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

Abstract

The invention provides a smart green landscape illumination adaptive control method and system based on deep learning, and relates to the technical field of illumination control, and the method comprises the steps: collecting parameters through an environment sensor, fusing features through a parallel residual network, constructing a layered adaptive neural network, and generating an illumination scene; and calculating an optimal control strategy by using a packet collaborative learning network to realize illumination adjustment. According to the invention, the lighting parameters can be automatically adjusted according to the environmental change, the visual comfort is improved, the energy consumption is reduced, and the intellectualization and greenization of landscape lighting are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of lighting control, and particularly to an intelligent green landscape lighting adaptive control method and system based on deep learning. Background Art

[0002] With the acceleration of the urbanization process and the improvement of the awareness of ecological environmental protection, intelligent green landscape lighting, as an important part of urban construction, has become a key factor in enhancing the urban image and improving the quality of life of residents. Traditional landscape lighting systems usually adopt fixed patterns or simple time control strategies, and cannot adjust intelligently according to environmental conditions and changes in the flow of people, resulting in energy waste and affecting the visual experience. In recent years, with the development of artificial intelligence and deep learning technologies, intelligent lighting control systems have begun to be applied to the field of urban landscape lighting, achieving a balance between energy conservation and emission reduction and visual comfort by perceiving environmental changes in real time and adaptively adjusting lighting parameters.

[0003] There are some problems in the actual application of existing landscape lighting systems. Most systems lack the ability to effectively fuse multi-source environmental data, and cannot comprehensively consider the complex correlations of multi-dimensional environmental factors such as light intensity, population density, and meteorological conditions, resulting in single and insufficiently adaptable lighting control decisions; traditional lighting control algorithms are difficult to accurately evaluate the visual comfort in different lighting scenarios, and fail to scientifically match the human visual perception characteristics with the light flux distribution, and there are over-bright or over-dark phenomena in some areas, affecting the user experience; most existing lighting systems adopt an independent control mode, lacking a cooperative control mechanism between lighting units, and failing to fully consider the spatial distribution relationship and dynamic changes in energy consumption, and cannot achieve optimal energy utilization while ensuring the visual effect. Summary of the Invention

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

[0005] In the first aspect of the embodiments of the present invention, an intelligent green landscape lighting adaptive control method based on deep learning is provided, including: Collecting light intensity, population density, and meteorological data of a target area through environmental sensors to obtain environmental parameter data; Performing standardized preprocessing on the environmental parameter data, fusing spatial-temporal features by using a parallel residual network to obtain multi-modal features, performing transfer learning on the multi-modal features by using a bidirectional mapping, determining the number of factors based on a dynamic correlation matrix, and generating an environmental feature vector through variational inference and kernel density estimation; Construct a hierarchical adaptive neural network based on environmental feature vectors, generate scenes according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculate visual comfort through the matching degree between the light flux distribution and the human visual response characteristics, and iteratively optimize and generate lighting scene parameters using a dynamic programming algorithm; Input the lighting scene parameters into a grouped collaborative learning network, divide control groups according to the spatial distribution relationship of lighting units, adopt an optimization mechanism of intra-group local competition and inter-group global cooperation, and calculate the optimal control strategy in combination with an energy consumption dynamic prediction model; Generate adjustment instructions for lighting equipment based on the optimal control strategy and execute lighting adjustment.

[0006] In an alternative embodiment, perform standardized preprocessing on environmental parameter data, use a parallel residual network to fuse spatio-temporal features to obtain multi-modal features, perform transfer learning on the multi-modal features using bidirectional mapping, determine the number of factors based on a dynamic correlation matrix, and generate environmental feature vectors through variational inference and kernel density estimation, including: Preprocess and standardize the environmental parameter data to obtain standardized environmental data; Input the standardized environmental data into the spatial feature extraction branch of the parallel residual network, obtain spatial features through a pyramid-shaped convolutional structure, and at the same time input the standardized environmental data into the temporal feature extraction branch of the parallel residual network, obtain temporal features through a causal convolutional structure, and fuse the spatial features and temporal features according to attention weights to generate multi-modal non-linear features; Input the multi-modal non-linear features into the source domain feature space and the target domain feature space for bidirectional mapping, establish a cyclic consistency constraint, calculate the uncertainty distribution of the mapped features, reorganize the mapped features according to the uncertainty distribution, and obtain transfer features; Input the transfer features into a mutual information calculation unit to construct a dynamic correlation matrix, and use a spectral clustering method to determine the number of factors from the dynamic correlation matrix; Substitute the number of factors into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings, perform non-linear rotation on the factor loadings using adaptive kernel density estimation and weight them based on feature discrimination, and output environmental feature vectors.

[0007] In an alternative embodiment, substitute the number of factors into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings, perform non-linear rotation on the factor loadings using adaptive kernel density estimation and weight them based on feature discrimination, and output environmental feature vectors, including: Input the number of factors into the variational inference algorithm, construct a Gaussian prior distribution of the factor loadings, set a variational posterior approximation in the form of a normal distribution for the variational parameters, and iteratively update the mean parameter and variance parameter of the variational posterior distribution by minimizing the KL divergence criterion to obtain the optimal factor loadings; Construct an adaptive Gaussian kernel function for the optimal factor loadings, dynamically adjust the kernel function bandwidth parameter according to the sample distribution characteristics, calculate the kernel density estimate of the factor loadings, construct a rotation optimization objective function based on the kernel density estimate, iteratively optimize the rotation matrix through gradient ascent, perform a non-linear rotation transformation on the optimal factor loadings to obtain the rotated factor loadings; Calculate the between-class scatter matrix and the within-class scatter matrix of the rotated factor loadings, determine the feature discrimination degree according to the trace ratio of the scatter matrices, set the weight coefficients for the rotated factor loadings based on the feature discrimination degree, and input the weighted factor loadings into a non-linear feature mapping function to generate the environmental feature vector.

[0008] In an alternative embodiment, construct a hierarchical adaptive neural network based on the environmental feature vector, generate a scene according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculate the visual comfort degree through the matching degree between the light flux distribution and the human visual response characteristics, and iteratively optimize the generated lighting scene parameters using a dynamic programming algorithm, including: Construct a hierarchical adaptive neural network based on the environmental feature vector, dynamically adjust the weight coefficients of the nodes in each layer of the neural network through the feature correlation matrix, and map and transform the environmental feature vector to obtain the scene feature representation; Use the scene feature representation to construct a light flux spatial distribution function, calculate the initial light flux distribution of each lighting unit in the target area according to the light flux spatial distribution function, and calculate the light flux superposition coefficient between lighting units by combining the three-dimensional spatial coordinates, light emission direction, and occlusion coefficient of the lighting units, and perform a combined operation on the initial light flux distribution and the light flux superposition coefficient to generate a light flux distribution map of the target area; Extract the spectral energy distribution characteristics from the light flux distribution map of the target area, calculate the dark adaptation threshold and the light adaptation threshold according to the pre-determined spectral response curve of the human cone cells, establish a visual comfort evaluation function, and calculate the matching degree of the spectral energy distribution characteristics to obtain the visual comfort parameters; Map the visual comfort parameters to state transition probabilities, construct a state transition matrix of the lighting scene parameters, and perform dynamic programming iterative calculations on the lighting scene parameters through the state transition matrix to generate the optimal lighting scene parameters.

[0009] In an alternative embodiment, extract the spectral energy distribution characteristics from the light flux distribution map of the target area, calculate the dark adaptation threshold and the light adaptation threshold according to the pre-determined spectral response curve of the human cone cells, establish a visual comfort evaluation function, and calculate the matching degree of the spectral energy distribution characteristics to obtain the visual comfort parameters, including: Perform a two-dimensional discrete wavelet transform on the light flux distribution map of the target area to obtain a multi-scale spectral coefficient matrix. Fit the multi-scale spectral coefficient matrix with a Gaussian mixture distribution, and through the expectation-maximization iteration, calculate and generate a spectral energy probability density function. Based on the spectral energy probability density function, calculate and determine the spectral energy distribution characteristics; Perform a two-way competition operation on the spectral response curve of the human cone cells and multiple groups of historical visual response data. Obtain a dynamic correction coefficient through the mutual game between the response parameters and the historical parameters. Perform a convolution operation on the spectral response curve based on the dynamic correction coefficient to generate a corrected response curve, and calculate the dark adaptation threshold and the light adaptation threshold according to the corrected response curve; Perform multi-layer feature extraction calculations on the spectral energy distribution characteristics, the dark adaptation threshold, and the light adaptation threshold to obtain a visual energy adaptation coefficient, a visual brightness response coefficient, and a visual time adjustment coefficient, and construct a visual comfort evaluation function; Perform dynamic weight iteration optimization on the spectral energy distribution characteristics and the visual comfort evaluation function, and calculate the visual comfort parameter.

