Street lamp intelligent sensing and energy-saving control optimization method based on Internet of Things

Abstract

The invention discloses a street lamp intelligent perception and energy-saving control optimization method based on the Internet of Things, and the method comprises the following steps: S1, collecting environment data through an Internet of Things terminal, and generating a multi-source local feature sub-sample set; s2, constructing a deep forest prediction model based on the multi-source local feature sub-sample set; s3, optimizing the depth forest prediction model by adopting an artificial immune algorithm to generate an optimized depth forest prediction model; s4, obtaining a control strategy instruction set based on the optimized deep forest prediction model; s5, the control strategy instruction set is issued to the street lamp intelligent terminal, and a dynamic on-off action and a brightness adjusting action are executed; and S6, collecting energy consumption feedback data and new environment change data after execution of the street lamp intelligent terminal, and outputting a depth forest prediction model after adaptive optimization. According to the method, the artificial immune algorithm and the deep forest model are fused, and intelligent prediction and dynamic energy-saving control optimization of the illumination demand of the street lamp are realized based on the environment perception data of the Internet of Things.

Description

Technical Field

[0001] The present invention relates to the technical field of smart cities and Internet of Things applications, and particularly to an intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things. Background Art

[0002] With the wide application of Internet of Things (IoT) technology, the intelligent level of urban infrastructure has been continuously improved. Especially in the construction of smart cities, the street lamp system, as an important part of urban public lighting, has gradually become an important carrier for intelligent perception and energy-saving control. Traditional street lamp control systems mostly rely on simple time scheduling or manual management, and it is difficult to dynamically adjust according to environmental changes, resulting in high energy consumption, serious resource waste, and lack of adaptive adjustment ability. With the development of environmental perception technology, sensor networks have been deployed in scenarios such as urban roads, blocks, parks, etc. to collect multi-source environmental information such as light intensity, weather changes, traffic flow, and pedestrian density, providing basic data support for more refined intelligent lighting control.

[0003] Existing energy-saving control methods for street lamps based on the Internet of Things usually adopt a simple rule decision model driven by perception data. For example, a fixed threshold is set, and when the environmental light intensity is lower than a certain level, the street lamp is automatically turned on, and when it is higher than a certain level, the street lamp is turned off, or the brightness of the street lamp is simply adjusted according to the changes in the density of people and vehicles. Although these methods improve the automation level of control to a certain extent, due to the lack of the ability to predict the trend of environmental changes and the depth of understanding of multi-source data fusion, they often lead to response lags and rigid control strategies, and are unable to adapt to the rapidly changing urban environment. Especially in scenarios where the light changes violently (such as dusk, early morning, rainy days) and the traffic flow fluctuates greatly (such as holidays, peak periods), traditional control strategies are difficult to make optimized responses in a timely manner, which not only affects lighting safety but also causes unnecessary energy consumption.

[0004] Methods for predicting environmental changes have gradually been proposed. Some studies introduce machine learning models, such as support vector machine (SVM), decision tree (DT), random forest (RF), etc., to speculate on future lighting requirements based on historical perception data. However, these traditional machine learning models have two main limitations: one is that the model capacity is limited and it is difficult to process complex non-linear and multi-source heterogeneous environmental data, resulting in insufficient prediction accuracy; the other is that the model is fixed and lacks the ability to update online, making it difficult to adapt to long-term environmental evolution and urban development changes. For this reason, some studies try to use deep learning methods, such as convolutional neural network (CNN), long short-term memory network (LSTM), etc. for time series prediction, but these methods usually require a large amount of labeled data for training and consume a large amount of computing resources, and there are great difficulties in deploying them on edge devices.

[0005] In recent years, with the proposal of the gcForest model, it provides deep feature enhancement and integrated learning capabilities with a lightweight and non-neural network structure, and can achieve prediction performance comparable to that of deep neural networks in small-sample and low-computing-power environments. The deep forest enhances features layer by layer through cascading random forests and completely random forests, gradually refining data features, and is suitable for application in the scenario of Internet of Things street lamp environment perception. However, most of the currently publicized applications of deep forests focus on general tasks such as classification and regression, lacking customized designs for street lamp lighting demand prediction and energy-saving control optimization, and failing to fully combine the spatio-temporal characteristics of Internet of Things environment perception with the energy-saving target-oriented features.

