Energy consumption prediction method and system based on green smart lighting for cultural tourism

Abstract

The present invention provides a method and system for predicting energy consumption of green intelligent lighting based on culture and tourism, which relates to the technical field of energy consumption prediction, including obtaining historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and constructing a deep probability graph model based on the preprocessed influencing factor data set using a variational inference algorithm, combining the pre-trained lighting energy consumption model with the deep probability graph model to obtain a prediction model adapted to the current scene; using a graph neural network algorithm, taking the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene as input, constructing a scene knowledge graph, generating energy consumption prediction values ​​through a multi-scale prediction algorithm, and generating green intelligent lighting control strategies based on the energy consumption prediction values ​​and in combination with the lighting equipment characteristic information in the scene knowledge graph using an intelligent decision-making algorithm, applying the generated green intelligent lighting control strategies to the lighting systems of cultural and tourism scenes to achieve intelligent control of lighting equipment.

Description

Technical Field

[0001] The present invention relates to energy consumption prediction technology, and in particular to a method and system for predicting energy consumption based on cultural tourism green smart lighting. Background Art

[0002] In the modern cultural and tourism industry, the management of lighting energy consumption has increasingly become a key factor in improving resource utilization efficiency and achieving sustainable development. Cultural and tourism scenes usually have complex functional zoning and dynamic tourist flows, which puts higher demands on the design and regulation of lighting systems. Existing lighting energy consumption prediction methods mostly rely on simple statistical models and lack the ability to deeply explore and dynamically adapt to influencing factors. Therefore, how to effectively integrate historical energy consumption data, real-time energy consumption data, and multi-dimensional influencing factor data, and make accurate predictions through advanced machine learning and graph neural network technologies, has become an urgent problem to be solved in the industry. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for predicting energy consumption of green intelligent lighting based on cultural tourism, which can solve the problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention,

[0005] A method for predicting energy consumption of green smart lighting based on cultural tourism is provided, including:

[0006] Obtain historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and use data cleaning algorithms, interpolation algorithms and anomaly detection algorithms to preprocess the data to obtain preprocessed energy consumption data sets and influencing factor data sets. Based on the preprocessed historical lighting energy consumption data, use the time series decomposition algorithm to extract features and generate energy consumption time series feature sets;

[0007] According to the preprocessed influencing factor data set, a multi-layer attention network is used to calculate the weights of the influencing factors of different time scales and spatial scales to obtain a weighted influencing factor feature set. The energy consumption time series feature set and the weighted influencing factor feature set are input into the multimodal feature fusion network to generate a unified feature representation. Based on the unified feature representation, a deep probabilistic graph model is constructed using a variational inference algorithm. The pre-trained lighting energy consumption model is combined with the deep probabilistic graph model to obtain a prediction model that adapts to the current scene.

[0008] Using the graph neural network algorithm, the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene are taken as input to construct a scene knowledge graph. The prediction model and scene knowledge graph adapted to the current scene are taken as input, and energy consumption prediction values ​​are generated through a multi-scale prediction algorithm. Based on the energy consumption prediction values ​​and combined with the lighting equipment characteristic information in the scene knowledge graph, an intelligent decision-making algorithm is used to generate a green smart lighting control strategy. The generated green smart lighting control strategy is applied to the lighting system of the cultural and tourism scene to realize the intelligent control of lighting equipment.

[0009] In an optional embodiment,

[0010] Based on the unified feature representation, the variational inference algorithm is used to build a deep probability graph model. The pre-trained lighting energy consumption model is combined with the deep probability graph model to obtain a prediction model that adapts to the current scene, including:

[0011] Based on the unified feature representation, a deep probabilistic graphical model containing latent variables and observed variables is constructed. The latent variables and context information related to the observed variables are input into a multi-layer perceptron, and the attention weight vector corresponding to each observed variable is calculated. The latent variables are weighted summed using the attention weight vector, and the observed variables are generated through nonlinear activation function processing.

[0012] The generated observed variables are taken as the mean of the Gaussian distribution, and the distribution variance is obtained by learning the neural network. The probability distribution of the observed variables is defined, and the prior distribution of the latent variables is multiplied by the conditional probability distribution of all observed variables to construct the joint distribution of the deep probability graph model. Based on the constructed joint distribution, the log-likelihood function is defined, and the standard normal distribution is selected as the basic distribution. Multiple reversible neural network layers including a normalization layer, a reversible mixing layer, and an affine coupling layer are designed. The basic distribution is transformed through the reversible neural network layer to construct a variational distribution.

[0013] Using the constructed variational distribution and joint distribution, derive the variational lower bound expression, calculate the variational lower bound, take the calculated variational lower bound as the optimization target, obtain a pre-trained large-scale lighting energy consumption model, combine the pre-trained large-scale lighting energy consumption model with the deep probability graph model, construct an initial combined prediction model, input the unified feature representation into the initial combined prediction model, and obtain the prediction result;

[0014] Based on the prediction results and actual energy consumption data, the prediction loss is calculated. The overall optimization objective function is constructed by combining the prediction loss and the calculated variational lower bound. Based on the overall optimization objective function, the stochastic gradient descent optimization algorithm is used to iteratively update the parameters of the combined prediction model. The iteration is repeated until the maximum number of iterations is met to obtain the final prediction model that adapts to the current scenario.

[0015] In an optional embodiment,

[0016] Combining the prediction loss and the calculated variational lower bound, the calculation formula for the overall optimization objective function is as follows:

[0017] ;

[0018] Among them, L represents the overall optimization objective function, N represents the number of data samples, and Y n represents the true value of the nth observed variable, represents the predicted value of the nth observed variable, λ represents the weight parameter, E q(Z|Y) logp(Z|Y) represents the expected value of the log-likelihood, q(Z|Y) represents the posterior distribution of the latent variable Z given the observed variable Y, and D KL (·) represents the KL divergence, p(Z|Y) represents the probability of the latent variable Z given the observed variable Y, and p(Z) represents the prior distribution.

[0019] In an optional embodiment,

[0020] Using the graph neural network algorithm, the topological structure information, functional zoning information and lighting equipment characteristics of the cultural and tourism scene are taken as input to build a scene knowledge graph. The prediction model and scene knowledge graph adapted to the current scene are taken as input, and the energy consumption prediction value is generated through the multi-scale prediction algorithm, including:

[0021] The topological structure information, functional zoning information and lighting equipment characteristic information of the cultural tourism scene are obtained as input data. The graph neural network algorithm is used, and the attention mechanism is introduced. By learning the importance weights of different neighbor nodes, the feature information of neighbor nodes is adaptively aggregated. At the same time, convolution operations are introduced in each layer of the graph neural network. Through the feature aggregation and transformation of the local neighborhood, the high-order feature representation of the node is extracted, and the scene knowledge graph containing the scene structured representation and semantic information is generated.

[0022] Obtain historical energy consumption data and influencing factor data of cultural and tourism scenarios, combine the scenario knowledge graph to input a prediction model that adapts to the current scenario, and use a multi-scale prediction algorithm to decompose the output results of the prediction model in the time dimension to obtain short-term, medium-term, and long-term prediction subsequences respectively;

[0023] Combined with the structured representation and semantic information in the scene knowledge graph, the sequence prediction model based on the attention mechanism is applied to the short-term prediction subsequence to generate the short-term lighting energy consumption prediction results. The functional partitioning information in the scene knowledge graph is integrated to apply the prediction model based on time series decomposition and periodicity detection to the medium-term prediction subsequence to generate the medium-term lighting energy consumption prediction results. The topological structure information in the scene knowledge graph is integrated to apply the prediction model based on causal inference and scenario analysis to the long-term prediction subsequence to generate the long-term lighting energy consumption prediction results.

[0024] An adaptive multi-scale fusion mechanism is introduced to learn the importance weights of short-term, medium-term and long-term lighting energy consumption prediction results, and adaptively fuse the prediction results of different scales to obtain the final energy consumption prediction value.