[0010] In an optional embodiment, input the lighting scene parameters into a grouped collaborative learning network. Divide the control groups according to the spatial distribution relationship of the lighting units. Adopt an optimization mechanism of local competition within the group and global cooperation between the groups, and combine an energy consumption dynamic prediction model to calculate the optimal control strategy, including: Obtain the position coordinates, light flux parameters, and light emission angle parameters of each lighting unit in the lighting scene parameters. Calculate and determine the spatial distance matrix and the light influence matrix between the lighting units. Perform an exponential decay operation according to the spatial distance matrix and the light influence matrix to obtain the association strength between the lighting units. Divide the lighting units with an association strength greater than the preset association degree threshold into the same control group; Perform a combined operation on the energy consumption parameters, lighting effect parameters, and intra-group interference degree parameters of the lighting units within each control group to obtain an intra-group optimization target value. Adopt a two-way competition mechanism to make the control parameters of the lighting units confront and adjust each other, and correct the parameter values through the intra-group energy balance constraint to generate the initial control parameters of the lighting units within the group; Perform a combined calculation on the intra-group optimization target values of each control group and the cross-interference degree between the control groups to obtain an inter-group collaborative target value. Optimize and solve the inter-group collaborative target value by using the distributed alternating direction multiplier method, and output the inter-group collaborative control parameters; Perform time-series feature extraction on the historical scene parameters, load characteristic parameters, and operating state parameters to obtain predicted energy consumption data. Perform dynamic weight iteration calculation on the predicted energy consumption data and the inter-group collaborative control parameters to generate the optimal control strategy for the lighting units.

[0011] In an alternative embodiment, the intra-group optimization target values of each control group and the cross-interference degrees between the control groups are combined and calculated to obtain the inter-group cooperation target value. The distributed alternating direction multiplier method is used to optimize and solve the inter-group cooperation target value, and the output inter-group cooperation control parameters include: Calculate the luminous flux, included angle, and distance parameters of the lighting units between different control groups to obtain an inter-group cross-interference degree matrix, generate a dynamic weight matrix through exponential normalization operation, and calculate the inter-group coupling coefficient of each control group; Perform compensation calculation on the intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient to obtain the inter-group cooperation target value; Extract the time-series characteristics of the environmental temperature parameter, population density parameter, and meteorological change parameter to obtain a time-series sensitive matrix. Introduce local variables, global variables, and Lagrange multipliers into the inter-group cooperation target value to construct a first augmented Lagrangian function, and perform response compensation on the time-series sensitive matrix and the first augmented Lagrangian function to obtain a second augmented Lagrangian function; For the second augmented Lagrangian function, fix the global variables and Lagrange multipliers, and each control group independently performs gradient descent to obtain local variables; fix the local variables and Lagrange multipliers, and perform gradient descent in combination with global constraints to obtain global variables; perform bidirectional mapping operation to generate a variable mapping matrix, and update the Lagrange multipliers; Perform difference operation on the local variables and global variables to obtain the original residual, calculate the difference of adjacent iteration global variables to obtain the dual residual, adjust the penalty parameter according to the ratio of the original residual to the dual residual, and input it into the second augmented Lagrangian function for iterative operation; When the original residual is less than the first preset threshold and the dual residual is less than the second preset threshold, terminate the iteration, and calculate the inter-group cooperation control parameters according to the local variables and global variables.

[0012] In the second aspect of the embodiments of the present invention, a wisdom green landscape lighting adaptive control system based on deep learning is provided, including: The first unit is used to collect the light intensity, population 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 the environmental parameter data, fuse spatial and temporal characteristics by using a parallel residual network to obtain multi-modal characteristics, perform transfer learning on the multi-modal characteristics by using bidirectional mapping, determine the number of factors based on a dynamic association matrix, and generate an environmental feature vector through variational inference and kernel density estimation; The third unit is used to construct a hierarchical adaptive neural network based on the environmental feature vector, generate a scene according to the regional illuminance distribution law and the luminous flux superposition relationship between lighting units, calculate the visual comfort according to the matching degree between the luminous flux distribution and the human visual response characteristics, and iteratively optimize and generate lighting scene parameters by using a dynamic programming algorithm; The fourth unit is used to input lighting scene parameters into a grouped collaborative learning network, divide control groups according to the spatial distribution relationship of lighting units, adopt an optimization mechanism of intra-group local competition and inter-group global cooperation, and calculate the optimal control strategy in combination with an energy consumption dynamic prediction model; The fifth unit is used to generate adjustment instructions for lighting devices based on the optimal control strategy and execute lighting adjustment.

[0013] In a third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0014] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0015] In the embodiments of the present invention, through the environmental sensor data acquisition and parallel residual network fusion technology, the deep fusion of multi-source data and accurate feature extraction are realized. The 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 the perception ability and environmental adaptability; The hierarchical adaptive neural network structure and dynamic programming algorithm are adopted to realize the refined control of the lighting scene according to the human visual characteristics and the light flux distribution relationship. Through the visual comfort evaluation mechanism, the humanized design of the light environment is realized, so that the lighting effect better meets the human visual needs and improves the user experience and environmental comfort; The grouped collaborative learning network and the energy consumption dynamic prediction model are introduced to minimize the energy consumption while ensuring the lighting effect. The control strategy is optimized through the intra-group local competition and inter-group global cooperation mechanism, and the parameters of the lighting device are intelligently adjusted, reducing the energy consumption compared with the traditional control method, and realizing the balance between energy conservation and environmental protection and visual effect of landscape lighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flow chart of the intelligent green landscape lighting adaptive control method based on deep learning in the embodiments of the present invention; Figure 2 It is a bar chart comparing the performance of the environmental feature extraction method based on multi-modal fusion and transfer learning in the embodiments of the present invention; Figure 3 It is the process of generating an environmental feature vector based on factor load rotation optimization in the embodiments of the present invention; Figure 4Schematic diagram for comparing the response time of visual comfort parameters in the embodiments of the present invention. Detailed implementation manners

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] The technical solutions of the present invention will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0019] Figure 1 Schematic flowchart of the intelligent green landscape lighting adaptive control method based on deep learning in the embodiments of the present invention, as Figure 1 shown, the method includes: Collecting the light intensity, pedestrian flow density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data; Performing standardized preprocessing on the environmental parameter data, using a parallel residual network to fuse spatial-temporal features to obtain multi-modal features, performing transfer learning on the multi-modal features using bidirectional mapping, determining the number of factors based on a dynamic correlation matrix, and generating an environmental feature vector through variational inference and kernel density estimation; Constructing a hierarchical adaptive neural network based on the environmental feature vector, generating a scene according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculating the visual comfort through the matching degree between the light flux distribution and the human visual response characteristics, and iteratively optimizing to generate lighting scene parameters using a dynamic programming algorithm; Inputting the lighting scene parameters into a grouped collaborative learning network, dividing control groups according to the spatial distribution relationship of lighting units, adopting an optimization mechanism of intra-group local competition and inter-group global cooperation, and calculating an optimal control strategy in combination with an energy consumption dynamic prediction model; Generating an adjustment instruction for lighting equipment based on the optimal control strategy and performing lighting adjustment.

[0020] In an alternative embodiment, performing standardized preprocessing on the environmental parameter data, using a parallel residual network to fuse spatial-temporal features to obtain multi-modal features, performing transfer learning on the multi-modal features using bidirectional mapping, determining the number of factors based on a dynamic correlation matrix, and generating an environmental feature vector through variational inference and kernel density estimation includes: Performing preprocessing and standardization on the environmental parameter data to obtain standardized environmental data; Input the standardized environmental data into the spatial feature extraction branch of the parallel residual network to obtain spatial features through a pyramidal convolution structure. At the same time, input the standardized environmental data into the temporal feature extraction branch of the parallel residual network to obtain temporal features through a causal convolution structure. Then, fuse the spatial features and temporal features according to the attention weights to generate multimodal non-linear features; Input the multimodal non-linear features into the source domain feature space and the target domain feature space respectively for bidirectional mapping, establish a cycle consistency constraint, calculate the uncertainty distribution of the mapped features, and reorganize the mapped features according to the uncertainty distribution to obtain transfer features; Input the transfer features into the mutual information calculation unit to construct a dynamic correlation matrix, and use the spectral clustering method to determine the number of factors from the dynamic correlation matrix; Substitute the number of factors into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings, perform non-linear rotation on the factor loadings using adaptive kernel density estimation and weight them based on feature discrimination, and output the environmental feature vector.

[0021] In a specific embodiment, preprocess the collected environmental parameter data by standardization. Specifically, collect multi-dimensional environmental parameter data including temperature, humidity, light, air pressure, noise, etc., fill in the missing values of these data, use the moving average method to process the outliers, and then use the Z-score standardization method to standardize the data of 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], after standardization, it becomes [0.21, 1.37, -0.95, -0.63].

[0022] The standardized environmental data is input into the parallel residual network structure, which includes a spatial feature extraction branch and a temporal feature extraction branch. In the spatial feature extraction branch, a pyramidal convolution structure is adopted, specifically including 3 convolutional layers. The first layer uses 32 3×3 convolutional kernels, the second layer uses 64 3×3 convolutional kernels, and the third layer uses 128 3×3 convolutional kernels. After each layer, batch normalization and ReLU activation functions are connected, and the original information is maintained through residual connections. For a five-dimensional data matrix such as temperature and humidity, a 128-dimensional spatial feature vector is obtained after spatial feature extraction.

