, and operating status parameters. The system collects nearly 90 days of lighting system operating data, including lighting unit on / off status, brightness levels, and energy consumption data. Time series analysis extracts seasonal patterns, cyclical changes, and trend characteristics to construct a prediction model. Using an example scenario, the model predicts energy consumption for the next hour. The total energy consumption is 320W, with the control groups' energy consumption being 74W for G1, 92W for G2, 85W for G3, and 69W for G4, respectively.

[0141] The optimal control strategy is generated by combining predicted energy consumption data with inter-group collaborative control parameters, and the final control strategy is determined through dynamic weighted iterative calculations. Based on the current scenario requirements and predicted energy consumption, the system dynamically adjusts the weighting between energy efficiency and lighting performance. When energy is scarce, the system increases the weighting of energy efficiency; when lighting quality requirements are high, the weighting of lighting performance is increased. Through five rounds of dynamic weighted iterative calculations, the final lighting unit control strategy is generated, including the brightness level, angle parameters, and switching timing of each lighting unit, achieving an optimal balance between lighting performance and energy consumption.

[0142] In this embodiment, control groups are divided based on the correlation intensity of spatial distance and illumination influence, so that lighting units with large mutual interference are grouped together to facilitate group management and collaborative optimization; within each control group, two-way competition and energy balance constraints are carried out in combination with energy consumption, lighting effect and interference degree to achieve the optimal initial setting of control parameters within the group; the optimization objectives of each group are integrated with the cross-interference degree and solved using the distributed alternating direction multiplier method to ensure the coordination and consistency of control parameters between groups and reduce the negative impact between adjacent groups; the time series characteristics of historical scenarios, loads and operating states are used to predict future energy consumption, and the dynamic weighted iteration with the collaborative control parameters between groups is used to adjust the control strategy in real time to improve energy saving and lighting quality; from unit correlation analysis, intra-group optimization, inter-group collaboration to dynamic prediction, a closed-loop optimization process is constructed to ensure that the final control strategy is globally optimal in terms of overall energy consumption, lighting effect and system stability.

[0143] In an optional embodiment, the intra-group optimization target value of each control group and the cross-interference degree between the control groups are combined and calculated to obtain the inter-group coordination target value. The inter-group coordination target value is optimized and solved by the distributed alternating direction multiplier method. The output inter-group coordination control parameters include:

[0144] The luminous flux, angle, and distance parameters of the lighting units between different control groups are calculated to obtain the inter-group cross-interference matrix. The dynamic weight matrix is ​​generated through exponential normalization operation to calculate the inter-group coupling coefficient of each control group.

[0145] The intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient are compensated and calculated to obtain the inter-group synergy target value;

[0146] The time series features of ambient temperature parameters, crowd density parameters, and meteorological change parameters are extracted to obtain a time series sensitivity matrix. Local variables, global variables, and Lagrange multipliers are introduced into the inter-group collaborative target value to construct the first augmented Lagrangian function. The time series sensitivity matrix and the first augmented Lagrangian function are compensated for the response to obtain the second augmented Lagrangian function.

[0147] For the second augmented Lagrangian function, the global variables and Lagrange multipliers are fixed, and each control group independently performs gradient descent to obtain local variables; the local variables and Lagrange multipliers are fixed, and gradient descent is performed in combination with the global constraints to obtain global variables; a bidirectional mapping operation is performed to generate a variable mapping matrix and update the Lagrange multipliers;

[0148] Perform difference calculation on local variables and global variables to obtain the original residual, calculate the difference of global variables in adjacent iterations to obtain the dual residual, adjust the penalty parameter according to the ratio of the original residual to the dual residual, and input the second augmented Lagrangian function for iterative calculation;

[0149] When the original residual is less than a first preset threshold and the dual residual is less than a second preset threshold, the iteration is terminated, and the inter-group collaborative control parameters are calculated based on the local variables and the global variables.

[0150] In one specific embodiment, the cross-interference matrix of lighting units between different control groups is calculated. For N control groups, each control group contains several lighting units, and the luminous flux parameters, installation angle parameters, and position distance parameters of each lighting unit are extracted separately. For example, for lighting unit a of control group A and lighting unit b of control group B, assuming that the luminous flux of a is 1000 lumens and the luminous flux of b is 800 lumens, the angle between the two is 45 degrees, and the distance is 3 meters, the cross-interference degree can be calculated as 176.8 based on the inverse law of light intensity and the cosine attenuation formula. The interference degree of the lighting units between all control groups is calculated pairwise to form an N×N dimensional cross-interference degree matrix D.