[0006] In the existing technology, for the problem of model hyperparameter optimization, methods such as grid search, Bayesian optimization, or genetic algorithms are usually used to adjust model parameters. However, these methods still have significant limitations in a dynamically changing environment. Grid search and Bayesian optimization require a large amount of offline training overhead and are difficult to adapt to environmental changes in real time; although genetic algorithm optimization has a certain degree of self-adaptability, it is prone to falling into local optima in a high-dimensional complex parameter space, and the stability of the evolution process is insufficient. As an intelligent optimization method inspired by the biological immune system, the artificial immune algorithm has the advantages of strong high-dimensional search ability, good population diversity, and great potential for global optimal search. In theory, it is more suitable for the hyperparameter optimization of deep forests in dynamic and complex environments. However, there are few studies in the existing literature that combine the artificial immune algorithm with the deep forest model for the field of Internet of Things intelligent perception and energy-saving control optimization, and lack a complete optimization closed-loop mechanism.

[0007] In addition, the existing technology generally lacks an adaptive model update mechanism based on energy consumption feedback. Once the traditional predictive control method is deployed, during the long-term operation process, due to factors such as environmental changes, equipment aging, and sensor drift, the prediction accuracy and energy-saving control effect gradually decline. However, the system lacks effective feedback collection and model fine-tuning means, and can only be retrained manually regularly, increasing the maintenance cost and having a slow response. In the application of intelligent street lamp energy-saving control, how to construct a low-latency, highly self-adaptive, and continuously optimized model learning mechanism has become a key issue in improving the reliability and energy-saving efficiency of the system.

[0008] Therefore, how to provide an Internet of Things-based intelligent perception and energy-saving control optimization method for street lamps is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose an intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things. The present invention makes full use of the Internet of Things environment perception technology, deep forest prediction model and artificial immune algorithm, combines the adaptive fine-tuning mechanism driven by energy consumption feedback, and details the overall process of intelligently realizing accurate prediction of lighting requirements, dynamic street lamp control and energy-saving optimization, with the advantages of high energy-saving efficiency, strong environmental adaptability, high prediction accuracy and good system robustness.

[0010] The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect environmental data through Internet of Things terminals, construct an original environmental perception data set, and use a sliding window mechanism to perform multi-granularity feature extraction on the original environmental perception data set to generate a multi-source local feature sub-sample set;

[0012] S2. Construct a deep forest prediction model, input the multi-source local feature sub-sample set into multiple groups of random forests and completely random forests for preliminary processing, and combine the cascade deep forest structure to perform layer-by-layer feature enhancement and lighting requirement prediction;

[0013] S3. Optimize the deep forest prediction model using the artificial immune algorithm, encode the hyperparameters of the deep forest prediction model into the antibody structure, define the affinity function, perform cloning amplification and mutation on the high-affinity antibodies, and select the antibody with the highest affinity to update the hyperparameters of the deep forest prediction model to generate an optimized deep forest prediction model;

[0014] S4. Based on the lighting requirement prediction results output by the optimized deep forest prediction model, formulate street lamp switch control strategies and brightness adjustment strategies to obtain a control strategy instruction set;

[0015] S5. Send the control strategy instruction set to the street lamp intelligent terminal to perform dynamic switch actions and brightness adjustment actions;

[0016] S6. Collect the energy consumption feedback data and new environmental change data after the street lamp intelligent terminal executes, update the parameters of the deep forest prediction model through the fine-tuning mechanism, and output an adaptively optimized deep forest prediction model.

[0017] Optionally, the original environmental perception data set includes environmental light intensity data, weather condition data, traffic flow data, time information data and historical energy consumption data.

[0018] Optionally, the generation of the multi-source local feature sub-sample set includes applying a sliding window mechanism to the original environmental perception data set, setting the window size to w and the step size to s. At each sliding position k, w consecutive perception data vectors within the window are extracted to generate a local feature sub-sample. The local feature sub-samples generated at all sliding positions k are numbered to form a multi-source local feature sub-sample set.

[0019] Optionally, S2 specifically includes:

[0020] S21. Input the multi-source local feature sub-sample set into the primary random forest ensemble for training to obtain a primary random forest enhanced feature set, which is used to represent environmental features such as light, weather, and traffic in the Internet of Things perception data;

[0021] S22. Input the multi-source local feature sub-sample set into the completely random forest ensemble for training to obtain a primary completely random forest enhanced feature set. The primary completely random forest enhanced feature set supplements diverse environmental feature representations to improve the accuracy of the light demand prediction;

[0022] S23. Perform feature-level concatenation on the primary random forest enhanced feature set and the primary completely random forest enhanced feature set to form a primary enhanced feature set, which is used as the input for the subsequent light demand prediction model;

[0023] S24. Use the primary enhanced feature set as the input for the first-level cascade forest to train the first-level deep forest. The primary enhanced feature set is processed by multiple random forest classifiers and multiple completely random forest classifiers respectively to extract the class probability distribution features, and the primary enhanced feature set is concatenated with the class probability distribution features output by each forest to generate a first-level enhanced feature set, and a first-level light demand prediction distribution is generated based on the first-level enhanced feature set to predict the street lamp lighting demand for the corresponding time period and area;