[0023] In the temporal feature extraction branch, a causal convolution structure is adopted to capture the temporal dependence relationship, including 3 causal convolutional layers, and the convolutional kernels with dilation rates of 1, 2, and 4 are used in each layer to ensure that the current output only depends on the historical input. For example, for the environmental data collected continuously for 7 days, after extracting the temporal features through causal convolution, a 96-dimensional temporal feature vector is formed.

[0024] Fuse spatial features and temporal features through an attention mechanism. Calculate the correlation scores of the two types of features to generate an attention weight matrix, where the weight values range from 0 to 1. For example, the attention weight of the spatial feature of temperature and the temporal feature of humidity is 0.75, indicating a high correlation between them. According to the attention weights, the spatial feature vector and the temporal feature vector are weighted and fused to generate a 160-dimensional multimodal non-linear feature vector.

[0025] In the transfer learning stage, construct a bidirectional mapping network between the source domain feature space and the target domain feature space. The source domain can be indoor environment data with sufficient annotations, and the target domain is the new environment data to be processed. The bidirectional mapping network contains two mapping functions: the mapping G from the source domain to the target domain and the mapping F from the target domain to the source domain. Each mapping function is implemented by a three-layer fully connected network, and the number of hidden layer nodes is 128, 64, and 32 respectively.

[0026] Establish a cycle consistency constraint, requiring that for the source domain feature x, after mapping through G and F, it should satisfy F(G(x))≈x; similarly, for the target domain feature y, it should satisfy G(F(y))≈y. Train the bidirectional mapping network 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 get [0.7, 0.2, 0.6], and then mapped back to the source domain to get [0.48, 0.32, 0.79], and the cycle consistency error is 0.032.

[0027] Calculate the uncertainty distribution of the mapped features. Using the Monte Carlo sampling method, perform 50 random samplings for each feature point, and calculate the variance as the uncertainty measure. For the feature points in the high-uncertainty region (variance greater than 0.2), reduce the uncertainty through neighborhood smoothing. Reorganize the mapped features according to the uncertainty distribution, increasing the proportion of low-uncertainty features and decreasing the proportion of high-uncertainty features, and finally obtain 160-dimensional transfer features.

[0028] Input the transfer features into the mutual information calculation unit, calculate the mutual information coefficient between the features, and 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. Use the spectral clustering method to process the correlation matrix, and determine the optimal number of factors by calculating the eigenvalue distribution of the Laplacian matrix. In practical applications, the obtained optimal number of factors is 8.

[0029] Based on the determined number of factors, the variational inference algorithm is used to optimize the factor loadings. The factor loading matrix is initialized as a random matrix of 160×8, and is iteratively optimized through variational inference to maximize the marginal likelihood of the observed data. During the iteration process, it stops when the growth rate of the likelihood function is less than 0.001 or reaches the maximum number of iterations 100. The optimized factor loadings can explain approximately 85% of the variance of the environmental data.

[0030] To further improve the interpretability of the factor loadings, adaptive kernel density estimation is used to perform nonlinear rotation on the factor loadings. The Gaussian kernel function is selected, and the bandwidth parameter is adaptively adjusted according to the data distribution, with the bandwidth range between 0.1 and 0.5. For example, for the factor loadings related to temperature, a Gaussian kernel with a bandwidth of 0.3 is used for density estimation.

[0031] The rotated factor loadings are weighted based on the feature discriminability, and the discriminability is obtained by calculating the ratio of the between-class variance to the within-class variance. For example, the discriminability of the temperature feature is 0.85, and the discriminability of the humidity feature is 0.62, with the former having a higher weight. The weighted factor loadings form the final environmental feature vector, with a dimension of 8×1, and each dimension represents a key feature of the environment.

[0032] In this embodiment, by preprocessing and standardizing the environmental parameter data, the influence of dimensional differences and outliers is eliminated, ensuring the consistency and reliability of the data, and improving the stability and prediction accuracy of the model; the pyramid convolution structure and causal convolution structure of the parallel residual network are adopted to effectively capture the spatial and temporal features in the environmental data, enhancing 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, reducing the distribution differences between domains, and improving the adaptability of the model in different environments; the spectral clustering method is used to automatically determine the number of factors, and the factor loadings are optimized through variational inference. Combining with adaptive kernel density estimation, the nonlinear rotation and weighting of the factor loadings are realized, improving the accuracy and interpretability of environmental feature extraction. Through the output of the environmental feature vector, the model's perception ability of environmental changes is enhanced, providing more reliable and intuitive support for environmental analysis and decision-making.

[0033] As Figure 2As shown in the figure, the performance comparison of different methods on five evaluation metrics is presented. The horizontal axis represents the five different evaluation metrics, and the vertical axis represents the percentage value. The figure compares the performance of three different schemes: the traditional method, the basic deep learning method, and the method proposed in this paper. Specifically, in terms of the feature extraction accuracy, the method proposed in this paper reaches 93.2%, which is higher than 85.6% of the basic deep learning method and 78.4% of the traditional method; in terms of the time-series correlation recognition rate, the method proposed in this paper reaches 91.5%, which is also better than 79.3% of the basic deep learning method and 65.7% of the traditional method; in terms of the inter-domain transfer efficiency, the method proposed in this paper achieves 88.7%, significantly exceeding 73.2% of the basic deep learning method and 59.3% of the traditional method; in terms of the feature discrimination index, the method proposed in this paper reaches 90.3%, still better than 80.5% of the basic deep learning method and 72.1% of the traditional method; finally, in terms of the computational efficiency, the method proposed in this paper is 79.6%, slightly lower than 83.5% of the traditional method but higher than 71.8% of the basic deep learning method. Generally speaking, the method proposed in this paper shows obvious advantages in most evaluation metrics, especially in the key metrics such as feature extraction, time-series correlation, and inter-domain transfer.

[0034] In an optional implementation, substitute the number of factors into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings, perform non-linear rotation on the factor loadings using adaptive kernel density estimation and weight them based on the feature discrimination, and the output environmental feature vector includes: Input the number of factors into the variational inference algorithm, construct a Gaussian prior distribution of the factor loadings, set a variational posterior approximation in the form of a normal distribution for the variational parameters, and iteratively update the mean parameter and variance parameter of the variational posterior distribution by minimizing the KL divergence criterion to obtain the optimal factor loadings; Construct an adaptive Gaussian kernel function for the optimal factor loadings, dynamically adjust the kernel function bandwidth parameter according to the sample distribution characteristics, calculate the kernel density estimation value of the factor loadings, construct a rotation optimization objective function based on the kernel density estimation value, iteratively optimize the rotation matrix through gradient ascent, and perform non-linear rotation transformation on the optimal factor loadings to obtain the rotated factor loadings; Calculate the between-class scatter matrix and within-class scatter matrix of the rotated factor loadings, determine the feature discrimination based on the trace ratio of the scatter matrices, set a weight coefficient for the rotated factor loadings based on the feature discrimination, and input the weighted factor loadings into a non-linear feature mapping function to generate the environmental feature vector.

[0035] As Figure 3 shown, the method includes: In a specific embodiment, variational inference is performed to optimize the factor loadings. In practical applications, assume 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. Determine an appropriate number of factors k (e.g., k = 5) based on domain knowledge and construct a factor analysis model. During the variational inference optimization process, set a zero-mean Gaussian prior distribution 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, set the variational posterior distribution of the factor loadings to a multivariate Gaussian distribution. The initial mean matrix is randomly initialized with small random numbers from a standard normal distribution, e.g., the mean is in the range [-0.01, 0.01], and the initial variance parameter is set to 0.5.

[0036] During the variational inference iteration process, the stochastic gradient ascent method is used to update the variational parameters. Each iteration uses a mini-batch of 32 samples. The learning rate is initially set to 0.01, and the Adam optimizer is used for parameter update. 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. Set the number of iterations to 1000. When the change in the ELBO value is less than the threshold 1e-6 for 10 consecutive iterations, terminate the iteration early. After the iteration is completed, take the mean of the variational posterior distribution as the optimal factor loading matrix Λ*, with dimensions p×k.

[0037] After obtaining the optimal factor loadings, perform adaptive kernel density estimation for non-linear rotation. Calculate the standard deviation of the sample set in each dimension, and select the Scott rule 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. Design an adaptive bandwidth adjustment mechanism for the data distribution characteristics of different feature dimensions: for dimensions with skewness greater than 2 or kurtosis greater than 4, multiply the bandwidth parameter by 0.8; for data-dense regions (regions where the data point spacing is less than 0.1 times the standard deviation), multiply the bandwidth parameter by 0.9; for data-sparse regions (regions where the data point spacing is greater than 2 times the standard deviation), multiply the bandwidth parameter by 1.2.

[0038] A Gaussian kernel function was constructed based on the adjusted bandwidth parameters to calculate the kernel density estimates of each element of the factor loading matrix. For example, if the first element of the optimal factor loading is 0.732, its density estimate calculated by the adaptive kernel function is 0.253. After calculating the density estimates for all elements, a rotation optimization objective function was constructed, which consists of a weighted combination of a 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 of 0.05, a maximum number of iterations of 200, and a convergence threshold of 1e-4. The iteration was stopped when the objective function changed less than the threshold for five consecutive times. The obtained rotation matrix R was used to transform the optimal factor loading, and the rotated factor loading matrix Λ_r = Λ* × R was obtained.