[0151] The dynamic weight matrix W is generated by exponential normalization operation. The specific operation is to perform exponential transformation and normalization on each element in the cross interference matrix D. Assume that the interference degree of control group A to control group B is 176.8. After exponential transformation (e -176.8 / 1000) and normalized to obtain a weight value of 0.83. The inter-group coupling coefficient of each control group is calculated based on the weight matrix W. For example, the coupling coefficient of control group A is the weighted average of its weights with all other groups, which is assumed to be 0.75.

[0152] Compensate the intra-group optimization target value for each control group with the corresponding inter-group coupling coefficient to obtain the inter-group synergy target value. For control group A, for example, if its intra-group optimization target value is 500 lux and its inter-group coupling coefficient is 0.75, its inter-group synergy target value can be calculated as 500 × (1 + 0.75 × 0.2) = 575 lux, where 0.2 is the compensation coefficient. Similar calculations are performed for all N control groups to form a synergy target value vector.

[0153] Time series feature extraction is performed on ambient temperature parameters, crowd density parameters, and meteorological change parameters. For example, for an office area, 24 hours of environmental data is collected, one set per hour. This includes temperature (22-28°C), crowd density (0-100 people / m2), and light intensity (0-1000 lux). The sliding window method is used to extract the time series variation features of these parameters and generate a time series sensitivity matrix S.

[0154] Based on the distributed alternating direction multiplier method, the augmented Lagrangian function is constructed to solve the inter-group cooperative control problem. For N control groups, the local variables x={x1, x2, ..., x n} represents the control parameters of each control group, the global variable z represents the control parameters that satisfy the global constraint, and the Lagrangian multiplier y. A first augmented Lagrangian function is constructed, and the timing sensitivity matrix S is subjected to response compensation with this function to obtain the second augmented Lagrangian function L.

[0155] During the iterative solution process, the global variable z and Lagrange multiplier y are fixed, and each control group independently performs gradient descent to update its local variable x. Taking control group A as an example, assume that the initial control parameter x1 is [450 lux, 3000K] and is updated to [480 lux, 3200K] after gradient descent iterations. With all local variables x and Lagrange multipliers y fixed, gradient descent is performed to update the global variable z in accordance with global constraints (e.g., total power does not exceed 10,000W). A bidirectional mapping operation is performed to generate the variable mapping matrix M, which is used to update the Lagrange multiplier y.

[0156] Calculate the original residual r=|xz| and the dual residual s=|z (k+1) -z k For example, in a certain iteration, the local variable x = [480 lux, 3200K] and the global variable z = [470 lux, 3150K], then the original residual r = |(480, 3200)-(470, 3150)| = 56.8. If the global variables of two adjacent iterations are zk =[470lux, 3150K] and z (k+1) =[472lux, 3160K], then the dual residual s=|(472, 3160)-(470, 3150)|=14.1.

[0157] Adjust the penalty parameter ρ based on the ratio of the primal residual to the dual residual. If r / s > 10, increase ρ by a factor of 2; if s / r > 10, decrease ρ by a factor of 2; otherwise, keep ρ unchanged. Feed the updated penalty parameter ρ into the second augmented Lagrangian function L and continue the iteration.

[0158] The first preset threshold is set to 5, and the second preset threshold is set to 10. The iteration terminates when the original residual r is less than 5 and the dual residual s is less than 10. Finally, the inter-group collaborative control parameters are calculated based on the converged local variable x and global variable z. Taking control group A as an example, the final collaborative control parameters are illuminance 475 lux, color temperature 3175K, and power adjusted to 92% of the original. These parameters are sent to the corresponding control group actuators, achieving inter-group collaborative control of the lighting system and effectively balancing the lighting effect and energy consumption of each control group.