[0024] S25. Use the first-level enhanced feature set as the input to iteratively construct the l-level cascade forest, and sequentially generate the l-level enhanced feature set and the corresponding l-level light demand prediction distribution:

[0025]

[0026] where Y l is the l-level light demand prediction distribution, K is the number of random forest classifiers used in the current l-th layer, M is the number of completely random forest classifiers used in the current l-th layer, k is the random forest classifier index, is the weighted coefficient assigned to the k-th random forest classifier in the comprehensive prediction, p Rk (F l-1) is the probability distribution of the lighting demand output by the k-th random forest classifier for the enhanced feature set in the previous layer, F l-1 is the enhanced feature set output by the depth forest in the (l - 1)-th layer, m is the index of the completely random forest classifier, is the weighted coefficient assigned by the m-th completely random forest classifier in the comprehensive prediction, is the probability distribution of the lighting demand output by the m-th completely random forest classifier for the enhanced feature set in the previous layer;

[0027] The predicted lighting demand distribution Y at the l-th level l is used to gradually improve the prediction accuracy of the lighting demand of the road lamp under environmental changes;

[0028] S26. Based on the predicted lighting demand distribution at the (l - 1)-th level and the predicted lighting demand distribution at the l-th level, calculate the performance improvement rate. If the performance improvement rate is less than or equal to the set threshold ∈, terminate the cascade and output the depth forest prediction model obtained by the current training. Otherwise, continue to execute the S25 step.

[0029] Optionally, the S3 specifically includes:

[0030] S31. Encode the hyperparameter set of the depth forest prediction model into the antibody population P = {a1, a2, …, a N}, where each antibody a i represents a set of hyperparameter combinations, and the hyperparameter combinations include the number of random forests, the number of completely random forests, the number of decision trees inside each forest, and the maximum depth of each decision tree;

[0031] S32. Based on the hyperparameter configuration corresponding to each antibody a i , train the depth forest prediction model, and evaluate the prediction performance on the validation set, calculate the prediction mean square error and the energy saving rate respectively, and define the affinity:

[0032] Affinity(a i ) = w1 × (1 - MSE(a i )) + w2 × ER(a i );

[0033] Among them, Affinity(a i ) is the affinity of the depth forest prediction model corresponding to the antibody a i under the comprehensive evaluation of the lighting demand prediction and the energy saving effect, a i is a set of hyperparameter configurations of the depth forest model, i is the index number of the antibody, w1 is the prediction accuracy weight coefficient, MSE(a i ) is the antibody a iConfigure the mean square error of the light demand prediction of the trained deep forest model on the validation set. Let w2 be the energy-saving effect weight coefficient, and ER(a i ) be the use of antibody a i Configure the trained deep forest model;

[0034] S33. Clone the antibody population according to the affinity. The number of clones is proportional to the affinity to generate a clone antibody set;

[0035] S34. Perform a mutation operation on each antibody in the clone antibody set. The mutation method is to apply a small perturbation to the hyperparameters to generate a mutant antibody set;

[0036] S35. Retrain the deep forest prediction model based on the mutant antibody set and calculate the affinity. Combine the original population and the mutant population, and select the N antibodies with the highest affinity from them to update to the new generation population;

[0037] S36. Iteratively execute the steps S32 to S35 until the average affinity change rate of the antibody population for consecutive G generations meets the convergence condition, and output the optimized deep forest prediction model:

[0038]

[0039] where, is the average affinity of the g-th generation of artificial immune population, is the average affinity of the (g - 1)-th generation of artificial immune population. g is the generation number of the current artificial immune evolution iteration, and δ is the convergence error threshold; the optimized deep forest prediction model is used to predict the street lamp light demand based on the Internet of Things environment perception data and formulate an energy-saving control optimization strategy based on the prediction result.

[0040] Optionally, the S34 specifically includes:

[0041] S341. Select the clone antibody set, apply the perturbation mechanism to each antibody's hyperparameter set separately, and perform small perturbations on the number of random forests, the number of completely random forests, the number of internal decision trees, and the maximum depth of the decision tree according to the set mutation rules to generate preliminary mutant antibodies;

[0042] S342. Uniformly integrate the mutation results of the preliminary mutant antibodies to form standardized mutant antibodies, ensuring that each hyperparameter still meets the preset value range and legality requirements after perturbation;

[0043] S343. Calculate the mutation amplitude of each mutant antibody. The mutation amplitude is defined as the average of the normalized change rates of the original antibody and the mutant antibody on each hyperparameter;

[0044] S344. Determine whether the variation range of the variant antibody exceeds the maximum allowable variation range threshold. If the variation range of the variant antibody is greater than the maximum allowable variation range threshold, discard the variant antibody. If the variation range of the variant antibody is less than or equal to the maximum allowable variation range threshold, retain the variant antibody to form a variant antibody set.