[0025] In an optional embodiment,

[0026] Applying a prediction model based on causal inference and scenario analysis to the long-term prediction subsequence, the long-term lighting energy consumption prediction results include:

[0027] Obtain historical lighting energy consumption data, scenario factor data, and long-term forecast subsequences for cultural and tourism scenes, where the scenario factor data include weather conditions, time characteristics, and implementation of energy-saving measures;

[0028] Fill missing values ​​and process outliers on the historical lighting energy consumption data, and standardize the processed historical lighting energy consumption data and scenario factor data to obtain preprocessed input data, and input the preprocessed input data into a pre-constructed generative adversarial network model, wherein the generator model adopts a long short-term memory network structure, and the discriminator model adopts a convolutional neural network structure, and the generator and the discriminator are optimized by alternating training to obtain a trained generative adversarial network model;

[0029] Based on the generator of the trained adversarial generative network model, a causal graph model is constructed, situational factors are used as nodes, and the causal relationship between nodes is determined by a structural learning algorithm. The causal graph model is used to apply a structural equation model to quantify the causal strength between nodes to obtain a causal relationship quantification result, and the causal relationship quantification result is integrated into the generator model to finally obtain a causally enhanced long-term lighting energy consumption prediction model;

[0030] Based on historical data distribution and expert knowledge, multiple groups of scenario factor combinations are generated and organized into a scenario matrix, where each row represents a future scenario and each column corresponds to a scenario factor. The long-term prediction subsequence and scenario matrix are input into the causally enhanced long-term lighting energy consumption prediction model to generate long-term lighting energy consumption prediction results under multiple scenarios.

[0031] In an optional embodiment,

[0032] Based on the energy consumption prediction value and the lighting equipment characteristic information in the scene knowledge graph, the green smart lighting control strategy is generated by using the intelligent decision-making algorithm, including:

[0033] Extract lighting equipment features and relationship information from the scene knowledge graph to form a decision support knowledge base, and build an initial decision model for multi-objective optimization based on the decision support knowledge base and the obtained energy consumption forecast value. The objective functions include minimizing energy consumption, maximizing user comfort, and maximizing equipment life.

[0034] Design an intelligent optimization algorithm based on reinforcement learning, take historical energy consumption data, real-time environmental parameters and user demand data as state input, and lighting control parameters as action output, solve the initial decision model, generate an initial control strategy, and develop a simulation platform for the virtual model of the lighting system. Input the initial control strategy into the simulation platform, perform the first round of simulation, and generate the first round of simulation result data set;

[0035] A multi-dimensional evaluation index system is designed, including energy efficiency, user satisfaction, equipment health status and system response time, and a comprehensive evaluation is conducted on the first round of simulation result data set to obtain the first round of evaluation results. Based on the first round of evaluation results, the initial decision model is optimized using an adaptive parameter adjustment algorithm to generate an optimized decision model, which is applied to the simulation platform for multiple rounds of iterative simulation. After each round of simulation, a strategy evaluation is conducted based on the evaluation indicators and the optimized decision model is adjusted accordingly until the preset termination conditions are met, thereby obtaining the final decision model and the corresponding green smart lighting control strategy.

[0036] In an optional embodiment,

[0037] The calculation formula of the objective function is as follows:

[0038] ;

[0039] Among them, F(x) represents the objective function, β represents the temperature parameter, and r i (t) represents the reward value of the i-th goal in the t-th decision step, r j (t) represents the reward value of the jth goal in the tth decision step, f i (x) represents the i-th optimization objective, where f1(x) represents the energy consumption minimization objective, f2(x) represents the user comfort maximization objective, and f3(x) represents the equipment life maximization objective. represents the ideal point of the i-th target, represents the negative ideal point of the i-th objective.

[0040] According to a second aspect of the embodiments of the present invention,

[0041] Provided is a green smart lighting energy consumption prediction system based on cultural tourism, including:

[0042] The first unit is used to obtain historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and use data cleaning algorithm, interpolation algorithm and anomaly detection algorithm to preprocess the data to obtain preprocessed energy consumption data set and influencing factor data set. Based on the preprocessed historical lighting energy consumption data, the time series decomposition algorithm is used to extract features and generate energy consumption time series feature set;

[0043] The second unit is used to calculate the weights of the influencing factors of different time scales and spatial scales using a multi-layer attention network based on the preprocessed influencing factor data set, obtain a weighted influencing factor feature set, input the energy consumption time series feature set and the weighted influencing factor feature set into a multimodal feature fusion network to generate a unified feature representation, and based on the unified feature representation, use a variational inference algorithm to construct a deep probabilistic graph model, combine the pre-trained lighting energy consumption model with the deep probabilistic graph model, and obtain a prediction model adapted to the current scene;

[0044] The third unit is used to use the graph neural network algorithm, take the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene as input, build a scene knowledge graph, take the prediction model and scene knowledge graph adapted to the current scene as input, generate energy consumption prediction values ​​through a multi-scale prediction algorithm, based on the energy consumption prediction values, combined with the lighting equipment characteristic information in the scene knowledge graph, use the intelligent decision-making algorithm to generate a green smart lighting control strategy, and apply the generated green smart lighting control strategy to the lighting system of the cultural and tourism scene to realize the intelligent control of lighting equipment.

[0045] According to a third aspect of the embodiments of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the embodiments of the present invention is:

[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0052] In this embodiment, by introducing multi-source heterogeneous data and using algorithms such as time series decomposition and multi-layer attention networks for feature extraction and fusion, the spatiotemporal correlation and the mechanism of influencing factors contained in the data can be fully explored to improve the accuracy of energy consumption prediction. By using a pre-trained large-scale lighting energy consumption model and combining it with a deep probability graph model built for the current scene, knowledge can be transferred between different scenes to improve the generalization and robustness of the model. Through a multi-scale prediction algorithm, short-term, medium-term and long-term energy consumption prediction results can be generated to meet decision-making needs at different time scales, such as short-term predictions for real-time scheduling and long-term predictions for energy planning. Based on the energy consumption prediction results and the scene knowledge graph, the optimal lighting control strategy can be generated using an intelligent decision-making algorithm to minimize energy consumption while ensuring lighting quality, thereby achieving green smart lighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a method for predicting energy consumption of green smart lighting for cultural tourism according to an embodiment of the present invention;

[0054] Figure 2 It is a structural schematic diagram of the cultural tourism green smart lighting energy consumption prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0057] Figure 1 FIG. 1 is a flow chart of a method for predicting energy consumption of green smart lighting for cultural tourism according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] S101. Obtain historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and use data cleaning algorithm, interpolation algorithm and anomaly detection algorithm to preprocess the data to obtain preprocessed energy consumption data set and influencing factor data set. Based on the preprocessed historical lighting energy consumption data, use time series decomposition algorithm to extract features and generate energy consumption time series feature set.

[0059] Among them, historical lighting energy consumption data includes lighting equipment power consumption data at different time granularities, real-time energy consumption data is lighting equipment power consumption data at the current moment, and influencing factor data includes meteorological data, tourist number data, holiday data, etc. The data cleaning algorithm is used to clean the original data to remove outliers and missing values. For missing values, interpolation algorithms are used to fill them. Common interpolation algorithms include linear interpolation and spline interpolation. Outlier detection can use statistical methods, such as the 3σ principle, box plot method, etc. After data preprocessing, the preprocessed energy consumption data set and influencing factor data set are obtained.

[0060] Next, based on the preprocessed historical lighting energy consumption data, the time series decomposition algorithm is used to extract features and generate energy consumption time series feature sets. The time series decomposition algorithm can decompose the original time series data into trend items, seasonal items, and random items. The trend item reflects the long-term trend of the data, the seasonal item reflects the periodic change law of the data, and the random item is the residual item after removing the trend item and the seasonal item. Common time series decomposition algorithms include STL decomposition, HP filtering, etc.

[0061] In this embodiment, by introducing multi-source heterogeneous data and using algorithms such as time series decomposition and multi-layer attention network for feature extraction and fusion, the spatiotemporal correlation and the mechanism of influencing factors contained in the data can be fully explored to improve the accuracy of energy consumption prediction. In addition, by analyzing the influencing factors, managers can better predict and allocate resources and improve overall energy efficiency. These effects can not only promote the sustainable development of cultural and tourism scenes, but also provide a scientific basis for the formulation of relevant policies.

[0062] S102. According to the preprocessed influencing factor data set, a multi-layer attention network is used to calculate the weights of the influencing factors of different time scales and spatial scales to obtain a weighted influencing factor feature set. The energy consumption time series feature set and the weighted influencing factor feature set are input into the multimodal feature fusion network to generate a unified feature representation. Based on the unified feature representation, a deep probability graph model is constructed using a variational inference algorithm. The pre-trained lighting energy consumption model is combined with the deep probability graph model to obtain a prediction model adapted to the current scene.