[0159] Existing lighting control systems often use independent control of individual zones or simple centralized control, lacking comprehensive consideration of the interactions between different control groups. Traditional lighting systems typically employ schedule-based preset control strategies or simple feedback control based on light sensors. For example, when the illumination in a certain area is insufficient, the lighting intensity in that area is increased, while ignoring the light spillover effect in adjacent areas. Existing technologies also employ centralized optimization methods using a central controller, but the computational complexity increases exponentially with the number of control groups, making it difficult to achieve real-time control of large-scale lighting systems. In addition, traditional lighting control systems often respond instantaneously to environmental parameters and lack the ability to extract and predict temporal variation characteristics, resulting in control lag or oscillation when the environment changes rapidly. The method of this embodiment accurately quantifies the mutual influence between different lighting control groups by constructing an inter-group cross-interference matrix; introduces a distributed optimization framework to decompose the global optimization problem into multiple local sub-problems that can be solved in parallel, significantly improving computational efficiency; innovatively introduces a temporal sensitivity matrix into the augmented Lagrangian function, enabling the control system to make predictive responses to the temporal characteristics of environmental changes; and designs a dynamic penalty parameter adjustment mechanism to improve the convergence speed and stability of the algorithm.

[0160] In multi-control group scenarios, compared with traditional independent control methods, it saves energy consumption; shortens response time, and significantly improves the speed of adapting to environmental changes; and reduces the algorithm computational complexity from O(n³) of traditional centralized control to O(n).

[0161] The smart green landscape lighting adaptive control system based on deep learning in an embodiment of the present invention includes:

[0162] The first unit is used to collect light intensity, crowd density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data;

[0163] The second unit is used to perform standardized preprocessing on environmental parameter data, fuse spatial and temporal features using a parallel residual network to obtain multimodal features, perform transfer learning on multimodal features using bidirectional mapping, determine the number of factors based on the dynamic correlation matrix, and generate environmental feature vectors through variational inference and kernel density estimation;

[0164] The third unit is used to build a hierarchical adaptive neural network based on the environmental feature vector, generate scenes according to the regional illumination distribution law and the superposition relationship between the luminous flux of lighting units, calculate visual comfort by matching the luminous flux distribution with the human visual response characteristics, and use a dynamic programming algorithm to iteratively optimize and generate lighting scene parameters;

[0165] The fourth unit is used to input lighting scene parameters into the group collaborative learning network, divide the control groups according to the spatial distribution relationship of the lighting units, adopt the optimization mechanism of local competition within the group and global collaboration between groups, and combine it with the dynamic energy consumption prediction model to calculate the optimal control strategy;

[0166] The fifth unit is used to generate an adjustment instruction for the lighting device based on the optimal control strategy and perform lighting adjustment.

[0167] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0168] processor;

[0169] a memory for storing processor-executable instructions;

[0170] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0171] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0172] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart green landscape lighting adaptive control method based on deep learning, characterized by: include: Collect light intensity, crowd density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data; The environmental parameter data is preprocessed by standardization, and the spatial and temporal features are fused using a parallel residual network to obtain multimodal features, including: Preprocess and standardize environmental parameter data to obtain standardized environmental data; The standardized environmental data is input into the spatial feature extraction branch of the parallel residual network, and the spatial features are obtained through the pyramid convolution structure. At the same time, the standardized environmental data is input into the temporal feature extraction branch of the parallel residual network, and the temporal features are obtained through the causal convolution structure. The spatial features and temporal features are fused according to the attention weight to generate multimodal nonlinear features. Bidirectional mapping is used to transfer multimodal features, the number of factors is determined based on the dynamic correlation matrix, and the environment feature vector is generated through variational inference and kernel density estimation; A hierarchical adaptive neural network is constructed based on the environmental feature vector. Scenes are generated based on the regional illumination distribution law and the luminous flux superposition relationship between lighting units. Visual comfort is calculated by the matching degree between the luminous flux distribution and the human visual response characteristics. A dynamic programming algorithm is used to iteratively optimize and generate lighting scene parameters, including: Constructing a hierarchical adaptive neural network based on the environmental feature vector, dynamically adjusting the weight coefficients of the nodes in each layer of the neural network through the feature correlation matrix, and mapping and transforming the environmental feature vector to obtain a scene feature representation; A luminous flux spatial distribution function is constructed using the scene feature representation. An initial luminous flux distribution of each lighting unit in the target area is calculated based on the luminous flux spatial distribution function. A luminous flux superposition coefficient between the lighting units is calculated based on the spatial three-dimensional coordinates, luminous direction, and shading coefficient of the lighting units. The initial luminous flux distribution and the luminous flux superposition coefficient are combined to generate a luminous flux distribution map of the target area. Extracting spectral energy distribution characteristics from the luminous flux distribution diagram of the target area, calculating dark adaptation thresholds and light adaptation thresholds according to a predetermined spectral response curve of human cone cells, establishing a visual comfort evaluation function, and performing matching calculation on the spectral energy distribution characteristics to obtain visual comfort parameters; Mapping the visual comfort parameters into state transition probabilities, constructing a state transition matrix of lighting scene parameters, and performing dynamic programming iterative calculations on the lighting scene parameters using the state transition matrix to generate optimal lighting scene parameters; The lighting scene parameters are input into the group collaborative learning network, and the control groups are divided according to the spatial distribution relationship of the lighting units. The optimization mechanism of local competition within the group and global collaboration between groups is adopted, and the optimal control strategy is calculated in combination with the dynamic energy consumption prediction model. Generate lighting equipment adjustment instructions based on the optimal control strategy and execute lighting adjustment.