[0045] Optionally, the S4 specifically includes:

[0046] S41. Using the optimized deep forest prediction model as the input, based on the input environmental perception data set, generate a lighting demand prediction result Y, where the lighting demand prediction result Y = {y1, y2, …, y T}, where y t represents the predicted value of the street lamp lighting demand at time step t;

[0047] S42. Based on the lighting demand prediction result Y, set the turn-on threshold and turn-off threshold to generate the street lamp switch control instruction C t for time step t;

[0048] S43. For the street lamps in the on state, generate a brightness adjustment instruction according to the predicted lighting demand at the corresponding time step t:

[0049]

[0050] where L t is the street lamp brightness adjustment instruction at time step t, t is the time index, clip(·) is a limiting function, y t is the predicted lighting demand value output by the optimized deep forest prediction model at time step t, θ off is the preset turn-on threshold, and θ off is the preset turn-off threshold;

[0051] S44. Combine the switch control instruction C t corresponding to time step t with the brightness adjustment instruction L t to form a complete street lamp control instruction set U.

[0052] Optionally, the street lamp switch control instruction includes, at each time step t, judging according to the predicted lighting demand and the set turn-on threshold and turn-off threshold. When the predicted lighting demand is greater than or equal to the turn-on threshold, set the street lamp switch control instruction to the on state; when the predicted lighting demand is less than or equal to the turn-off threshold, set the street lamp switch control instruction to the off state.

[0053] Optionally, the S6 specifically includes:

[0054] S61. During the execution of the street lamp control instruction set U, collect the actual energy consumption data and actual environmental perception data corresponding to each time step t, and generate a feedback data set;

[0055] S62. Based on the feedback data set, calculate the energy consumption deviation set for each time step according to the theoretically calculated energy consumption data corresponding to the execution of the control instruction set:

[0056]

[0057] where ΔE t is the energy consumption deviation at time step t, t is the time index, is the actual street lamp energy consumption value collected in real time at time step t by the Internet of Things environmental perception device, P base is the fixed standby power consumption of the street lamp device in the off state or the basic on state, C t is the switch control instruction, P max is the maximum power consumption of the street lamp when working at the maximum brightness, L t is the brightness adjustment instruction;

[0058] S63. Based on the energy consumption deviation set, calculate the average energy consumption deviation, and set the energy consumption deviation correction threshold to ∈ E , if the average energy consumption deviation is greater than the energy consumption deviation correction threshold, trigger the optimization of the prediction model;

[0059] S64. In the case of triggering the optimization of the prediction model, retrain the deep forest prediction model based on the feedback data set, update the model parameters through the fine-tuning mechanism, and output the adaptively optimized deep forest prediction model.

[0060] Optionally, S64 specifically includes:

[0061] S641. In the case of triggering the optimization of the prediction model, based on the feedback data set, extract the feature subset;

[0062] S642. Input the feature subset into the currently optimized deep forest prediction model. In the currently optimized deep forest prediction model, select the sub-forest models with the contribution rate of the overall energy consumption prediction error greater than the preset threshold σ M , and form a set of sub-models to be fine-tuned by all the sub-forest models that meet the conditions;

[0063] S643. For each sub-model in the set of sub-models to be fine-tuned, perform incremental learning based on the feedback data set, adjust the original parameter set, and superimpose the parameter update amount obtained by training with the feedback data on the original parameters to generate the fine-tuned sub-model parameter set;

[0064] S644. Replace the corresponding sub-model set to be fine-tuned in the currently optimized deep forest prediction model with the fine-tuned sub-model set, keep the structures and parameters of other unadjusted sub-models unchanged, and output the adaptively optimized deep forest prediction model.

[0065] The beneficial effects of the present invention are as follows:

[0066] First, different from the traditional Internet of Things (IoT)-based street lamp control method that simply turns on and off according to the environmental light threshold, the present invention proposes an intelligent perception and energy-saving control optimization method for street lamps based on the IoT. By using a sliding window mechanism to extract local spatio-temporal feature sub-samples and combining with a deep forest model for layer-by-layer feature enhancement and light demand prediction, it can effectively model the environmental change trend in a small-sample and heterogeneous data environment, significantly improving the response sensitivity and prediction accuracy of the street lamp control system to complex urban environments. Through the dynamic generation of local feature sub-samples, the present invention realizes local perception and trend capture of environmental information in the feature extraction stage, and has stronger adaptability and generalization ability compared with traditional static feature modeling methods.