[0063] First, the influencing factor dataset includes a variety of variables that may affect lighting energy consumption, such as weather, time, activity type, etc. These influencing factors may show different impact strengths at different time scales (such as hours, days, weeks) and spatial scales (such as different locations, regions). Therefore, the use of a multi-layer attention network can effectively capture the importance of these factors and calculate their specific contribution to energy consumption. The core of the attention mechanism is that it can dynamically adjust the weights according to the characteristics of the input data, thereby highlighting important factors and suppressing unimportant factors. The preprocessed influencing factor data is input into the multi-layer attention network. The network calculates the weights of each influencing factor through multiple levels of operations to generate a weighted influencing factor feature set. This feature set effectively reflects the impact of each factor on energy consumption under different time and space conditions.

[0064] Subsequently, the weighted influencing factor feature set is combined with the energy consumption time series feature set and input into the multimodal feature fusion network. The goal of this network is to integrate data from different sources into a unified feature representation that can simultaneously capture the dynamic characteristics of the time series and the weighted effects of spatial influencing factors. Through fusion, a more comprehensive and representative feature set can be obtained.

[0065] In actual operation, suppose that data on factors such as weather and tourist flow in a certain cultural and tourism scene are collected. Through the multi-layer attention network, it can be found that the impact of weather factors on energy consumption in different seasons is significantly higher than that of tourist flow. After weighting, these features are integrated into a unified feature representation, and then combined with the existing lighting energy consumption model, the variational inference algorithm is used to adjust the parameters of the model, and finally a new lighting energy consumption prediction model that adapts to the current scene is formed.

[0066] In an optional embodiment,

[0067] Based on the unified feature representation, the variational inference algorithm is used to build a deep probability graph model. The pre-trained lighting energy consumption model is combined with the deep probability graph model to obtain a prediction model that adapts to the current scene, including:

[0068] Based on the unified feature representation, a deep probabilistic graphical model containing latent variables and observed variables is constructed. The latent variables and context information related to the observed variables are input into a multi-layer perceptron, and the attention weight vector corresponding to each observed variable is calculated. The latent variables are weighted summed using the attention weight vector, and the observed variables are generated through nonlinear activation function processing.

[0069] The generated observed variables are taken as the mean of the Gaussian distribution, and the distribution variance is obtained by learning the neural network. The probability distribution of the observed variables is defined, and the prior distribution of the latent variables is multiplied by the conditional probability distribution of all observed variables to construct the joint distribution of the deep probability graph model. Based on the constructed joint distribution, the log-likelihood function is defined, and the standard normal distribution is selected as the basic distribution. Multiple reversible neural network layers including a normalization layer, a reversible mixing layer, and an affine coupling layer are designed. The basic distribution is transformed through the reversible neural network layer to construct a variational distribution.

[0070] Using the constructed variational distribution and joint distribution, derive the variational lower bound expression, calculate the variational lower bound, take the calculated variational lower bound as the optimization target, obtain a pre-trained large-scale lighting energy consumption model, combine the pre-trained large-scale lighting energy consumption model with the deep probability graph model, construct an initial combined prediction model, input the unified feature representation into the initial combined prediction model, and obtain the prediction result;

[0071] Based on the prediction results and actual energy consumption data, the prediction loss is calculated. The overall optimization objective function is constructed by combining the prediction loss and the calculated variational lower bound. Based on the overall optimization objective function, the stochastic gradient descent optimization algorithm is used to iteratively update the parameters of the combined prediction model. The iteration is repeated until the maximum number of iterations is met to obtain the final prediction model that adapts to the current scenario.

[0072] Exemplarily, firstly, based on the unified feature representation, a deep probabilistic graph model containing latent variables and observed variables is constructed. In this model, latent variables represent the potential pattern of energy consumption and the implicit role of influencing factors, and observed variables represent energy consumption characteristics and influencing factor characteristics. In order to capture the dynamic dependency between observed variables, an attention mechanism is introduced to calculate the attention weight vector corresponding to each observed variable. Specifically, the latent variables and contextual information related to the observed variables (such as time, spatial position, etc.) are input into a multi-layer perceptron, and the attention weight vector is calculated by nonlinear transformation and softmax function. Then, the latent variables are weighted and summed using the attention weight vector, and processed by a nonlinear activation function (such as ReLU) to generate observed variables. This generation process based on the attention mechanism can adaptively adjust the influence of latent variables on different observed variables and improve the expressiveness of the model.

[0073] When defining the probability distribution of the observed variables, the generated observed variables are taken as the mean of the Gaussian distribution, and the variance of the distribution is learned using the neural network. In this way, the conditional probability distribution of the observed variables can be expressed as a Gaussian distribution with the generated value as the mean and the learned variance. At the same time, the prior distribution of the latent variables is assumed to be a standard normal distribution. By multiplying the prior distribution of the latent variables with the conditional probability distribution of all observed variables, the joint distribution of the deep probability graph model is constructed.

[0074] In order to infer and learn the model, the variational inference method is adopted. First, based on the constructed joint distribution, the log-likelihood function, that is, the log probability of the observed variable, is defined. However, due to the existence of hidden variables, it is difficult to directly calculate the log-likelihood function. Therefore, a variational distribution is introduced to approximate the posterior distribution of the hidden variables. In order to construct a more flexible and accurate variational distribution, the standard normal distribution is selected as the base distribution, and then a variational distribution generation process containing multiple reversible neural network layers is designed. Specifically, the input is normalized using the Normalization Layer, and then the normalized input is transformed through the Invertible Mixing Layer and the Affine Coupling Layer. Through the combination of these reversible neural network layers, the base distribution is transformed into a complex variational distribution to better fit the true posterior distribution of the hidden variables.

[0075] In the process of learning the model, the variational distribution and joint distribution constructed are used to derive the expression of the variational lower bound (ELBO). The variational lower bound is a lower bound of the log-likelihood function, which consists of the KL divergence between the variational distribution and the joint distribution and the expected log-likelihood of the observed variable under the variational distribution. By maximizing the variational lower bound, the difference between the variational distribution and the true posterior distribution can be minimized and the log-likelihood of the observed variable can be maximized at the same time.

[0076] In order to further improve the prediction performance, a pre-trained large-scale lighting energy consumption model is introduced. The pre-trained model is combined with the deep probability graph model to construct an initial combined prediction model. Specifically, the unified feature representation is input into the initial combined prediction model, and the lighting energy consumption prediction result is obtained through the feature extraction of the pre-trained model and the inference process of the deep probability graph model.

[0077] During the optimization process of the model, the prediction loss (such as mean absolute error or mean square error) is calculated based on the prediction results and actual energy consumption data. At the same time, the overall optimization objective function is constructed by combining the prediction loss and the variational lower bound calculated previously. This objective function takes into account both the prediction performance and the generation ability and inference quality of the model. Finally, the stochastic gradient descent optimization algorithm is used to iteratively update the parameters of the combined prediction model based on the overall optimization objective function. The iterative process is repeated until the preset maximum number of iterations is reached or other termination conditions are met, and the final prediction model adapted to the current scenario is obtained.

[0078] In an optional embodiment,

[0079] Combining the prediction loss and the calculated variational lower bound, the calculation formula for the overall optimization objective function is as follows:

[0080] ;

[0081] Among them, L represents the overall optimization objective function, N represents the number of data samples, and Y n represents the true value of the nth observed variable, represents the predicted value of the nth observed variable, λ represents the weight parameter, E q(Z|Y) logp(Z|Y) represents the expected value of the log-likelihood, q(Z|Y) represents the posterior distribution of the latent variable Z given the observed variable Y, and D KL (·) represents the KL divergence, p(Z|Y) represents the probability of the latent variable Z given the observed variable Y, and p(Z) represents the prior distribution.

[0082] In this embodiment, by introducing the attention mechanism, the influence of latent variables on different observed variables is adaptively adjusted to improve the expressiveness and interpretability of the model. A reversible neural network layer is used to construct a variational distribution to improve the flexibility and accuracy of variational inference. A pre-trained model is introduced to improve the prediction performance and generalization ability through transfer learning. An overall optimization objective function is constructed to take into account both the prediction performance and the generation ability of the model to improve the overall performance of the model. Improve the accuracy and reliability of energy consumption prediction, enhance the adaptability of the model to different situations, and help decision makers formulate more effective energy management strategies. In addition, the deep learning characteristics of the model can also self-learn and optimize, so that when new data arrives in the future, it can adjust and adapt more quickly. This not only helps to achieve the rational use of resources, but also promotes the sustainable development of cultural and tourism scenarios.