[0039] The rotated factor loadings are weighted based on the feature discrimination. Assuming that the environmental samples come from different categories (such as indoor and outdoor environments), the inter-class dispersion matrix S_b and intra-class dispersion matrix S_w corresponding to the rotated factor loadings are calculated. The inter-class dispersion calculates the sum of the squares of the differences between the mean of each category and the overall mean, and the intra-class dispersion calculates the sum of the squares of the differences between each sample and the mean of the category to which it belongs. The feature discrimination is calculated using the Fisher criterion: d_i = tr(S_b_i) / tr(S_w_i), where tr represents the trace of the matrix, S_b_i and S_w_i are the inter-class dispersion and intra-class dispersion of the i-th feature, respectively. The calculated discrimination is normalized to obtain the normalized discrimination d'_i.

[0040] Based on the normalized discrimination, the weight coefficient w_i = exp(α × d'_i) is designed, where α is the adjustment parameter and is set to 2.0 by default. The weight coefficient is applied to the rotated factor loading to obtain the weighted factor loading Λ_w, where Λ_w[i, j] = Λ_r[i, j] × w_i.

[0041] The weighted factor loads are passed through a nonlinear feature mapping function to generate an environmental feature vector. The ReLU activation function is selected as the nonlinear part of the mapping function, and a residual connection is added to retain 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.

[0042] In the prior art, landscape lighting control mostly adopts a fixed schedule or a simple light sensor triggering mechanism, lacking 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 transformation, which is difficult to capture the non-linear relationships in the data and is sensitive to noise. The method of this embodiment replaces the traditional maximum likelihood estimation with a variational inference algorithm, improving the stability and generalization ability of factor analysis; introducing an adaptive Gaussian kernel function and a non-linear rotation transformation to enhance the ability to extract non-linear environmental features; dynamically adjusting the feature weights through between-class scatter and within-class scatter analysis, improving the identification accuracy of key environmental factors.

[0043] In an alternative embodiment, a hierarchical adaptive neural network is constructed based on the environmental feature vector, scene generation is performed according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, the visual comfort is calculated by the matching degree between the light flux distribution and the human visual response characteristics, and the lighting scene parameters are iteratively optimized by using a dynamic programming algorithm, including: A hierarchical adaptive neural network is constructed based on the environmental feature vector, and the weight coefficients of each layer of nodes in the neural network are dynamically adjusted through the feature correlation matrix, and the environmental feature vector is mapped and transformed to obtain a scene feature representation; A light flux spatial distribution function is constructed by using the scene feature representation, the initial light flux distribution of each lighting unit in the target area is calculated according to the light flux spatial distribution function, and the light flux superposition coefficient between lighting units is calculated by combining the three-dimensional spatial coordinates, light emission direction, and occlusion coefficient of the lighting unit, and the initial light flux distribution and the light flux superposition coefficient are combined and calculated to generate a light flux distribution map of the target area; The spectral energy distribution characteristics are extracted from the light flux distribution map of the target area, the dark adaptation threshold and the light adaptation threshold are calculated according to the pre-determined spectral response curve of the human cone cells, a visual comfort evaluation function is established, and the visual comfort parameters are obtained by calculating the matching degree of the spectral energy distribution characteristics; The visual comfort parameters are mapped into state transition probabilities, a state transition matrix of the lighting scene parameters is constructed, and the lighting scene parameters are iteratively calculated by dynamic programming through the state transition matrix to generate optimal lighting scene parameters.

[0044] In a specific embodiment, the lighting scene generation system based on the environmental feature vector starts to work by collecting environmental feature data. The environmental feature vector includes key parameters such as indoor temperature, humidity, illuminance, personnel density, and time period type. These parameters are collected in real time through a sensor network, and outliers are eliminated through data preprocessing and standardized to form a feature vector with a dimension of 15.

[0045] Construct a four - layer adaptive neural network, including an input layer, two hidden layers, and an output layer. The number of nodes in the input layer matches the dimension of the environmental feature vector, which is 15. The first hidden layer is set with 32 nodes, the second hidden layer with 16 nodes, and the output layer with 8 nodes corresponding to the scene feature representation. Dynamically adjust the network weights by calculating the feature correlation matrix. When the correlation coefficient between temperature and light intensity is detected to reach 0.85, automatically enhance the connection weight between these two nodes to 1.2 times. For feature pairs with a correlation below 0.2, such as humidity and personnel density, the connection weight is reduced to 0.8 times, thereby optimizing the network structure. The initial weights are randomly generated in the interval [-0.5, 0.5], use ReLU as the activation function, and apply batch normalization technology to improve the training stability.

[0046] After completing the feature mapping, construct a luminous flux spatial distribution function using the obtained 8 - dimensional scene feature representation. This function uses an improved Gaussian distribution model to describe the scattering law of light in space. For an office area with an 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; for the lighting unit located at (7, 2), it is set to 750 lumens. Further consider the spatial position information of the lighting units, such as the height of ceiling - mounted lighting equipment is 2.8 meters, and the height of wall - mounted lighting equipment is 1.5 meters, and record the luminous direction angle.

[0047] Establish a luminous flux superposition model to calculate the influence between lighting units. When there is an obstacle between two lighting units, the occlusion coefficient is set to 0.6; when there is no occlusion, the coefficient is 1.0. Through iterative calculation, the system generates a luminous flux distribution map of the target area, showing that the luminous flux in the central area reaches 950 lumens and in the edge area is 650 lumens, forming a smooth transition lighting effect.

[0048] After the luminous flux distribution map is generated, the system extracts the spectral energy distribution characteristics, including the energy distribution curve in the visible light spectrum range. The system has built - in spectral response curve data of three types of human cone cells (S - type, M - type, L - type). Under dark environmental conditions, the response threshold of S - type cells is set to 10 cd / m 2 ,the thresholds of M - type and L - type cells are 15 cd / m 2 and 12 cd / m 2 respectively; under bright environmental conditions, the adaptation thresholds of these three types of cells are increased to 100 cd / m 2 、120 cd / m 2 and 110 cd / m 2 respectively.

[0049] The established visual comfort evaluation function comprehensively considers three key indicators: the matching degree between the spectral energy distribution and the human visual response, the illuminance uniformity, and the glare index. For an ordinary office environment, the weight of the spectral matching degree is set to 0.5, the weight of the illuminance uniformity is 0.3, and the weight of the glare index is 0.2. The calculated visual comfort parameter of the current light flux distribution is 0.78, which is close to but does not reach the ideal value of 0.85.

[0050] To optimize the lighting parameters, the visual comfort parameter is mapped to a state transition probability. The constructed state transition matrix includes three dimensions: the brightness level of the lighting unit, the color temperature parameter, and the irradiation angle. The brightness level is divided into 10 levels, the color temperature range is from 2700K to 6500K, divided into 8 grades; the irradiation angle is divided into 6 intervals within the range of 0° to 180°. The system uses the value iteration method for dynamic programming calculation, sets the discount factor to 0.95, and the maximum number of iterations to 100.

[0051] After 57 iterations, the generated optimal lighting scene parameters show that the brightness of the central lighting unit is adjusted to 920 lumens, the color temperature is set to 4200K, and the irradiation angle is 75°; the brightness of the edge lighting unit is adjusted to 680 lumens, the color temperature is set to 3800K, and the irradiation angle is 60°. The visual comfort parameter calculated under this set of parameters is increased to 0.87, exceeding the target value of 0.85, and at the same time, the energy consumption is reduced by 12%.

[0052] In this embodiment, by dynamically adjusting the node weights of each layer through a hierarchical adaptive neural network and a feature correlation matrix, the scene features mapped by the environmental feature vectors are made more discriminative and robust, improving the basic quality of subsequent light flux calculations; constructing a light flux spatial distribution function based on the scene features, and calculating the superposition coefficient by combining three-dimensional coordinates, light emission direction, and occlusion coefficient, realizing a high-precision simulation of the interaction of multiple light sources in the target area; extracting spectral energy characteristics from the light flux distribution map, calculating the dark / light adaptation thresholds with reference to the spectral response curve of the human cone cells, and quantifying the visual comfort through a matching degree function, making the evaluation closer to human perception; mapping the visual comfort parameter to a state transition probability, constructing a state transition matrix of the lighting scene parameters, and using dynamic programming iterative search to automatically obtain the global optimal lighting scheme under a given scene; forming a closed-loop adaptive process from environmental features to scene features, comfort evaluation, and then to parameter optimization; each step has clearly defined physical or physiological indicators, improving the interpretability and deployment trust of the method.