[0067] Second, in the process of optimizing the light demand prediction model, the present invention first organically combines the artificial immune algorithm with the deep forest structure, and dynamically optimizes the model hyperparameter configuration through immune cloning, affinity evaluation, and antibody mutation mechanisms. Compared with traditional grid search, random search, or genetic optimization methods, the artificial immune optimization mechanism can, while maintaining population diversity, quickly locate the optimal or near-optimal model hyperparameter combination in a way that combines local fine-tuning and global search in parallel, greatly improving the model training efficiency and final prediction performance. In particular, by comprehensively considering the light prediction error and actual energy-saving effect through the affinity function, the optimization process closely adheres to the application scenario requirements, avoiding the problem of simply pursuing prediction accuracy while ignoring the actual energy-saving benefits.

[0068] In the street lamp control decision-making stage, the present invention proposes a dynamic switching and brightness adjustment strategy based on the light demand prediction result, sets double thresholds for turning on and off the lights, and combines a brightness adjustment ratio mapping mechanism to realize the transformation from simple binary switching control to continuously adjustable lighting control. Through fine-grained control, it can dynamically adjust the street lamp brightness according to the actual environmental light demand, not only ensuring the lighting safety needs of pedestrians and vehicles, but also effectively reducing the overall energy consumption. Compared with traditional time-switching or control methods triggered based on a single light intensity, the control strategy of the present invention has higher intelligence, real-time performance, and energy-saving effect.

[0069] Meanwhile, the present invention first introduces an adaptive optimization closed-loop mechanism based on energy consumption feedback. By collecting actual energy consumption data after executing instructions in real time, comparing and analyzing it with the theoretically predicted energy consumption, dynamically calculating the energy consumption deviation, and then triggering the fine-tuning learning process of the deep forest model. The fine-tuning mechanism includes sub-forest retraining, threshold updating, feature re-enhancement, and a small amount of incremental training, which can quickly correct local prediction errors without destroying the overall structure of the original model, and continuously improve the robustness and energy-saving control ability of the system in a dynamic environment. Compared with the traditional method of periodically retraining the model offline, the present invention realizes true online adaptive optimization, greatly reducing the system maintenance cost and improving the long-term stability and reliability of the intelligent lighting system.

[0070] In addition, the present invention adopts a lightweight and modular design concept. The deep forest model has good scalability and low computing power adaptation ability, and is suitable for deployment on edge computing nodes or low-power Internet of Things devices, meeting the engineering application requirements of real-time response, low latency, and low energy consumption in the smart city scenario. Through the above technical innovation and optimization integration, the present invention can maximize energy consumption optimization while ensuring lighting safety, with significant practical application value and broad promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0072] Figure 1 is a flowchart of the intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things proposed by the present invention;

[0073] Figure 2 is a flowchart of the deep forest model for predicting lighting requirements and layer-by-layer feature enhancement in the intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things proposed by the present invention;

[0074] Figure 3 is a flowchart of the artificial immune algorithm for optimizing the hyperparameters of the deep forest in the intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0075] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0076] Refer to Figures 1-3 , the intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things includes the following steps:

[0077] S1. Collect environmental data through IoT terminals, construct an original environmental perception dataset, and use a sliding window mechanism to extract multi-granularity features from the original environmental perception dataset to generate a multi-source local feature subsample set;

[0078] S2. Construct a deep forest prediction model, input the multi-source local feature subsample set into multiple groups of random forests and completely random forests for preliminary processing, and combine a cascaded deep forest structure for layer-by-layer feature enhancement and lighting demand prediction;

[0079] S3. Optimize the deep forest prediction model using an artificial immune algorithm, encode the hyperparameters of the deep forest prediction model as antibody structures, define an affinity function, perform clone amplification and mutation on high-affinity antibodies, and select the antibody with the highest affinity to update the hyperparameters of the deep forest prediction model to generate an optimized deep forest prediction model;

[0080] S4. Based on the lighting demand prediction results output by the optimized deep forest prediction model, formulate a street lamp switch control strategy and a brightness adjustment strategy to obtain a control strategy instruction set;

[0081] S5. Send the control strategy instruction set to the street lamp intelligent terminal to execute dynamic switch actions and brightness adjustment actions;

[0082] S6. Collect the energy consumption feedback data and new environmental change data after the street lamp intelligent terminal executes, update the parameters of the deep forest prediction model through a fine-tuning mechanism, and output an adaptively optimized deep forest prediction model.

[0083] The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things provided by the present invention breaks through the limitations of traditional street lamp control relying on fixed thresholds and static strategies. By introducing sliding window local feature extraction, deep forest lighting demand prediction, and artificial immune optimization mechanisms, it realizes intelligent modeling and adaptive decision-making for dynamic control of street lamps in complex urban environments. Combined with an energy consumption feedback fine-tuning mechanism, the system can continuously optimize the model performance, quickly adapt to environmental changes, and effectively improve the energy-saving rate and lighting safety level. This method has high environmental adaptability, energy consumption control ability, and system stability, and has broad application prospects and promotion value in smart cities.