[0083] S103. Using the graph neural network algorithm, the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene are taken as input to construct a scene knowledge graph. The prediction model and scene knowledge graph adapted to the current scene are taken as input, and energy consumption prediction values ​​are generated through a multi-scale prediction algorithm. Based on the energy consumption prediction values ​​and the lighting equipment characteristic information in the scene knowledge graph, a green smart lighting control strategy is generated using an intelligent decision-making algorithm. The generated green smart lighting control strategy is applied to the lighting system of the cultural and tourism scene to realize intelligent control of lighting equipment.

[0084] First of all, the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene are the basis for constructing the scene knowledge graph. The topological structure information reflects the spatial relationship between the various elements in the scene (such as buildings, paths, etc.), the functional zoning information describes the functional use of each area in the scene (such as leisure area, exhibition area, etc.), and the lighting equipment characteristic information contains the technical parameters of each lighting equipment (such as power, light intensity, energy efficiency, etc.). This information is organized in the form of a graph to form a multidimensional knowledge graph, which provides a structured basis for subsequent data analysis and prediction.

[0085] Next, the scene knowledge graph is used as input, combined with the energy consumption prediction model adapted to the current scene, and analyzed through the graph neural network. Graph neural network is a deep learning model that specializes in processing graph structure data and can effectively capture the complex relationships between nodes (such as lighting equipment). Through this process, the model can learn the correlation features between different nodes from the graph, thereby improving the accuracy of energy consumption prediction.

[0086] After obtaining the predicted energy consumption values, the intelligent decision-making algorithm is used to generate a green smart lighting control strategy based on these values ​​and the lighting equipment characteristics information in the scene knowledge graph. The intelligent decision-making algorithm aims to formulate a reasonable lighting control plan based on the predicted energy consumption values ​​and the characteristics of existing equipment. This may include measures such as adjusting the brightness, switch status or timing control of the equipment to achieve optimal energy utilization and reduce carbon emissions.

[0087] , suppose that in a certain cultural tourism scenario, the topology shows the flow of tourists in different areas, and the functional zoning information shows the high usage rate of a certain area during a specific period of time. Through the graph neural network, it may be found that the lighting demand in the area fluctuates with the change of tourist flow, and then a prediction value is generated. If the prediction shows that the energy consumption of the area will increase significantly between 8 and 9 pm, the intelligent decision-making algorithm may develop a strategy to enhance lighting during this period, and gradually reduce the brightness after the number of tourists decreases, thereby achieving the purpose of energy saving.

[0088] In an optional embodiment,

[0089] Using the graph neural network algorithm, the topological structure information, functional zoning information and lighting equipment characteristics of the cultural and tourism scene are taken as input to build a scene knowledge graph. The prediction model and scene knowledge graph adapted to the current scene are taken as input, and the energy consumption prediction value is generated through the multi-scale prediction algorithm, including:

[0090] The topological structure information, functional zoning information and lighting equipment characteristic information of the cultural tourism scene are obtained as input data. The graph neural network algorithm is used, and the attention mechanism is introduced. By learning the importance weights of different neighbor nodes, the feature information of neighbor nodes is adaptively aggregated. At the same time, convolution operations are introduced in each layer of the graph neural network. Through the feature aggregation and transformation of the local neighborhood, the high-order feature representation of the node is extracted, and the scene knowledge graph containing the scene structured representation and semantic information is generated.

[0091] Obtain historical energy consumption data and influencing factor data of cultural and tourism scenarios, combine the scenario knowledge graph to input a prediction model that adapts to the current scenario, and use a multi-scale prediction algorithm to decompose the output results of the prediction model in the time dimension to obtain short-term, medium-term, and long-term prediction subsequences respectively;

[0092] Combined with the structured representation and semantic information in the scene knowledge graph, the sequence prediction model based on the attention mechanism is applied to the short-term prediction subsequence to generate the short-term lighting energy consumption prediction results. The functional partitioning information in the scene knowledge graph is integrated to apply the prediction model based on time series decomposition and periodicity detection to the medium-term prediction subsequence to generate the medium-term lighting energy consumption prediction results. The topological structure information in the scene knowledge graph is integrated to apply the prediction model based on causal inference and scenario analysis to the long-term prediction subsequence to generate the long-term lighting energy consumption prediction results.

[0093] An adaptive multi-scale fusion mechanism is introduced to learn the importance weights of short-term, medium-term and long-term lighting energy consumption prediction results, and adaptively fuse the prediction results of different scales to obtain the final energy consumption prediction value.

[0094] For example, the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural tourism scene are first obtained as input data. The topological structure information describes the connection relationship between various areas and devices in the scene, the functional zoning information reflects the functional attributes and uses of different areas in the scene, and the lighting equipment characteristic information includes the type, power, light source parameters, etc. of the lamp.

[0095] The input data is modeled using a graph neural network algorithm, and the scene is abstracted into a graph structure, where nodes represent areas or devices in the scene, and edges represent the associations between them. The attention mechanism is introduced in the message passing and aggregation operations of the graph neural network, and the feature information of neighboring nodes is adaptively aggregated by learning the importance weights of different neighboring nodes. At the same time, convolution operations are introduced in each layer of the graph neural network to extract high-order feature representations of nodes through feature aggregation and transformation of local neighborhoods. After learning and reasoning of multi-layer graph neural networks, a scene knowledge graph containing scene structured representations and semantic information is generated.

[0096] Obtain historical energy consumption data and influencing factor data of cultural and tourism scenes, such as weather, temperature, humidity, and traffic, and combine them with scene knowledge graphs to input prediction models that adapt to the current scene. Decompose the output results of the prediction model in the time dimension through a multi-scale prediction algorithm to obtain short-term, medium-term, and long-term prediction subsequences. Short-term prediction focuses on energy consumption changes in the next few hours to a day, medium-term prediction focuses on energy consumption changes in the next few days to weeks, and long-term prediction focuses on energy consumption changes in the next few months to a year.

[0097] Combining the structured representation and semantic information in the scene knowledge graph, a sequence prediction model based on the attention mechanism is applied to the short-term prediction subsequence. The model dynamically focuses on the key time steps in the historical energy consumption data through the attention mechanism, captures the characteristics and patterns of short-term energy consumption changes, and generates short-term lighting energy consumption prediction results. Integrating the functional partition information in the scene knowledge graph, a prediction model based on time series decomposition and periodicity detection is applied to the medium-term prediction subsequence. The model decomposes the medium-term energy consumption data into trend terms, periodic terms, and residual terms, analyzes the periodic laws of energy consumption in different functional areas, and combines scene knowledge for prediction to generate medium-term lighting energy consumption prediction results. Integrating the topological structure information in the scene knowledge graph, a prediction model based on causal inference and scenario analysis is applied to the long-term prediction subsequence. The model uses the topological structure and association relationship of the scene to infer the causal dependency between different areas and devices, and combines scenario analysis and hypothesis verification to predict the trend of long-term lighting energy consumption changes.

[0098] An adaptive multi-scale fusion mechanism is introduced to adaptively fuse the prediction results of different scales by learning the importance weights of short-term, medium-term and long-term lighting energy consumption prediction results. The attention mechanism is used to dynamically adjust the weights of the prediction results of each time scale. The calculation of the weights takes into account factors such as the confidence of the prediction results, historical performance and scene characteristics, and finally the fused energy consumption prediction value is obtained.

[0099] In this embodiment, by integrating graph neural networks and knowledge graphs, the structured information and semantic information of the scene are fully utilized to improve the accuracy and interpretability of energy consumption prediction. A multi-scale prediction method is adopted to design adaptive prediction models for the characteristics of short-term, medium-term and long-term energy consumption changes, thereby improving the time scale coverage and adaptability of the prediction. The attention mechanism and adaptive fusion strategy are introduced to dynamically focus on key information and prediction results of different scales, thereby improving the adaptability and robustness of the prediction model. Combined with causal inference and scenario analysis, the impact of internal factors and external environmental factors on energy consumption is considered to enhance the interpretability and decision support capabilities of the prediction results.