[0053] In an alternative implementation, extracting the spectral energy distribution characteristics from the light flux distribution map of the target area, calculating the dark adaptation threshold and the light adaptation threshold according to the pre-determined spectral response curve of the human cone cells, establishing a visual comfort evaluation function, and calculating the matching degree of the spectral energy distribution characteristics to obtain the visual comfort parameter includes: Perform a two-dimensional discrete wavelet transform on the target area luminous flux distribution map to obtain a multi-scale spectral coefficient matrix. Fit the multi-scale spectral coefficient matrix with a Gaussian mixture distribution, and through expectation-maximization iteration, calculate and generate a spectral energy probability density function. Based on the spectral energy probability density function, calculate and determine the spectral energy distribution characteristics; Perform a two-way competition operation on the spectral response curve of the human cone cells and multiple groups of historical visual response data. Obtain a dynamic correction coefficient through the mutual game between the response parameters and the historical parameters. Perform a convolution operation on the spectral response curve based on the dynamic correction coefficient to generate a corrected response curve. Calculate the dark adaptation threshold and the light adaptation threshold according to the corrected response curve; Perform multi-layer feature extraction calculations on the spectral energy distribution characteristics, the dark adaptation threshold, and the light adaptation threshold to obtain a visual energy adaptation coefficient, a visual brightness response coefficient, and a visual time adjustment coefficient, and construct a visual comfort evaluation function; Perform dynamic weight iterative optimization on the spectral energy distribution characteristics and the visual comfort evaluation function, and calculate the visual comfort parameter.

[0054] In a specific embodiment, process the target area luminous flux distribution map, perform a two-dimensional discrete wavelet transform operation, use the Haar wavelet basis function to perform multi-layer decomposition on the luminous flux distribution map, and respectively obtain the detail coefficients in the horizontal, vertical, and diagonal directions. In actual operation, select three-layer wavelet decomposition. The high-frequency component matrices obtained after each layer of decomposition can be represented by three directions: H (horizontal), V (vertical), and D (diagonal), and the low-frequency component is represented by A. After three-layer decomposition, a total of ten sub-bands are finally obtained: A3, H3, V3, D3, H2, V2, D2, H1, V1, D1. These sub-bands together constitute the multi-scale spectral coefficient matrix, representing the energy characteristics of the luminous flux distribution at different scales and directions. For example, for a 1024×1024 pixel luminous flux distribution map, the dimension of the third-layer low-frequency coefficient matrix A3 is 128×128.

[0055] The obtained multi-scale spectral coefficient matrix is fitted with Gaussian mixture distribution. A three-group Gaussian distribution model is used for mixture fitting, and each Gaussian distribution has different mean and covariance parameters. Iterative optimization is carried out through the expectation-maximization algorithm, with the iteration threshold set to 0.001 and the maximum number of iterations set to 100 times. In a typical case, after 72 iterations of convergence, the weights of the three Gaussian components are 0.45, 0.35, and 0.2 respectively, the mean vectors are [0.78, 0.62], [0.41, 0.58], and [0.25, 0.31], and the covariance matrices represent the distribution characteristics of each component. Based on these parameters, a spectral energy probability density function is constructed. 100 equally spaced sampling points are taken in the sample spectral energy space, the probability density values of each point are calculated, and a spectral energy distribution feature vector is obtained through normalization. The dimension of this vector is 100, which characterizes the statistical distribution characteristics of the spectral energy in the target area.

[0056] In the aspect of processing the spectral response curves of human cone cells, the standard response curve data of three known types of cone cells, namely L-type, M-type, and S-type, to the visible spectrum (380nm - 780nm) are used, with a sampling interval of 5nm, forming three 81-dimensional vectors. Combining multiple sets of collected historical visual response data (including 50 sets of human eye adaptation parameters under different lighting environments), a two-way competition operation is performed. Specifically, the similarity scores are calculated for each set of historical data and the current response curve, and the top 10 historical data with the highest similarity rankings are selected for weighted fusion, and the weight coefficients are distributed 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 non-linear distribution in the range of 380 - 780nm, with a larger correction coefficient in the middle wavelength band (500 - 600nm) (average value 1.15) and smaller correction coefficients at both ends of the wavelength band (average value 0.92). The original spectral response curve is convolved with this dynamic correction coefficient to generate a corrected response curve. In an actual case, 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.

[0057] According to the corrected response curve, the dark adaptation threshold and the light adaptation threshold are calculated. The dark adaptation threshold is calculated using the increasing brightness detection method. Starting from a brightness value close to zero (0.01cd / m 2 ), increasing in steps of 0.01cd / m 2 , the activation degree of the three types of cone cells is calculated for each brightness value. When the activation degree first exceeds the set threshold (0.15), the corresponding brightness value is the dark adaptation threshold. In the case, the obtained dark adaptation threshold is 0.13cd / m 2 . The light adaptation threshold is calculated using a similar method, but starting from a high brightness (10000cd / m 2) Decremental calculation: When the oversaturation response degree of the three types of cone cells is lower than the set threshold (0.85), the corresponding brightness value is the light adaptation threshold. In the case, the obtained light adaptation threshold is 2450 cd / m 2 .

[0058] The construction process of the visual comfort evaluation function involves multi-layer feature extraction. First, calculate the visual energy adaptation coefficient based on the spectral energy distribution feature and the adaptation threshold. The specific method is to perform a dot product operation on the distribution value of the spectral energy at each wavelength point and the response curve value of the three types of cone cells at that wavelength point, and then accumulate to obtain the total stimulus amount. Then compare it with the difference between the dark adaptation threshold and the light adaptation threshold to generate a normalized visual energy adaptation coefficient, with a value range of [0, 1]. In the case, the calculated adaptation coefficient is 0.78. The visual brightness response coefficient is obtained by calculating the ratio of the weighted mean of the spectral energy to the light adaptation threshold. In the case, this coefficient is 0.62. The visual time adjustment coefficient considers the time characteristics of the change in the light environment and is obtained by calculating the standard deviation of the spectral data sampled continuously for 10 seconds and then performing a non-linear mapping. In the case, this coefficient is 0.85.

[0059] When constructing the visual comfort evaluation function, the above three coefficients are integrated through a weighted combination method, and the weight coefficients are 0.4, 0.35, and 0.25 respectively to form a complete evaluation function. This function has non-linear characteristics and different sensitivities to the changes of the three parameters. Perform dynamic weight iterative optimization on the spectral energy distribution feature and the visual comfort evaluation function, using the gradient descent method, setting the learning rate to 0.05, and iterating 50 times. Each iteration adjusts the weights to make the evaluation result more in line with the actual visual comfort feeling. In the case, the finally obtained visual comfort parameter is 0.726, indicating that the current light environment has good visual comfort.

[0060] Existing visual comfort evaluation technologies mainly adopt simple illuminance threshold methods or brightness uniformity evaluation methods, only considering the physical quantities of the lighting system and ignoring the physiological characteristics of the human visual system and its dynamic adaptation process. For example, traditional methods usually use 300 - 500 lux as a fixed standard for indoor lighting comfort evaluation, or evaluate uniformity using the ratio of maximum brightness to minimum brightness. These methods lack consideration of spectral energy distribution and understanding of the complexity of the human visual system. In the existing technologies, Fourier transform is mostly used for spectral analysis, which is difficult to capture local features and multi-scale information in the light environment; visual response models mostly use CIE standard observer functions with fixed parameters and cannot adapt to visual response differences under different individuals and environmental conditions. The method of this embodiment replaces traditional Fourier analysis with two-dimensional discrete wavelet transform to more precisely capture the multi-scale and directional features of the light flux distribution; introduces a Gaussian mixture distribution model to probabilistically model spectral coefficients to more comprehensively represent the statistical characteristics of spectral energy; innovatively proposes a two-way competition operation mechanism to achieve dynamic correction of the cone cell response curve, enabling the visual evaluation model to adapt to different individuals and environmental conditions; constructs a multi-level visual comfort evaluation function, comprehensively considering factors such as energy adaptation, brightness response, and time adjustment, which is more in line with the physiological mechanism of the formation of human visual comfort.

[0061] As Figure 4 shown, the computational response time performance of four different algorithms under the condition of environmental light intensity varying from 0 lx to 5000 lx is presented. Different marker symbols are used in the figure to represent different algorithm implementations: circular markers represent the original HMC Monte Carlo algorithm, which is the most basic probability sampling method with the longest computational response time, gradually decreasing from 67.8 ms at 0 lx to 52.4 ms at 5000 lx; triangular markers represent the Markov chain sampling algorithm, which improves computational efficiency by optimizing the sampling strategy, and the response time decreases from 53.4 ms at 0 lx to 35.2 ms at 5000 lx; diamond markers represent the optimization algorithm based on Bayesian inference, which introduces prior knowledge to guide the sampling process, improving the computational response time from 42.7 ms at 0 lx to 23.9 ms at 5000 lx; star markers represent the adaptive Gaussian kernel density estimation algorithm proposed in this paper, which achieves optimal computational efficiency by dynamically adjusting kernel function parameters and sampling weights, and the response time significantly decreases from 28.5 ms under 0 lx conditions to 12.8 ms at 5000 lx. From the overall trend, the response times of all algorithms decrease with the increase of environmental light intensity, indicating 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.