[0084] In this embodiment, the original environmental perception dataset includes environmental light intensity data, weather condition data, traffic flow data, time information data, and historical energy consumption data.

[0085] By integrating multi-source data such as ambient light intensity, weather conditions, traffic flow, time information, and historical energy consumption, the present invention constructs a rich and comprehensive original dataset for environmental perception, providing reliable data support for accurate prediction of lighting demand and energy-saving control of streetlights. This method effectively improves the model's perception and response capabilities to environmental changes, realizes more intelligent and efficient dynamic management of streetlights, and has good system adaptability and practical application value.

[0086] In this embodiment, the generation of the multi-source local feature sub-sample set includes applying a sliding window mechanism to the original environmental perception dataset, setting the window size as w and the step size as s. At each sliding position k, w consecutive perception data vectors within the window are extracted to generate local feature sub-samples; the local feature sub-samples generated at all sliding positions k are numbered to form a multi-source local feature sub-sample set.

[0087] By applying a sliding window mechanism to the original environmental perception dataset, the present invention dynamically extracts multi-granularity local feature sub-samples, systematically capturing environmental change trends and local detail features. This method not only improves the spatio-temporal integrity and continuity of feature representation, but also enhances the input richness and accuracy of the subsequent lighting demand prediction model, significantly improving the environmental perception ability and dynamic response level of the intelligent streetlight control system.

[0088] In this embodiment, the specific steps of S2 include:

[0089] S21: Input the multi-source local feature sub-sample set into a primary random forest ensemble for training to obtain a primary random forest enhanced feature set, which is used to represent environmental features such as light, weather, and traffic in the Internet of Things perception data;

[0090] S22: Input the multi-source local feature sub-sample set into a completely random forest ensemble for training to obtain a primary completely random forest enhanced feature set, which supplements diverse environmental feature representations to improve the accuracy of lighting demand prediction;

[0091] S23: Perform feature-level splicing on the primary random forest enhanced feature set and the primary completely random forest enhanced feature set to form a primary enhanced feature set, which is used as the input for the subsequent lighting demand prediction model.

[0092] S24. Use the primary enhanced feature set as the input of the first-level cascade forest, train the first-level deep forest, process the primary enhanced feature set through multiple random forest classifiers and multiple completely random forest classifiers respectively, extract the class probability distribution features, splice the primary enhanced feature set with the class probability distribution features output by each forest, generate the first-level enhanced feature set, and generate the first-level lighting demand prediction distribution based on the first-level enhanced feature set to predict the street lamp lighting demand for the corresponding time period and area;

[0093] S25. Use the first-level enhanced feature set as the input, iteratively construct the l-th level cascade forest, and sequentially generate the l-th level enhanced feature set and the corresponding l-th level lighting demand prediction distribution:

[0094]

[0095] where Y l is the l-th level lighting demand prediction distribution, K is the number of random forest classifiers used in the current l-th layer, M is the number of completely random forest classifiers used in the current l-th layer, k is the random forest classifier index, is the weighting coefficient assigned to the k-th random forest classifier in the comprehensive prediction, is the lighting demand probability distribution output by the k-th random forest classifier for the enhanced feature set of the previous layer, F l-1 is the enhanced feature set output by the (l - 1)-th level deep forest, m is the completely random forest classifier index, is the weighting coefficient assigned to the m-th completely random forest classifier in the comprehensive prediction, is the lighting demand probability distribution output by the m-th completely random forest classifier for the enhanced feature set of the previous layer;

[0096] The l-th level lighting demand prediction distribution Y l is used to gradually improve the prediction accuracy of the street lamp lighting demand under environmental changes;

[0097] S26. Calculate the performance improvement rate based on the (l - 1)-th level lighting demand prediction distribution and the l-th level lighting demand prediction distribution. If the performance improvement rate is less than or equal to the set threshold ∈, terminate the cascade and output the currently trained deep forest prediction model; otherwise, continue to execute step S25.

[0098] The present invention constructs a phased and multi-level deep forest prediction model to fully exploit the spatio-temporal correlation and potential structural features of multi-source local feature sub-samples. Through the parallel feature enhancement of the primary random forest ensemble and the completely random forest ensemble, as well as the iterative construction and performance improvement rate control mechanism of the hierarchical cascaded deep forest, the continuous optimization of the feature expressiveness of the lighting demand prediction and the adaptive growth of the model are realized. This method not only effectively alleviates the overfitting problem of the traditional single-layer random forest in the high-dimensional complex feature space, but also improves the generalization ability and stability of the model in dynamic environmental changes, ensuring the continuous, accurate and low-energy consumption operation of the intelligent street lamp system in complex urban environments. The overall system has the advantages of flexible structure, efficient training, controllable performance, etc., significantly enhancing the practical application effect of street lamp intelligent perception and energy-saving control optimization.