[0100] In an optional embodiment,

[0101] Applying a prediction model based on causal inference and scenario analysis to the long-term prediction subsequence, the long-term lighting energy consumption prediction results include:

[0102] Obtain historical lighting energy consumption data, scenario factor data, and long-term forecast subsequences for cultural and tourism scenes, where the scenario factor data include weather conditions, time characteristics, and implementation of energy-saving measures;

[0103] Fill missing values ​​and process outliers on the historical lighting energy consumption data, and standardize the processed historical lighting energy consumption data and scenario factor data to obtain preprocessed input data, and input the preprocessed input data into a pre-constructed generative adversarial network model, wherein the generator model adopts a long short-term memory network structure, and the discriminator model adopts a convolutional neural network structure, and the generator and the discriminator are optimized by alternating training to obtain a trained generative adversarial network model;

[0104] Based on the generator of the trained adversarial generative network model, a causal graph model is constructed, situational factors are used as nodes, and the causal relationship between nodes is determined by a structural learning algorithm. The causal graph model is used to apply a structural equation model to quantify the causal strength between nodes to obtain a causal relationship quantification result, and the causal relationship quantification result is integrated into the generator model to finally obtain a causally enhanced long-term lighting energy consumption prediction model;

[0105] Based on historical data distribution and expert knowledge, multiple groups of scenario factor combinations are generated and organized into a scenario matrix, where each row represents a future scenario and each column corresponds to a scenario factor. The long-term prediction subsequence and scenario matrix are input into the causally enhanced long-term lighting energy consumption prediction model to generate long-term lighting energy consumption prediction results under multiple scenarios.

[0106] For example, the historical lighting energy consumption data is first preprocessed, including missing value filling, outlier processing and data standardization, to obtain preprocessed input data. Then, the preprocessed input data is input into a pre-built adversarial generative network model for training, where the generator model adopts a long short-term memory network structure and the discriminator model adopts a convolutional neural network structure. The generator and the discriminator are optimized by alternating training, and finally a trained adversarial generative network model is obtained.

[0107] On this basis, the generator of the trained adversarial generative network model is further used to construct a causal graph model, taking the situational factors as nodes and determining the causal relationship between nodes through the structural learning algorithm. The structural learning algorithm infers the causal relationship between nodes by analyzing the conditional independence and dependence between nodes and generates a causal graph. Common structural learning algorithms include constraint-based algorithms, score-based algorithms, and Bayesian-based algorithms. The constraint-based PC algorithm is used for causal structure learning. The algorithm gradually deletes the edges in the causal graph through the conditional independence test, and finally obtains a causal graph that satisfies all conditional independence constraints.

[0108] After obtaining the causal graph, the structural equation model is used to quantify the causal strength between nodes. The structural equation model is a mathematical model used to describe the causal relationship between variables. It represents each node as a linear combination of its parent nodes and introduces a noise term to represent the influence of unobserved factors. By estimating the coefficients in the structural equation model, the causal strength between nodes can be quantified. The maximum likelihood estimation method is used to estimate the parameters of the structural equation model to obtain the quantitative results of the causal relationship.

[0109] Next, the causal relationship quantification results are integrated into the generator model to obtain a causally enhanced long-term lighting energy consumption prediction model. Specifically, the causal relationship quantification results are introduced as prior knowledge into the loss function of the generator model, and the generator model is constrained in the form of a regularization term so that the generated data conforms to the causal relationship constraints. In this way, the generator model can not only learn the distribution characteristics of historical data, but also consider the causal relationship between situational factors, thereby improving the accuracy and interpretability of the prediction.

[0110] Finally, in order to generate long-term lighting energy consumption forecast results under multiple scenarios, multiple sets of scenario factor combinations are generated based on historical data distribution and expert knowledge, and organized into a scenario matrix, with each row representing a future scenario and each column corresponding to a scenario factor. Then, the long-term forecast subsequence and scenario matrix are input into the causal enhanced long-term lighting energy consumption forecast model to generate long-term lighting energy consumption forecast results under the corresponding scenarios. In this way, forecast results under multiple scenarios can be obtained, providing a more comprehensive reference for decision-making.

[0111] Suppose there is a set of historical lighting energy consumption data, including daily lighting energy consumption values ​​and corresponding situational factors, such as weather, temperature, holidays, etc. First, preprocess this set of data, fill in missing values, remove outliers, and perform standardization to obtain preprocessed input data. Then, input the preprocessed input data into the pre-built adversarial generative network model for training. Through the alternating optimization of the generator and the discriminator, a trained adversarial generative network model is obtained.

[0112] Next, the trained generator model is used to construct a causal graph model. The situational factors such as weather, temperature, and holidays are used as nodes, and the PC algorithm is used to learn the causal structure to obtain a causal relationship graph between situational factors. Then, the structural equation model is used to quantify the causal relationship and estimate the causal strength between each node and its parent node. The causal relationship quantification results are integrated into the generator model to obtain a causally enhanced long-term lighting energy consumption prediction model.

[0113] Finally, based on the historical data distribution and expert knowledge, multiple combinations of scenario factors are generated, such as different weather conditions, temperature ranges, holiday arrangements, etc., and organized into a scenario matrix. The long-term prediction subsequence (such as the time series for the next month) and the scenario matrix are input into the causal enhancement long-term lighting energy consumption prediction model to generate the long-term lighting energy consumption prediction results under the corresponding scenarios. In this way, the lighting energy consumption prediction values ​​under different scenarios can be obtained, providing a basis for the optimization control of the lighting system.

[0114] In this embodiment, by preprocessing the historical lighting energy consumption data and training it with the adversarial generative network model, the distribution characteristics of the historical data can be effectively learned, the expressive power and generalization ability of the prediction model can be improved, and the accuracy of the long-term lighting energy consumption prediction can be improved. By introducing the causal graph model and the structural equation model, the causal relationship between the scenario factors can be explicitly modeled, and the causal strength can be quantified, so that the prediction model can not only give the prediction results, but also explain the reasons behind the prediction results, enhance the interpretability of the prediction results, and improve the credibility of the model. By generating multiple groups of scenario factor combinations and inputting them into the causal enhancement prediction model, the long-term lighting energy consumption prediction results under different scenarios can be obtained, providing decision makers with more comprehensive reference information, supporting prediction analysis under multiple scenarios, and improving the scientificity and accuracy of decision-making. By adjusting the length of the long-term prediction subsequence, the time granularity of the prediction can be flexibly controlled, and both long-term trend predictions and short-term fine predictions can be generated to meet different decision-making needs. By integrating the causal enhanced long-term lighting energy consumption prediction model into the lighting optimization control system, a prediction basis can be provided for the optimal scheduling of the lighting system. By combining real-time monitoring data and prediction results, the operation strategy of the lighting system can be dynamically optimized, energy utilization efficiency can be improved, and smart lighting can be realized.

[0115] In an optional embodiment,

[0116] Based on the energy consumption prediction value and the lighting equipment characteristic information in the scene knowledge graph, the green smart lighting control strategy is generated by using the intelligent decision-making algorithm, including:

[0117] Extract lighting equipment features and relationship information from the scene knowledge graph to form a decision support knowledge base, and build an initial decision model for multi-objective optimization based on the decision support knowledge base and the obtained energy consumption forecast value. The objective functions include minimizing energy consumption, maximizing user comfort, and maximizing equipment life.

[0118] Design an intelligent optimization algorithm based on reinforcement learning, take historical energy consumption data, real-time environmental parameters and user demand data as state input, and lighting control parameters as action output, solve the initial decision model, generate an initial control strategy, and develop a simulation platform for the virtual model of the lighting system. Input the initial control strategy into the simulation platform, perform the first round of simulation, and generate the first round of simulation result data set;

[0119] A multi-dimensional evaluation index system is designed, including energy efficiency, user satisfaction, equipment health status and system response time, and a comprehensive evaluation is conducted on the first round of simulation result data set to obtain the first round of evaluation results. Based on the first round of evaluation results, the initial decision model is optimized using an adaptive parameter adjustment algorithm to generate an optimized decision model, which is applied to the simulation platform for multiple rounds of iterative simulation. After each round of simulation, a strategy evaluation is conducted based on the evaluation indicators and the optimized decision model is adjusted accordingly until the preset termination conditions are met, thereby obtaining the final decision model and the corresponding green smart lighting control strategy.