[0062] In an alternative embodiment, the lighting scene parameters are input into a grouped collaborative learning network. The control groups are divided according to the spatial distribution relationship of the lighting units. An optimization mechanism of intra-group local competition and inter-group global cooperation is adopted. Combining with the energy consumption dynamic prediction model, the calculation of the optimal control strategy includes: 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 light influence matrix between the lighting units. Perform an exponential decay operation according to the spatial distance matrix and the light influence matrix to obtain the correlation intensity between the lighting units. Divide the lighting units with a correlation intensity greater than the preset correlation degree threshold into the same control group; Perform a combined operation on the energy consumption parameters, lighting effect parameters, and intra-group interference degree parameters of the lighting units in each control group to obtain the intra-group optimization target value. Adopt a two-way competition mechanism to make the control parameters of the lighting units confront and adjust each other, and correct the parameter values through the intra-group energy balance constraint to generate the initial control parameters of the lighting units within the group; Perform a combined calculation on the intra-group optimization target values of each control group and the cross-interference degree between the control groups to obtain the inter-group cooperation target value. Optimize and solve the inter-group cooperation target value by the distributed alternating direction multiplier method, and output the inter-group cooperation control parameters; Extract the temporal characteristics of the historical scene parameters, load characteristic parameters, and operating state parameters to obtain the predicted energy consumption data. Perform dynamic weight iteration calculation on the predicted energy consumption data and the inter-group cooperation control parameters to generate the optimal control strategy for the lighting units.

[0063] In a specific embodiment, the acquisition of the 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 units through a sensor array and records it as three-dimensional coordinates (x, y, z); measures the luminous flux value of each lighting unit through a photometer, and the typical value range is 300 - 1500 lumens; the luminous angle parameter is obtained by measuring the main optical axis direction and the divergence angle of the lighting unit, and generally the divergence angle is between 30° and 120°. Taking Office Area A as an example, 20 LED lighting units are deployed in this area, numbered L1 - L20, where L1 is located at the coordinates (2.5, 3.0, 2.8), the luminous flux is 800 lumens, and the main luminous angle is 45°.

[0064] The spatial distance matrix is calculated using the Euclidean distance formula to calculate the physical distance between any two lighting units. For the lighting units L1(2.5, 3.0, 2.8) and L2(5.0, 3.5, 2.8) in the example, the calculated spatial distance is 2.55 meters. Calculate pairwise for all 20 lighting units to form a 20×20 spatial distance matrix D.

[0065] The calculation of the light influence matrix takes into account the luminous flux, luminous angle, and occlusion factors of the lighting units. Based on the inverse-square law of light attenuation and combined with the cosine correction of the luminous angle, the system calculates the light influence intensity between any two lighting units. Taking the influence of L1 on L2 as an example, considering the luminous flux of L1 is 800 lumens, the distance between the two is 2.55 meters, and the included angle of the luminous angles is 15°, the calculated light influence intensity is 75.2 lumens per square meter. Similar calculations are performed for all lighting units to form a 20×20 light influence matrix I.

[0066] The calculation of the correlation intensity uses an exponential decay function. For each element of the distance matrix D and the light influence matrix I, exponential transformations are respectively applied with attenuation coefficients α = 0.3 and β = 0.5, and then the elements of the two transformed matrices are weighted and summed with a weight ratio of 0.4 for the distance factor and 0.6 for the light influence factor to obtain the correlation intensity matrix R between the lighting units. The range of the correlation intensity value is between 0 and 1, and the larger the value, the closer the correlation.

[0067] For the control group division, a correlation threshold of 0.65 is set, and the lighting units with a correlation intensity greater than this threshold are grouped into the same control group. By analyzing the correlation intensity matrix R through a clustering algorithm, the 20 lighting units are finally divided into 4 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}.

[0068] The calculation of the within-group optimization target value comprehensively considers the energy consumption parameter, lighting effect parameter, and within-group interference parameter. The energy consumption parameter includes the product of the rated power and the current brightness level of each lighting unit; the lighting effect parameter includes the target illuminance achievement rate and the illuminance uniformity; the within-group interference parameter characterizes the degree of mutual influence between the lighting units within the group. Taking the 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; the within-group interference parameter is 0.15. Through combined operations, a weight of 0.3 is given to the energy consumption parameter, a weight of 0.5 is given to the lighting effect parameter, and a weight of 0.2 is given to the within-group interference parameter, resulting in a within-group optimization target value of 0.82 for G1.

[0069] The two-way competition mechanism is a process of mutually opposing and adjusting the control parameters of the lighting units within a group. The system adopts an iterative game algorithm, enabling each lighting unit to adjust its brightness level and angle parameters according to its own state and the influence of other units. In each round of iteration, the lighting units play against each other based on the requirements of the target area and the energy consumption situation, dynamically adjusting the parameter values. Taking Group G1 as an example, after 15 rounds of iteration, the brightness level of L1 is adjusted from the initial 85% to 78%, and the emission angle is slightly adjusted from 45° to 42°; the brightness level of L3 is adjusted from 80% to 76%, and the angle remains unchanged; similar adjustments are also made to other units, ultimately achieving energy balance within the group.

[0070] The calculation of the inter-group cooperation target value combines the intra-group optimization target values of each control group and the cross-interference degree between control groups. The cross-interference degree is calculated by analyzing the light superposition effect between the lighting units of different control groups. For the four control groups in the example, the inter-group cross-interference degree matrix is a 4×4 matrix, and the element values range from 0 to 0.5. The cross-interference degree between G1 and G2 is 0.25, indicating a certain degree of light influence between the two groups. The optimization target values of each group are weighted and combined with the cross-interference degree to obtain the inter-group cooperation objective function.

[0071] The distributed alternating direction method of multipliers is used to solve the inter-group cooperation optimization problem. This method decomposes the global optimization problem into multiple local sub-problems. Each control group solves the sub-problems locally and then exchanges information, achieving global convergence by introducing Lagrange multipliers and augmented terms. During the iteration process, each control group updates its own control strategy based on the current parameters and the information of other groups, iterating until convergence. Through 30 iterations, the system obtains the globally approximate optimal inter-group cooperation control parameters.

[0072] The energy consumption prediction model is constructed based on historical scenario parameters, load characteristic parameters, and operating state parameters. The system collects the operation data of the lighting system for the past 90 days, including the on-off state, brightness level, energy consumption data, etc. of the lighting units. Seasonal patterns, periodic changes, and trend characteristics are extracted through time series analysis to construct the prediction model. Predicting the energy consumption for the next hour in the example scenario, the model predicts that the total energy consumption is 320W, and the energy consumption of each control group is G1: 74W, G2: 92W, G3: 85W, G4: 69W.

[0073] The optimal control strategy generation combines the predicted energy consumption data with the inter-group collaborative control parameters, and determines the final control strategy through dynamic weight iterative calculation. According to the current scenario requirements and predicted energy consumption, the system dynamically adjusts the weight ratio of energy efficiency and lighting effect. When energy is in short supply, the system increases the weight of energy efficiency; when high lighting quality is required, the weight of lighting effect is increased. Through 5 rounds of dynamic weight iterative calculation, the final control strategy for lighting units is generated, including the brightness level, angle parameters and switching timing of each lighting unit, achieving the optimal balance between lighting effect and energy consumption.

[0074] In this embodiment, control groups are divided based on the correlation intensity of spatial distance and lighting influence, so that lighting units with large mutual interference are grouped together, facilitating grouped management and collaborative optimization; within each control group, two-way competition and energy balance constraints are combined with energy consumption, lighting effect and interference degree to achieve the optimal initial setting of the intra-group control parameters; the optimization objectives of each group and the cross-interference degree are integrated, and the distributed alternating direction multiplier method is used to solve, ensuring the consistency of the inter-group control parameters and reducing the negative impact between adjacent groups; using the temporal characteristics of historical scenarios, loads and operating states, future energy consumption is predicted, and it is dynamically weighted and iterated with the inter-group collaborative control parameters to adjust the control strategy in real time, improving energy conservation 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 the global optimality of the final control strategy in terms of overall energy consumption, lighting effect and system stability.

[0075] In an alternative implementation manner, the intra-group optimization target values of each control group and the cross-interference degree between control groups are combined and calculated to obtain the inter-group collaborative target value, and the distributed alternating direction multiplier method is used to optimize and solve the inter-group collaborative target value, and the output inter-group collaborative control parameters include: Calculate the luminous flux, included angle, and distance parameters of lighting units between different control groups to obtain the inter-group cross-interference degree matrix, generate a dynamic weight matrix through exponential normalization operation, and calculate the inter-group coupling coefficient of each control group; Perform compensation calculation on the intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient to obtain the inter-group collaborative target value; Extract the temporal characteristics of the environmental temperature parameter, population density parameter, and meteorological change parameter to obtain the temporal sensitivity matrix, introduce local variables, global variables and Lagrange multipliers to the inter-group collaborative target value to construct the first augmented Lagrangian function, and perform response compensation on the temporal sensitivity matrix and the first augmented Lagrangian function to obtain the second augmented Lagrangian function; For the second augmented Lagrangian function, fix the global variables and Lagrange multipliers, and each control group independently performs gradient descent to obtain local variables; fix the local variables and Lagrange multipliers, and perform gradient descent in combination with global constraints to obtain global variables; perform a bi-directional mapping operation to generate a variable mapping matrix and update the Lagrange multipliers. Perform a difference operation on the local variables and global variables to obtain the original residual, calculate the difference between adjacent iterative global variables to obtain the dual residual, adjust the penalty parameter according to the ratio of the original residual to the dual residual, and input it into the second augmented Lagrangian function for iterative operation. Terminate the iteration when the original residual is less than the first preset threshold and the dual residual is less than the second preset threshold, and calculate the inter-group collaborative control parameter according to the local variables and global variables.