2. The method according to claim 1, characterized in that The environmental parameter data is preprocessed by standardization, and the spatial and temporal features are fused using a parallel residual network to obtain multimodal features. The multimodal features are transferred using bidirectional mapping. The number of factors is determined based on the dynamic correlation matrix, and the environmental feature vector is generated through variational inference and kernel density estimation. The multimodal nonlinear features are input into the source domain feature space and the target domain feature space for bidirectional mapping, cycle consistency constraints are established, the uncertainty distribution of the mapped features is calculated, and the mapped features are reorganized according to the uncertainty distribution to obtain the migration features. The migration features are input into the mutual information calculation unit to construct a dynamic correlation matrix, and the number of factors is determined from the dynamic correlation matrix using the spectral clustering method; The number of factors is substituted into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings. The adaptive kernel density estimation is used to perform nonlinear rotation on the factor loadings and weight them based on the feature discrimination to output the environmental feature vector.

3. The method according to claim 2, characterized in that Substitute the number of factors into the variational inference algorithm to optimize the factor loadings and obtain the optimal factor loadings. Adaptive kernel density estimation is used to perform nonlinear rotation on the factor loadings and weight them based on feature discrimination. The output environment feature vector includes: The number of factors is input into the variational inference algorithm to construct a Gaussian prior distribution of the factor loads. A variational posterior approximation in the form of a normal distribution is set for the variational parameters. By minimizing the KL divergence criterion, the mean parameter and variance parameter of the variational posterior distribution are iteratively updated to obtain the optimal factor loads. An adaptive Gaussian kernel function is constructed for the optimal factor loading. The kernel function bandwidth parameter is dynamically adjusted according to the sample distribution characteristics. The kernel density estimate of the factor loading is calculated. A rotation optimization objective function is constructed based on the kernel density estimate. The rotation matrix is ​​iteratively optimized through gradient ascent. A nonlinear rotation transformation is performed on the optimal factor loading to obtain the rotated factor loading. The inter-class scatter matrix and intra-class scatter matrix of the rotated factor loadings are calculated, the feature discrimination is determined according to the trace ratio of the scatter matrix, the weight coefficients are set for the rotated factor loadings based on the feature discrimination, and the weighted factor loadings are input into the nonlinear feature mapping function to generate the environmental feature vector.

4. The method according to claim 1, wherein Extracting spectral energy distribution characteristics from the target area luminous flux distribution diagram, calculating dark adaptation thresholds and light adaptation thresholds based on a predetermined spectral response curve of human cone cells, establishing a visual comfort evaluation function, and performing matching calculation on the spectral energy distribution characteristics to obtain visual comfort parameters include: A two-dimensional discrete wavelet transform is performed on the luminous flux distribution map of the target area to obtain a multi-scale spectral coefficient matrix. A Gaussian mixture distribution is fitted to the multi-scale spectral coefficient matrix, and the spectral energy probability density function is calculated and generated through expectation maximization iteration. The spectral energy distribution characteristics are determined based on the spectral energy probability density function. The spectral response curve of human cone cells is subjected to a two-way competitive operation with multiple sets of historical visual response data. A dynamic correction coefficient is obtained through the mutual game between the response parameter and the historical parameter. A convolution operation based on the dynamic correction coefficient is performed on the spectral response curve to generate a correction response curve. The dark adaptation threshold and the light adaptation threshold are calculated based on the correction response curve. Multi-layer feature extraction is performed on the spectral energy distribution characteristics, dark adaptation threshold and light adaptation threshold to obtain the visual energy adaptation coefficient, visual brightness response coefficient and visual time adjustment coefficient, and the visual comfort evaluation function is constructed. The spectral energy distribution characteristics and visual comfort evaluation function are dynamically weighted and iteratively optimized to calculate the visual comfort parameters.