[0120] Exemplarily, it is necessary to first establish a sound decision support knowledge base. Using knowledge graph technology, a semantic association network of scene lighting equipment is constructed to form a structured, machine-understandable knowledge representation. Specifically, through manual annotation, automatic extraction and other methods, the attribute characteristics of lighting equipment, such as model, power, luminous flux, color temperature, color rendering index, etc., as well as the topological connection between devices, control logic and other relationship information are obtained from heterogeneous data sources such as equipment parameters, spatial layout, and instructions for use. Using the ontology modeling method, the core concepts, attributes and relationships in the field of lighting are defined, and the ontology model of the scene knowledge graph is constructed. Then, the extracted equipment characteristics and relationship information are mapped to the ontology model to form an instantiated knowledge graph.

[0121] Based on the constructed scene knowledge graph, the implicit device association and optimization potential can be further explored and inferred. For example, through graph reasoning technology, the energy consumption characteristics and usage patterns of different types of lighting equipment in different regions can be analyzed to find optimizable dimming combinations and control strategies; through graph representation learning technology, the low-dimensional semantic representation of device nodes can be learned to characterize the similarities and complementarities between devices, providing semantically enhanced state features for subsequent intelligent optimization.

[0122] Finally, the scene knowledge graph formed by integrating multi-dimensional information such as energy consumption statistics, equipment characteristics, and associated semantics constitutes the decision support knowledge foundation for the generation of smart lighting control strategies. Based on this knowledge foundation, the environmental perception ability, user demand understanding ability, and global coordination ability of lighting optimization decisions can be significantly improved.

[0123] On this basis, an initial lighting control decision model for multiple objectives was further constructed. Specifically, the objective function of the decision model includes three aspects: minimizing energy consumption, maximizing user comfort, and maximizing equipment life. Among them, minimizing energy consumption aims to minimize the power consumption of the lighting system and achieve green energy-saving operation; maximizing user comfort aims to create a light environment that meets user lighting needs and is visually comfortable, thereby improving user experience; maximizing equipment life aims to extend the healthy use cycle of lighting equipment and reduce maintenance and replacement costs.

[0124] The key to constructing the initial decision model is to reasonably set the weight coefficients of each optimization target based on the energy consumption prediction results and scene knowledge, and balance the priority and dependency between different targets. Specifically, the power, luminous flux and other parameters of lighting equipment are extracted from the scene knowledge graph, and the spatial illumination distribution and energy consumption level are estimated based on the physical model; the comfort of different lighting strategies is measured by combining user preferences, visual health standards and other knowledge; the health status of lighting equipment under different control modes is estimated based on the equipment life curve, maintenance history and other knowledge. On this basis, the weighted summation, hierarchical analysis and other methods are used to construct a multi-objective function, and the weight coefficients of each target are dynamically adjusted according to the energy consumption prediction trend, scene semantics and other factors to generate an initial decision model with scene adaptability. The target weights in the decision model can be flexibly configured according to the focus of the actual scene. For example, for industrial workshops that focus on energy saving, energy consumption targets can be given a higher weight; for elderly activity centers that focus on comfort, the priority of comfort targets can be appropriately increased; for museum exhibition halls that require frequent dimming and are sensitive to life, the impact of equipment life factors can be strengthened. Through knowledge-driven personalized decision modeling, the initial decision model can take into account both the general optimization framework and scenario-specific needs, forming the basis for a regulatory strategy tailored to local conditions.

[0125] In order to further improve the adaptability of smart lighting systems to complex dynamic environments, an intelligent optimization algorithm based on reinforcement learning is designed to achieve autonomous learning and dynamic evolution of lighting control strategies. Specifically, the smart lighting control problem is modeled as a Markov decision process, and the optimal lighting control strategy is learned through the continuous interaction between the intelligent agent and the environment. The state space contains historical energy consumption data, real-time environmental parameters (such as illuminance, personnel activities, etc.) and user demand data (such as lighting preferences, visual health, etc.), reflecting the complex scenes in which the lighting system is located. The action space is the lighting control parameters, such as lamp switching, brightness adjustment, color temperature adjustment, etc. The reward function is based on multi-dimensional indicators such as energy consumption, comfort, and life span, guiding the intelligent agent to learn a control strategy that takes into account multiple objectives. On this basis, deep reinforcement learning algorithms such as Deep Q-Network (DQN) and Deep Deterministic Policy Gradient (DDPG) are used to directly learn the mapping relationship from state to action through end-to-end neural network approximation to generate continuous and fine-grained lighting control strategies.

[0126] In order to accelerate the convergence and robustness of policy learning, a knowledge-based enhancement mechanism is introduced in reinforcement learning optimization. On the one hand, the scene knowledge graph is embedded in the state representation to provide the agent with global semantic information, accelerate environmental understanding and policy generalization; on the other hand, the initial decision model is integrated into the exploration strategy to guide the agent to search in the potential policy space and avoid blind attempts. In addition, a model-based simulation platform is designed to simulate the operation of the lighting system in a virtual environment, and perform policy iteration optimization safely and efficiently offline to reduce the cost risk of real deployment.

[0127] After multiple rounds of reinforcement learning iterations, the intelligent agent continuously optimizes the control strategy from environmental feedback, and finally obtains the optimal lighting control strategy that meets multiple objectives. It is worth mentioning that the learned strategy can be adaptively adjusted according to real-time scene changes without human intervention. For example, when lighting conditions change, the intelligent agent can autonomously adjust the brightness of lamps to maintain constant illumination; when personnel activities decrease, the intelligent agent can automatically lower the lighting power to save energy; when a certain area has long-term low illumination, the intelligent agent can appropriately increase the lighting in that area to protect visual health. Through knowledge-guided continuous learning optimization, the smart lighting system can respond to environmental changes in real time and intelligently, and always maintain the optimal state of energy saving, comfort, and health.

[0128] In order to objectively evaluate the performance of smart lighting control strategies, a comprehensive and fine-grained indicator evaluation system needs to be constructed. A set of indicators covering four dimensions, namely energy efficiency, comfort, health, and intelligence, is designed to form a quantitative basis for strategy evaluation and continuous optimization. In the energy efficiency dimension, indicators such as lighting power density (LPD) and power efficiency (LPW) are introduced to reflect the lighting energy consumption level per unit area and the luminous flux output per unit energy consumption, and to evaluate the energy-saving performance of the lighting system. In the comfort dimension, indicators such as illuminance uniformity, glare index (UGR), and color tolerance (CRI) are introduced to measure the spatial uniformity of light distribution, the visual comfort of the light source, and the fidelity of light color restoration, and to evaluate the comfort of lighting quality. In the health dimension, indicators such as flicker index (FI), blue light hazard (BHR), and circadian rhythm (CLA) are introduced to measure the flicker safety of the light source, the impact of blue light on vision, and the interference of lighting on the biological clock, and to evaluate the health friendliness of the lighting solution. The intelligent dimension introduces indicators such as response delay, dimming accuracy, and optimization effect to examine the system's real-time response capability to changes in demand, the accuracy of executing control instructions, and the level of balanced optimization of multiple objectives, thereby evaluating the degree of intelligence of lighting control.

[0129] After obtaining the results of the strategy evaluation, it is necessary to further quantify the gap between the performance of each indicator and the optimization target, and generate an explainable and executable strategy improvement plan. For example, if the energy consumption does not meet expectations, it is possible to diagnose energy consumption hotspots, locate the low-energy-efficiency time and space intervals and control parameters, and guide the optimization algorithm to improve the energy saving efforts in the problem intervals; if the glare problem is prominent, the glare source can be analyzed, the high-brightness light spots can be located, and the optimization algorithm can be guided to reduce the lighting intensity of the problem light spots; if the biological clock interference exceeds the standard, the circadian rhythm curve can be traced, the time period of day and night deviation can be located, and the optimization algorithm can be guided to increase the color temperature adjustment amplitude in the problem period. Through fine-grained strategy evaluation and feedback, a closed-loop optimization guidance mechanism is formed, which can significantly improve the pertinence and effectiveness of the optimization algorithm.