[0076] In a specific embodiment, calculate the cross-interference matrix of the lighting units between different control groups. 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 respectively. For example, for the lighting unit a of control group A and the lighting unit b of control group B, assume that the luminous flux of a is 1000 lumens, the luminous flux of b is 800 lumens, the angle between them is 45 degrees, and the distance is 3 meters. Then the cross-interference degree can be calculated as 176.8 according to the inverse law of light intensity and the cosine attenuation formula. Calculate the interference degree pairwise for all lighting units between control groups to form an N×N-dimensional cross-interference matrix D.

[0077] Generate a dynamic weight matrix W through 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 on control group B is 176.8. After exponential transformation (e -176.8 / 1000 ) and normalization processing, the weight value 0.83 is obtained. Calculate the inter-group coupling coefficient of each control group according to the weight matrix W. For example, the coupling coefficient of control group A is the weighted average of its weights with all other groups, assumed to be 0.75.

[0078] Perform a compensation calculation on the intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient to obtain the inter-group collaborative target value. Taking control group A as an example, if its intra-group optimization target value is illuminance 500 lux and the inter-group coupling coefficient is 0.75, then its inter-group collaborative target value can be calculated as 500×(1 + 0.75×0.2)=575 lux, where 0.2 is the compensation coefficient. Perform similar operations on all N control groups to form a collaborative target value vector.

[0079] Extract the temporal characteristics of environmental temperature parameters, population density parameters, and meteorological change parameters. Taking an office area as an example, collect environmental data for 24 hours, with one group per hour, including temperature (22 - 28°C), population density (0 - 100 people per square meter), light intensity (0 - 1000 lux), etc. Extract the temporal change characteristics of these parameters through the sliding window method to generate a temporal sensitivity matrix S.

[0080] Based on the distributed alternating direction multiplier method, construct an augmented Lagrangian function to solve the inter-group collaborative control problem. For N control groups, introduce local variables x = {x1, x2,..., x n} to represent the control parameters of each control group, the global variable z to represent the control parameters that satisfy the global constraints, and the Lagrange multiplier y. Construct the first augmented Lagrangian function, and perform response compensation on the temporal sensitivity matrix S and this function to obtain the second augmented Lagrangian function L.

[0081] In the iterative solution process, fix the global variable z and the Lagrange multiplier y, and each control group independently executes gradient descent to update its own local variable x. Taking control group A as an example, assume that the initial control parameter x1 is [450 lux, 3000 K], and after gradient descent iteration, it is updated to [480 lux, 3200 K]. Fix all local variables x and the Lagrange multiplier y, and perform gradient descent to update the global variable z in combination with the global constraints (such as the total power not exceeding 10000 W). Perform two-way mapping operations to generate a variable mapping matrix M for updating the Lagrange multiplier y.

[0082] Calculate the original residual r = |x - z| and the dual residual s = |z (k+1) - z k |. For example, in a certain iteration, the local variable x = [480 lux, 3200 K], and the global variable z = [470 lux, 3150 K], then the original residual r = |(480, 3200) - (470, 3150)| = 56.8. If the global variables in two adjacent iterations are z k = [470 lux, 3150 K] and z (k+1) = [472 lux, 3160 K], then the dual residual s = |(472, 3160) - (470, 3150)| = 14.1.

[0083] Adjust the penalty parameter ρ according to the ratio of the original residual to the dual residual. If r / s > 10, double ρ; if s / r > 10, halve ρ; otherwise, keep ρ unchanged. Input the updated penalty parameter ρ into the second augmented Lagrangian function L to continue the iteration.

[0084] Set the first preset threshold to 5 and the second preset threshold to 10, and terminate the iteration when the original residual r < 5 and the dual residual s < 10. Finally, based on the converged local variable x and global variable z, the inter-group cooperative control parameters are calculated. Taking control group A as an example, the final cooperative control parameters are illuminance 475 lux, color temperature 3175 K, and the power is adjusted to 92% of the original. These parameters will be sent to the corresponding control group actuators to achieve the inter-group cooperative control of the lighting system, effectively balancing the lighting effects and energy consumption of each control group.

[0085] Most existing lighting control systems adopt partition-independent control or simple centralized control methods, lacking comprehensive consideration of the interactive effects between different control groups. Traditional lighting systems usually adopt preset control strategies based on schedules or simple feedback control based on light sensors. For example, when the illuminance in a certain area is insufficient, the lighting intensity in that area is increased, while ignoring the light spillover effect in adjacent areas. There are also methods in the prior art that use a central controller for centralized optimization, 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, the response of traditional lighting control systems to environmental parameters is mostly an instantaneous reaction, lacking the ability to extract and predict the characteristics of temporal changes, resulting in control lags or oscillations when the environment changes rapidly. The method of this embodiment constructs an inter-group cross-interference degree matrix to accurately quantify the mutual influence between different lighting control groups; introduces a distributed optimization framework to decompose the global optimization problem into multiple local sub-problems that can be solved in parallel, greatly improving the computational efficiency; innovatively introduces a time-sensitive matrix into the augmented Lagrangian function to enable the control system to make predictive responses to the temporal characteristics of environmental changes; designs a dynamic penalty parameter adjustment mechanism to improve the convergence speed and stability of the algorithm.

[0086] In a multi-control group scenario, it saves power consumption compared with the traditional independent control method; the response time is shortened, and the adaptation speed to environmental changes is significantly improved; the computational complexity of the algorithm is reduced from O(n³) of the traditional centralized control to O(n).

[0087] The intelligent green landscape lighting adaptive control system based on deep learning in the embodiments of the present invention includes: The first unit is used to collect the light intensity, pedestrian flow 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 the environmental parameter data, fuse spatial and temporal features using a parallel residual network to obtain multi-modal features, perform transfer learning on the multi-modal features using bidirectional mapping, determine the number of factors based on a dynamic association matrix, and generate an environmental feature vector through variational inference and kernel density estimation; The third unit is used to construct a hierarchical adaptive neural network based on the environmental feature vector, generate a scene according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculate the visual comfort through the matching degree between the light flux distribution and the human visual response characteristics, and iteratively optimize and generate the lighting scene parameters by using the dynamic programming algorithm; The fourth unit is used to input the lighting scene parameters into the grouped 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 cooperation between groups, and calculate the optimal control strategy in combination with the energy consumption dynamic prediction model; The fifth unit is used to generate the adjustment instructions for the lighting equipment based on the optimal control strategy and execute the lighting adjustment.

[0088] In the third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0089] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0090] The present invention can be a method, a device, a system and / or a computer program product. The computer program product may include a computer-readable storage medium on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and 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. An adaptive control method for intelligent green landscape lighting based on deep learning, characterized in that Including: Collect the light intensity, pedestrian flow density, and meteorological data of the target area through environmental sensors to obtain environmental parameter data; Perform standardized preprocessing on the environmental parameter data, use a parallel residual network to fuse spatial-temporal features to obtain multi-modal features, perform transfer learning on the multi-modal features using bidirectional mapping, determine the number of factors based on the dynamic correlation matrix, and generate an environmental feature vector through variational inference and kernel density estimation; Construct a hierarchical adaptive neural network based on the environmental feature vector, generate a scene according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculate the visual comfort through the matching degree between the light flux distribution and the human visual response characteristics, and use a dynamic programming algorithm to iteratively optimize and generate lighting scene parameters; Input the lighting scene parameters into a grouped collaborative learning network, divide the control groups according to the spatial distribution relationship of the lighting units, adopt an optimization mechanism of local competition within the group and global cooperation between groups, and calculate the optimal control strategy in combination with the energy consumption dynamic prediction model; Generate an adjustment instruction for the lighting device based on the optimal control strategy and execute the lighting adjustment.

2. The method according to claim 1, wherein Performing standardized preprocessing on the environmental parameter data, using a parallel residual network to fuse spatial-temporal features to obtain multi-modal features, performing transfer learning on the multi-modal features using bidirectional mapping, determining the number of factors based on the dynamic correlation matrix, and generating an environmental feature vector through variational inference and kernel density estimation includes: Preprocess and standardize the environmental parameter data to obtain standardized environmental data; Input the standardized environmental data into the spatial feature extraction branch of the parallel residual network, obtain spatial features through a pyramidal convolution structure, and at the same time input the standardized environmental data into the temporal feature extraction branch of the parallel residual network, obtain temporal features through a causal convolution structure, and fuse the spatial features and temporal features according to the attention weights to generate multi-modal non-linear features; Input the multi-modal non-linear features into the source domain feature space and the target domain feature space respectively for bidirectional mapping, establish a cyclic consistency constraint, calculate the uncertainty distribution of the mapped features, and reorganize the mapped features according to the uncertainty distribution to obtain transfer features; Input the transfer features into the mutual information calculation unit to construct a dynamic correlation matrix, and use the spectral clustering method to determine the number of factors from the dynamic correlation matrix; Substitute the number of factors into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings, perform non-linear rotation on the factor loadings using adaptive kernel density estimation and weight them based on feature discrimination, and output the environmental feature vector.