[0099] In this embodiment, step S3 specifically includes:

[0100] S31. Encode the hyperparameter set of the deep forest prediction model into an antibody population P = {a1, a2, …, a N}, where each antibody a i represents a set of hyperparameter combinations, and the hyperparameter combinations include the number of random forests, the number of completely random forests, the number of decision trees inside each forest, and the maximum depth of each decision tree;

[0101] S32. Based on the hyperparameter configuration corresponding to each antibody a i , train the deep forest prediction model, and evaluate the prediction performance on the validation set, calculate the prediction mean square error and the energy-saving rate respectively, and define the affinity:

[0102] Affinity(a i ) = w1 × (1 - MSE(a i )) + w2 × ER(a i );

[0103] where Affinity(a i ) is the affinity of the deep forest prediction model corresponding to antibody a i under the comprehensive evaluation of lighting demand prediction and energy-saving effect, a i is a set of hyperparameter configurations of the deep forest model, i is the index number of the antibody, w1 is the prediction accuracy weight coefficient, MSE(a i ) is the mean square error of lighting demand prediction of the deep forest model trained with the configuration of antibody a i on the validation set, w2 is the energy-saving effect weight coefficient, and ER(a i ) is the deep forest model trained with the configuration of antibody a i ;

[0104] S33. Clone the antibody population according to the affinity, where the number of clones is proportional to the affinity, and generate a set of cloned antibodies;

[0105] S34. Perform a mutation operation on each antibody in the set of cloned antibodies. The mutation method is to apply a small perturbation to the hyperparameters to generate a set of mutant antibodies;

[0106] S35. Retrain the deep forest prediction model based on the set of mutant antibodies and calculate the affinity. Combine the original population and the mutant population, and select the N antibodies with the highest affinity from them to update to the new generation population;

[0107] S36. Iteratively execute the steps S32 to S35 until the average affinity change rate of the antibody population for consecutive G generations meets the convergence condition, and output the optimized deep forest prediction model:

[0108]

[0109] where, is the average affinity of the g-th generation of artificial immune population, is the average affinity of the (g - 1)-th generation of artificial immune population, g is the generation number of the current artificial immune evolution iteration, and δ is the convergence error threshold; the optimized deep forest prediction model is used to predict the street lamp lighting demand based on the Internet of Things environment perception data, and formulate an energy-saving control optimization strategy based on the prediction result.

[0110] In the present invention, by introducing an artificial immune algorithm to adaptively optimize the hyperparameters of the deep forest prediction model, the accuracy of lighting demand prediction and the adaptability of the model to dynamic environment changes are improved. By establishing an affinity evaluation system and combining the dual indicators of prediction accuracy and energy-saving effect, it is ensured that the optimization direction is more in line with the actual needs of intelligent energy-saving control. The cloning and mutation operations enhance the diversity of the search space and avoid falling into local optima. The population iteration and convergence determination mechanism effectively improve the optimization efficiency and reduce the consumption of training resources. The finally obtained optimized deep forest model can accurately predict the street lamp lighting demand in the Internet of Things environment, realize on-demand switching and brightness adjustment, significantly improve the energy-saving effect and intelligent response ability of the street lamp system, and has good application and promotion value.

[0111] In this embodiment, the S34 specifically includes:

[0112] S341. Select the set of cloned antibodies, apply the perturbation mechanism to each antibody's hyperparameter set separately, and perform small perturbations on the number of random forests, the number of completely random forests, the number of internal decision trees, and the maximum depth of the decision tree according to the set mutation rules to generate preliminary mutant antibodies;

[0113] S342. Uniformly integrate the mutation results of the preliminary mutant antibodies to form standardized mutant antibodies, ensuring that each hyperparameter still meets the preset value range and legality requirements after perturbation.

[0114] S343. Calculate the mutation amplitude of each mutant antibody. The mutation amplitude is defined as the average value of the normalized change rates of the original antibody and the mutant antibody on each hyperparameter.

[0115] S344. Determine whether the mutation amplitude of the mutant antibody exceeds the maximum allowable mutation amplitude threshold. If the mutation amplitude of the mutant antibody is greater than the maximum allowable mutation amplitude threshold, discard the mutant antibody. If the mutation amplitude of the mutant antibody is less than or equal to the maximum allowable mutation amplitude threshold, retain the mutant antibody to form a mutant antibody set.