[0130] After the simulation evaluation of the initial strategy, the initial decision model based on fixed parameters and weights can be found. In order to further improve the adaptability and generalization of lighting optimization, an adaptive parameter adjustment mechanism is designed to realize the online optimization and dynamic update of the decision model. Adaptive parameter adjustment driven by scene knowledge is the key to optimizing the decision model. On the one hand, through graph reasoning and semantic parsing, the changes in the scene environment, such as personnel flow, task switching, weather changes, etc., are perceived in real time, and the weight distribution of the optimization target is dynamically adjusted accordingly. For example, when the number of people in the conference room increases suddenly, the weight of the illumination uniformity target is increased to strengthen the optimization of local lighting; when sudden severe weather causes a sudden drop in outdoor light, the weight of the energy consumption target is increased to strengthen energy saving under the premise of ensuring lighting quality. On the other hand, through data mining and user feedback analysis, the dynamic changes of user needs are understood in real time, and the target vector of the optimization model is dynamically adjusted accordingly. For example, for elderly users with declining eyesight, the target values ​​of illumination level and color rendering index can be appropriately increased; for office workers who frequently work overtime at night, the minimum color temperature of night lighting can be appropriately increased to reduce interference with the biological clock.

[0131] After obtaining the adjustment plan for optimization parameters and objectives, it is necessary to further refine it into executable decision model optimization steps. It mainly includes three aspects: First, based on parameter sensitivity analysis, evaluate the contribution of different parameter adjustments to the optimization objectives, and optimize key parameters accordingly; second, based on optimization algorithms such as gradient descent, search for parameter combinations that meet new optimization objectives in the decision space, and evaluate and screen candidate solutions with the best optimization effect; third, based on online learning algorithms, continuously fit the relationship model between optimization parameters and optimization effects, and use newly collected data streams to implement incremental updates to the relationship model, continuously improving the real-time and accuracy of parameter optimization.

[0132] By integrating knowledge-based parameter tuning, gradient-based target optimization, and learning-based model updating, the adaptive parameter tuning mechanism can give the initial decision model the ability to continuously evolve, always keeping close contact with dynamic scenarios and real-time needs. The optimized decision model can be returned to the virtual simulation platform for closed testing and withstand the test of multiple scenarios and multiple needs. When the optimized model withstands comprehensive verification, it can be deployed in a real physical space to bring users lighting smart services that are tailored to local conditions and needs.

[0133] In an optional embodiment,

[0134] The calculation formula of the objective function is as follows:

[0135] ;

[0136] Among them, F(x) represents the objective function, β represents the temperature parameter, and r i(t) represents the reward value of the i-th goal in the t-th decision step, r j (t) represents the reward value of the jth goal in the tth decision step, f i (x) represents the i-th optimization objective, where f1(x) represents the energy consumption minimization objective, f2(x) represents the user comfort maximization objective, and f3(x) represents the equipment life maximization objective. represents the ideal point of the i-th target, represents the negative ideal point of the i-th objective.

[0137] In this embodiment, by preprocessing the historical lighting energy consumption data and training it with the adversarial generative network model, the distribution characteristics of the historical data can be effectively learned, the expressive power and generalization ability of the prediction model can be improved, and the accuracy of the long-term lighting energy consumption prediction can be improved. By introducing the causal graph model and the structural equation model, the causal relationship between the scenario factors can be explicitly modeled, and the causal strength can be quantified, so that the prediction model can not only give the prediction results, but also explain the reasons behind the prediction results, enhance the interpretability of the prediction results, and improve the credibility of the model. By generating multiple groups of scenario factor combinations and inputting them into the causal enhancement prediction model, the long-term lighting energy consumption prediction results under different scenarios can be obtained, providing decision makers with more comprehensive reference information, supporting prediction analysis under multiple scenarios, and improving the scientificity and accuracy of decision-making. By adjusting the length of the long-term prediction subsequence, the time granularity of the prediction can be flexibly controlled, and both long-term trend predictions and short-term fine predictions can be generated to meet different decision-making needs. By integrating the causal enhanced long-term lighting energy consumption prediction model into the lighting optimization control system, a prediction basis can be provided for the optimal scheduling of the lighting system. By combining real-time monitoring data and prediction results, the operation strategy of the lighting system can be dynamically optimized, energy utilization efficiency can be improved, and smart lighting can be realized.

[0138] Figure 2 FIG. 1 is a schematic diagram of a structure of a system for predicting energy consumption of green smart lighting for cultural tourism according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0139] The first unit is used to obtain historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and use data cleaning algorithm, interpolation algorithm and anomaly detection algorithm to preprocess the data to obtain preprocessed energy consumption data set and influencing factor data set. Based on the preprocessed historical lighting energy consumption data, the time series decomposition algorithm is used to extract features and generate energy consumption time series feature set;

[0140] The second unit is used to calculate the weights of the influencing factors of different time scales and spatial scales using a multi-layer attention network based on the preprocessed influencing factor data set, obtain a weighted influencing factor feature set, input the energy consumption time series feature set and the weighted influencing factor feature set into a multimodal feature fusion network to generate a unified feature representation, and based on the unified feature representation, use a variational inference algorithm to construct a deep probabilistic graph model, combine the pre-trained lighting energy consumption model with the deep probabilistic graph model, and obtain a prediction model adapted to the current scene;

[0141] The third unit is used to use the graph neural network algorithm, take the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene as input, build a scene knowledge graph, take the prediction model and scene knowledge graph adapted to the current scene as input, generate energy consumption prediction values ​​through a multi-scale prediction algorithm, based on the energy consumption prediction values, combined with the lighting equipment characteristic information in the scene knowledge graph, use the intelligent decision-making algorithm to generate a green smart lighting control strategy, and apply the generated green smart lighting control strategy to the lighting system of the cultural and tourism scene to realize the intelligent control of lighting equipment.

[0142] According to a third aspect of the embodiments of the present invention,

[0143] An electronic device is provided, comprising:

[0144] processor;

[0145] a memory for storing processor-executable instructions;

[0146] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0147] A fourth aspect of the embodiments of the present invention is:

[0148] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0149] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting energy consumption of green intelligent lighting based on cultural tourism, characterized in that: include: Obtain historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and use data cleaning algorithms, interpolation algorithms and anomaly detection algorithms to preprocess the data to obtain preprocessed energy consumption data sets and influencing factor data sets. Based on the preprocessed historical lighting energy consumption data, use the time series decomposition algorithm to extract features and generate energy consumption time series feature sets. Among them, the influencing factor data include meteorological data, tourist number data, and holiday data; According to the preprocessed influencing factor data set, a multi-layer attention network is used to calculate the weights of the influencing factors of different time scales and spatial scales to obtain a weighted influencing factor feature set. The energy consumption time series feature set and the weighted influencing factor feature set are input into the multimodal feature fusion network to generate a unified feature representation. Based on the unified feature representation, a deep probabilistic graph model is constructed using a variational inference algorithm. The pre-trained lighting energy consumption model is combined with the deep probabilistic graph model to obtain a prediction model that adapts to the current scene. Using the graph neural network algorithm, the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene are taken as input to construct a scene knowledge graph. The prediction model and scene knowledge graph adapted to the current scene are taken as input, and energy consumption prediction values ​​are generated through a multi-scale prediction algorithm. Based on the energy consumption prediction values ​​and the lighting equipment characteristic information in the scene knowledge graph, a green smart lighting control strategy is generated using an intelligent decision-making algorithm. The generated green smart lighting control strategy is applied to the lighting system of the cultural and tourism scene to realize intelligent control of lighting equipment. Based on the unified feature representation, the variational inference algorithm is used to build a deep probability graph model. The pre-trained lighting energy consumption model is combined with the deep probability graph model to obtain a prediction model that adapts to the current scene, including: Based on the unified feature representation, a deep probabilistic graphical model including latent variables and observed variables is constructed, the latent variables and context information related to the observed variables are input into a multi-layer perceptron, the attention weight vector corresponding to each observed variable is calculated, the latent variables are weighted summed using the attention weight vector, and the observed variables are generated through nonlinear activation function processing, wherein the latent variables represent the potential pattern of energy consumption and the implicit role of influencing factors, and the observed variables represent the energy consumption characteristics and the characteristics of the influencing factors; The generated observed variables are taken as the mean of the Gaussian distribution, and the distribution variance is obtained by learning the neural network. The probability distribution of the observed variables is defined, and the prior distribution of the latent variables is multiplied by the conditional probability distribution of all observed variables to construct the joint distribution of the deep probability graph model. Based on the constructed joint distribution, the log-likelihood function is defined, and the standard normal distribution is selected as the basic distribution. Multiple reversible neural network layers including a normalization layer, a reversible mixing layer, and an affine coupling layer are designed. The basic distribution is transformed through the reversible neural network layer to construct a variational distribution. Using the constructed variational distribution and joint distribution, derive the variational lower bound expression, calculate the variational lower bound, take the calculated variational lower bound as the optimization target, obtain a pre-trained large-scale lighting energy consumption model, combine the pre-trained large-scale lighting energy consumption model with the deep probability graph model, construct an initial combined prediction model, input the unified feature representation into the initial combined prediction model, and obtain the prediction result; Based on the prediction results and actual energy consumption data, the prediction loss is calculated, and the overall optimization objective function is constructed by combining the prediction loss and the calculated variational lower bound. Based on the overall optimization objective function, the stochastic gradient descent optimization algorithm is used to iteratively update the parameters of the combined prediction model, and the iteration is repeated until the maximum number of iterations is met to obtain the final prediction model that adapts to the current scenario; Using the graph neural network algorithm, the topological structure information, functional zoning information and lighting equipment characteristics of the cultural and tourism scene are taken as input to build a scene knowledge graph. The prediction model and scene knowledge graph adapted to the current scene are taken as input, and the energy consumption prediction value is generated through the multi-scale prediction algorithm, including: The topological structure information, functional zoning information and lighting equipment characteristic information of the cultural tourism scene are obtained as input data. The graph neural network algorithm is used, and the attention mechanism is introduced. By learning the importance weights of different neighbor nodes, the feature information of neighbor nodes is adaptively aggregated. At the same time, convolution operations are introduced in each layer of the graph neural network. Through the feature aggregation and transformation of the local neighborhood, the high-order feature representation of the node is extracted, and the scene knowledge graph containing the scene structured representation and semantic information is generated. Obtain historical energy consumption data and influencing factor data of cultural and tourism scenarios, combine the scenario knowledge graph to input a prediction model that adapts to the current scenario, and use a multi-scale prediction algorithm to decompose the output results of the prediction model in the time dimension to obtain short-term, medium-term, and long-term prediction subsequences respectively; Combined with the structured representation and semantic information in the scene knowledge graph, the sequence prediction model based on the attention mechanism is applied to the short-term prediction subsequence to generate the short-term lighting energy consumption prediction results. The functional partitioning information in the scene knowledge graph is integrated to apply the prediction model based on time series decomposition and periodicity detection to the medium-term prediction subsequence to generate the medium-term lighting energy consumption prediction results. The topological structure information in the scene knowledge graph is integrated to apply the prediction model based on causal inference and scenario analysis to the long-term prediction subsequence to generate the long-term lighting energy consumption prediction results. An adaptive multi-scale fusion mechanism is introduced to learn the importance weights of short-term, medium-term and long-term lighting energy consumption prediction results, and adaptively fuse the prediction results of different scales to obtain the final energy consumption prediction value.