5. The method according to claim 1, wherein The lighting scene parameters are input into the group collaborative learning network. The control groups are divided according to the spatial distribution relationship of the lighting units. The optimization mechanism of local competition within the group and global collaboration between groups is adopted. Combined with the dynamic energy consumption prediction model, the optimal control strategy is calculated, including: Obtain the position coordinates, luminous flux parameters, and luminous angle parameters of each lighting unit in the lighting scene parameters, calculate and determine the spatial distance matrix and the lighting influence matrix between the lighting units, perform an exponential decay operation based on the spatial distance matrix and the lighting influence matrix to obtain the correlation strength between the lighting units, and divide the lighting units with correlation strength greater than a preset correlation threshold into the same control group; The energy consumption parameters, lighting effect parameters, and group interference parameters of each lighting unit in the control group are combined to obtain the group optimization target value. A two-way competition mechanism is used to adjust the control parameters of the lighting units against each other and the parameter values ​​are corrected through the group energy balance constraint to generate the initial control parameters of the lighting units in the group. The intra-group optimization target value of each control group and the cross-interference degree between the control groups are combined to calculate the inter-group collaborative target value, and the inter-group collaborative target value is optimized and solved by the distributed alternating direction multiplier method to output the inter-group collaborative control parameters; The predicted energy consumption data is obtained by extracting the time series features of historical scene parameters, load characteristic parameters, and operating status parameters. The predicted energy consumption data is dynamically weighted iteratively calculated with the inter-group collaborative control parameters to generate the optimal control strategy for the lighting unit.

6. The method according to claim 5, characterized in that The inter-group coordination target value is obtained by combining the intra-group optimization target value of each control group and the cross-interference degree between the control groups. The inter-group coordination target value is optimized and solved by the distributed alternating direction multiplier method. The output inter-group coordination control parameters include: The luminous flux, angle, and distance parameters of the lighting units between different control groups are calculated to obtain the inter-group cross-interference matrix. The dynamic weight matrix is ​​generated through exponential normalization operation to calculate the inter-group coupling coefficient of each control group. The intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient are compensated and calculated to obtain the inter-group synergy target value; The time series features of ambient temperature parameters, crowd density parameters, and meteorological change parameters are extracted to obtain a time series sensitivity matrix. Local variables, global variables, and Lagrange multipliers are introduced into the inter-group collaborative target value to construct the first augmented Lagrangian function. The time series sensitivity matrix and the first augmented Lagrangian function are compensated for the response to obtain the second augmented Lagrangian function. For the second augmented Lagrangian function, the global variables and Lagrange multipliers are fixed, and each control group independently performs gradient descent to obtain local variables; the local variables and Lagrange multipliers are fixed, and gradient descent is performed in combination with the global constraints to obtain global variables; a bidirectional mapping operation is performed to generate a variable mapping matrix and update the Lagrange multipliers; Perform difference calculation on local variables and global variables to obtain the original residual, calculate the difference of global variables in adjacent iterations to obtain the dual residual, adjust the penalty parameter according to the ratio of the original residual to the dual residual, and input the second augmented Lagrangian function for iterative calculation; When the original residual is less than a first preset threshold and the dual residual is less than a second preset threshold, the iteration is terminated, and the inter-group collaborative control parameters are calculated based on the local variables and the global variables.

7. A smart green landscape lighting adaptive control system based on deep learning, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to collect light intensity, crowd density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data; The second unit is used to perform standardized preprocessing on environmental parameter data, fuse spatial and temporal features using a parallel residual network to obtain multimodal features, perform transfer learning on multimodal features using bidirectional mapping, determine the number of factors based on the dynamic correlation matrix, and generate environmental feature vectors through variational inference and kernel density estimation; The third unit is used to build a hierarchical adaptive neural network based on the environmental feature vector, generate scenes according to the regional illumination distribution law and the superposition relationship between the luminous flux of lighting units, calculate visual comfort by matching the luminous flux distribution with the human visual response characteristics, and use a dynamic programming algorithm to iteratively optimize and generate lighting scene parameters; The fourth unit is used to input lighting scene parameters into the group collaborative learning network, divide the control groups according to the spatial distribution relationship of the lighting units, adopt the optimization mechanism of local competition within the group and global collaboration between groups, and combine it with the dynamic energy consumption prediction model to calculate the optimal control strategy; The fifth unit is used to generate an adjustment instruction for the lighting device based on the optimal control strategy and perform lighting adjustment.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images