3. The method according to claim 2, wherein Substitute the number of factors into the variational inference algorithm to optimize the factor loadings to obtain the optimal factor loadings, perform non-linear rotation on the factor loadings using adaptive kernel density estimation and weight them based on feature discrimination, and output the environmental feature vector includes: Input the number of factors into the variational inference algorithm, construct a Gaussian prior distribution of the factor loadings, set a variational posterior approximation in the form of a normal distribution for the variational parameters, and iteratively update the mean parameter and variance parameter of the variational posterior distribution by minimizing the KL divergence criterion to obtain the optimal factor loadings; Construct an adaptive Gaussian kernel function for the optimal factor loadings, dynamically adjust the kernel bandwidth parameter according to the sample distribution characteristics, calculate the kernel density estimate of the factor loadings, construct a rotation optimization objective function based on the kernel density estimate, iteratively optimize the rotation matrix through gradient ascent, perform a non-linear rotation transformation on the optimal factor loadings to obtain the rotated factor loadings; Calculate the between-class scatter matrix and within-class scatter matrix of the rotated factor loadings, determine the feature discrimination degree according to the trace ratio of the scatter matrices, set weight coefficients for the rotated factor loadings based on the feature discrimination degree, and input the weighted factor loadings into a non-linear feature mapping function to generate environmental feature vectors.

4. The method according to claim 1, characterized in that Construct a hierarchical adaptive neural network based on the environmental feature vectors, generate scenes according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculate the visual comfort level through the matching degree between the light flux distribution and the human visual response characteristics, and use a dynamic programming algorithm to iteratively optimize and generate lighting scene parameters including: Construct a hierarchical adaptive neural network based on the environmental feature vectors, dynamically adjust the weight coefficients of each layer of nodes in the neural network through the feature correlation matrix, and map and transform the environmental feature vectors to obtain scene feature representations; Use the scene feature representations to construct a light flux spatial distribution function, calculate the initial light flux distribution of each lighting unit in the target area according to the light flux spatial distribution function, and combine the three-dimensional spatial coordinates, light emission direction, and occlusion coefficient of the lighting units to calculate the light flux superposition coefficient between the lighting units. Combine the initial light flux distribution and the light flux superposition coefficient to perform a combined operation to generate a light flux distribution map of the target area; Extract the spectral energy distribution characteristics from the light flux distribution map of the target area, calculate the dark adaptation threshold and light adaptation threshold according to the pre-determined spectral response curve of the human cone cells, establish a visual comfort evaluation function, and calculate the matching degree of the spectral energy distribution characteristics to obtain visual comfort parameters; Map the visual comfort parameters to state transition probabilities, construct a state transition matrix of the lighting scene parameters, and perform dynamic programming iterative calculations on the lighting scene parameters through the state transition matrix to generate optimal lighting scene parameters.

5. The method according to claim 4, wherein Extract the spectral energy distribution characteristics from the light flux distribution map of the target area, calculate the dark adaptation threshold and light adaptation threshold according to the pre-determined spectral response curve of the human cone cells, establish a visual comfort evaluation function, and calculate the matching degree of the spectral energy distribution characteristics to obtain visual comfort parameters including: Perform a two-dimensional discrete wavelet transform on the light flux distribution map of the target area to obtain a multi-scale spectral coefficient matrix, perform Gaussian mixture distribution fitting on the multi-scale spectral coefficient matrix, and calculate and generate a spectral energy probability density function through expectation maximization iteration, and calculate and determine the spectral energy distribution characteristics based on the spectral energy probability density function; Perform a two-way competitive operation on the spectral response curve of human cone cells and multiple sets of historical visual response data, obtain a dynamic correction coefficient through the mutual game between the response parameter and the historical parameter, perform a convolution operation based on the dynamic correction coefficient on the spectral response curve to generate a corrected response curve, and calculate the dark adaptation threshold and the light adaptation threshold according to the corrected response curve; Perform multi-layer feature extraction calculations on the spectral energy distribution characteristics, the dark adaptation threshold, and the light adaptation threshold to obtain the visual energy adaptation coefficient, the visual brightness response coefficient, and the visual time adjustment coefficient, and construct a visual comfort evaluation function; Perform dynamic weight iterative optimization on the spectral energy distribution characteristics and the visual comfort evaluation function, and calculate the visual comfort parameter.

6. The method according to claim 1, characterized in that, Input the lighting scene parameters into the grouped collaborative learning network, divide the control groups according to the spatial distribution relationship of the lighting units, adopt an optimization mechanism of intra-group local competition and inter-group global cooperation, and calculate the optimal control strategy in combination with the energy consumption dynamic prediction model, 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 light influence matrix between the lighting units, perform an exponential decay operation according to the spatial distance matrix and the light influence matrix to obtain the association strength between the lighting units, and divide the lighting units with an association strength greater than the preset association degree threshold into the same control group; Perform a combined operation on the energy consumption parameters, lighting effect parameters, and intra-group interference degree parameters of the lighting units in each control group to obtain the intra-group optimization target value, adopt a two-way competition mechanism to make the control parameters of the lighting units adjust against each other and correct the parameter values through the intra-group energy balance constraint, and generate the initial control parameters of the lighting units within the group; Perform a combined calculation on the intra-group optimization target values of each control group and the cross-interference degree between the control groups to obtain the inter-group cooperation target value, and optimize and solve the inter-group cooperation target value through the distributed alternating direction multiplier method, and output the inter-group cooperation control parameters; Perform time series feature extraction on the historical scene parameters, load characteristic parameters, and operating state parameters to obtain the predicted energy consumption data, and perform dynamic weight iterative calculation on the predicted energy consumption data and the inter-group cooperation control parameters to generate the optimal control strategy for the lighting units.

7. The method according to claim 6, wherein Perform a combined calculation on the intra-group optimization target values of each control group and the cross-interference degree between the control groups to obtain the inter-group cooperation target value, and optimize and solve the inter-group cooperation target value through the distributed alternating direction multiplier method, and output the inter-group cooperation control parameters, including: Calculate the luminous flux, included angle, and distance parameters of the lighting units between different control groups to obtain the inter-group cross-interference degree matrix, generate a dynamic weight matrix through exponential normalization operation, and calculate the inter-group coupling coefficient of each control group; Perform a compensation calculation on the intra-group optimization target value of each control group and the corresponding inter-group coupling coefficient to obtain the inter-group cooperation target value; Perform time series feature extraction on the environmental temperature parameter, the population density parameter, and the meteorological change parameter to obtain the time series sensitive matrix, introduce local variables, global variables, and Lagrange multipliers into the inter-group cooperation target value to construct the first augmented Lagrangian function, and perform response compensation on the time series sensitive matrix and the first augmented Lagrangian function to obtain the second augmented Lagrangian function; For the second augmented Lagrangian function, fix the global variables and Lagrange multipliers, and each control group independently performs gradient descent to obtain local variables; fix the local variables and Lagrange multipliers, and perform gradient descent in combination with global constraints to obtain global variables; perform two-way mapping operations to generate a variable mapping matrix and update the Lagrange multipliers. Perform a difference operation on the local variables and global variables to obtain the original residual, calculate the difference between adjacent iterative global variables to obtain the dual residual, adjust the penalty parameter according to the ratio of the original residual to the dual residual, and input it into the second augmented Lagrangian function for iterative operations. Terminate the iteration when the original residual is less than the first preset threshold and the dual residual is less than the second preset threshold, and calculate the inter-group collaborative control parameter based on the local variables and global variables.

8. An intelligent green landscape lighting adaptive control system based on deep learning, for implementing the method described in any one of the preceding claims 1-7, characterized in that, It includes: The first unit is used to collect the light intensity, pedestrian flow 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 the environmental parameter data, fuse spatial-temporal features using a parallel residual network to obtain multi-modal features, perform transfer learning on the multi-modal features using two-way mapping, determine the number of factors based on the dynamic association matrix, and generate an environmental feature vector through variational inference and kernel density estimation. The third unit is used to construct a hierarchical adaptive neural network based on the environmental feature vector, generate a scene according to the regional illuminance distribution law and the light flux superposition relationship between lighting units, calculate the visual comfort based on the matching degree between the light flux distribution and the human visual response characteristics, and iteratively optimize and generate lighting scene parameters using a dynamic programming algorithm. The fourth unit is used to input the lighting scene parameters into a grouped collaborative learning network, divide control groups according to the spatial distribution relationship of lighting units, adopt an optimization mechanism of intra-group local competition and inter-group global cooperation, and calculate the optimal control strategy in combination with an energy consumption dynamic prediction model. The fifth unit is used to generate an adjustment instruction for lighting equipment based on the optimal control strategy and perform lighting adjustment.

9. An electronic device, characterized in that, It includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images