2. The method according to claim 1, characterized in that Combining the prediction loss and the calculated variational lower bound, the calculation formula for the overall optimization objective function is as follows: Among them, L represents the overall optimization objective function, N represents the number of data samples, and Y n represents the true value of the nth observed variable, represents the predicted value of the nth observed variable, λ represents the weight parameter, E q(Z|Y) logp(Z|Y) represents the expected value of the log-likelihood, q(Z|Y) represents the posterior distribution of the latent variable Z given the observed variable Y, and D KL (·) represents the KL divergence, p(Z|Y) represents the probability of the latent variable Z given the observed variable Y, and p(Z) represents the prior distribution.

3. The method according to claim 1, characterized in that Applying a prediction model based on causal inference and scenario analysis to the long-term prediction subsequence, the long-term lighting energy consumption prediction results include: Obtain historical lighting energy consumption data, scenario factor data, and long-term forecast subsequences for cultural and tourism scenes, where the scenario factor data include weather conditions, time characteristics, and implementation of energy-saving measures; Fill missing values ​​and process outliers on the historical lighting energy consumption data, and standardize the processed historical lighting energy consumption data and scenario factor data to obtain preprocessed input data, and input the preprocessed input data into a pre-constructed generative adversarial network model, wherein the generator model adopts a long short-term memory network structure, and the discriminator model adopts a convolutional neural network structure, and the generator and the discriminator are optimized by alternating training to obtain a trained generative adversarial network model; Based on the generator of the trained adversarial generative network model, a causal graph model is constructed, situational factors are used as nodes, and the causal relationship between nodes is determined by a structural learning algorithm. The causal graph model is used to apply a structural equation model to quantify the causal strength between nodes to obtain a causal relationship quantification result, and the causal relationship quantification result is integrated into the generator model to finally obtain a causally enhanced long-term lighting energy consumption prediction model; Based on historical data distribution and expert knowledge, multiple groups of scenario factor combinations are generated and organized into a scenario matrix, where each row represents a future scenario and each column corresponds to a scenario factor. The long-term prediction subsequence and scenario matrix are input into the causally enhanced long-term lighting energy consumption prediction model to generate long-term lighting energy consumption prediction results under multiple scenarios.

4. The method according to claim 1, characterized in that: Based on the energy consumption prediction value and the lighting equipment characteristic information in the scene knowledge graph, the green smart lighting control strategy is generated by using the intelligent decision-making algorithm, including: Extract lighting equipment features and relationship information from the scene knowledge graph to form a decision support knowledge base, and build an initial decision model for multi-objective optimization based on the decision support knowledge base and the obtained energy consumption forecast value. The objective functions include minimizing energy consumption, maximizing user comfort, and maximizing equipment life. Design an intelligent optimization algorithm based on reinforcement learning, take historical energy consumption data, real-time environmental parameters and user demand data as state input, and lighting control parameters as action output, solve the initial decision model, generate an initial control strategy, and develop a simulation platform for the virtual model of the lighting system. Input the initial control strategy into the simulation platform, perform the first round of simulation, and generate the first round of simulation result data set; A multi-dimensional evaluation index system is designed, including energy efficiency, user satisfaction, equipment health status and system response time, and a comprehensive evaluation is conducted on the first round of simulation result data set to obtain the first round of evaluation results. Based on the first round of evaluation results, the initial decision model is optimized using an adaptive parameter adjustment algorithm to generate an optimized decision model, which is applied to the simulation platform for multiple rounds of iterative simulation. After each round of simulation, a strategy evaluation is conducted based on the evaluation indicators and the optimized decision model is adjusted accordingly until the preset termination conditions are met, thereby obtaining the final decision model and the corresponding green smart lighting control strategy.

5. The method according to claim 4, characterized in that The calculation formula of the objective function is as follows: Among them, F(x) represents the objective function, β represents the temperature parameter, and r i (t) represents the reward value of the i-th goal in the t-th decision step, r j (t) represents the reward value of the jth goal in the tth decision step, f i (x) represents the i-th optimization objective, where f1(x) represents the energy consumption minimization objective, f2(x) represents the user comfort maximization objective, and f3(x) represents the equipment life maximization objective. represents the ideal point of the i-th target, represents the negative ideal point of the i-th objective.

6. A green intelligent lighting energy consumption prediction system based on cultural tourism, used to implement any of the methods in claims 1-5, characterized in that: include: The first unit is used to obtain historical lighting energy consumption data, real-time energy consumption data and influencing factor data of cultural and tourism scenes, and use data cleaning algorithm, interpolation algorithm and anomaly detection algorithm to preprocess the data to obtain preprocessed energy consumption data set and influencing factor data set. Based on the preprocessed historical lighting energy consumption data, the time series decomposition algorithm is used to extract features and generate energy consumption time series feature set; The second unit is used to calculate the weights of the influencing factors of different time scales and spatial scales using a multi-layer attention network based on the preprocessed influencing factor data set, obtain a weighted influencing factor feature set, input the energy consumption time series feature set and the weighted influencing factor feature set into a multimodal feature fusion network to generate a unified feature representation, and based on the unified feature representation, use a variational inference algorithm to construct a deep probabilistic graph model, combine the pre-trained lighting energy consumption model with the deep probabilistic graph model, and obtain a prediction model adapted to the current scene; The third unit is used to use the graph neural network algorithm, take the topological structure information, functional zoning information and lighting equipment characteristic information of the cultural and tourism scene as input, build a scene knowledge graph, take the prediction model and scene knowledge graph adapted to the current scene as input, generate energy consumption prediction values ​​through a multi-scale prediction algorithm, based on the energy consumption prediction values, combined with the lighting equipment characteristic information in the scene knowledge graph, use the intelligent decision-making algorithm to generate a green smart lighting control strategy, and apply the generated green smart lighting control strategy to the lighting system of the cultural and tourism scene to realize the intelligent control of lighting equipment.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images