[0116] In the present invention, by performing fine-grained hyperparameter perturbation on the cloned antibody and combining the mutation amplitude evaluation and legality verification mechanism, the stability and self-adaptability of the hyperparameter optimization process of the deep forest model are improved. Through step-by-step perturbation and standardization processing, it is ensured that each hyperparameter remains within a reasonable range after mutation, avoiding model failure; by introducing mutation amplitude control, reasonable mutant antibodies are screened out, avoiding confusion in the search direction caused by excessive deviation. This mechanism effectively expands the diversity of the antibody population, enhances the local search and global exploration capabilities of the artificial immune algorithm, significantly improves the optimization efficiency and convergence performance of the light demand prediction model in the dynamic Internet of Things environment, and further supports the intelligent and efficient operation of the street lamp energy-saving control system.

[0117] In this embodiment, the specific steps of S4 are as follows:

[0118] S41. Take the optimized deep forest prediction model as the input, and generate the light demand prediction result Y based on the input environmental perception data set. The light demand prediction result Y = {y1, y2,..., y T}, where y t represents the predicted light demand value of the street lamp at time step t.

[0119] S42. Based on the light demand prediction result Y, set the turn-on threshold and turn-off threshold to generate the street lamp switch control instruction C t ;

[0120] S43. For the street lamps in the on state, generate a brightness adjustment instruction according to the predicted light demand at the corresponding time step t:

[0121]

[0122] where L t is the street lamp brightness adjustment instruction at time step t, t is the time index, clip(·) is a limiting function, and y tThe predicted value of the lighting demand output by the optimized deep forest prediction model at time step t, θ off Is the preset lighting-on threshold, θ off Is the preset lighting-off threshold;

[0123] S44. Combine the switch control instruction C t Corresponding to time step t with the brightness adjustment instruction L t To form a complete street lamp control instruction set U.

[0124] In the present invention, by generating a predicted result of the lighting demand based on the optimized deep forest prediction model and combining the setting of the lighting-on threshold and the lighting-off threshold, intelligent switch control of street lamps at the time step level is realized. By introducing a brightness adjustment mechanism, a brightness instruction is dynamically generated according to the predicted lighting demand, and a limiting function is used to ensure a reasonable brightness adjustment ratio, significantly improving the adaptive response ability of the street lamp system. The complete control instruction set constructs a dynamic management framework integrating switch control and brightness adjustment, effectively reducing energy consumption and improving environmental adaptability. This method can accurately match the lighting demand according to real-time perception data in a changing Internet of Things environment, realizing more efficient and energy-saving intelligent scheduling of urban street lamps, and has broad application and promotion prospects.

[0125] In this embodiment, the street lamp switch control instruction includes, at each time step t, judging according to the predicted lighting demand and the set lighting-on threshold and lighting-off threshold. When the predicted lighting demand is greater than or equal to the lighting-on threshold, the street lamp switch control instruction is set to the on state; when the predicted lighting demand is less than or equal to the lighting-off threshold, the street lamp switch control instruction is set to the off state.

[0126] In the present invention, by setting the lighting-on threshold and the lighting-off threshold, dynamic switch control of street lamps based on the predicted result of the lighting demand is realized. At each time step, the state of the street lamp is intelligently judged according to the predicted lighting demand, effectively avoiding the energy waste caused by traditional timing control. This method can flexibly switch the street lamp on and off according to real-time environmental changes, improving the intelligent level and energy-saving effect of the lighting system, and is especially suitable for variable and complex urban lighting scenarios in the Internet of Things environment.

[0127] In this embodiment, the specific content of S6 includes:

[0128] S61. During the execution of the street lamp control instruction set U, collect the actual energy consumption data and actual environmental perception data corresponding to each time step t to generate a feedback data set;

[0129] S62. Based on the feedback data set and according to the theoretically calculated energy consumption data corresponding to the executed control instruction set, calculate the energy consumption deviation set for each time step:

[0130]

[0131] Among them, ΔE t is the energy consumption deviation at time step t, where t is the time index, is the actual street lamp energy consumption value collected in real time by the Internet of Things environment perception device at time step t, P base is the fixed standby power consumption of the street lamp device in the off state or the basic on state, C t is the switch control instruction, P max is the maximum power consumption when the street lamp works at the maximum brightness, L t is the brightness adjustment instruction;

[0132] S63. Based on the energy consumption deviation set, calculate the average energy consumption deviation, and set the energy consumption deviation correction threshold to ∈ E , if the average energy consumption deviation is greater than the energy consumption deviation correction threshold, trigger the optimization of the prediction model;

[0133] S64. In the case of triggering the optimization of the prediction model, retrain the deep forest prediction model based on the feedback data set, update the model parameters through the fine-tuning mechanism, and output the adaptively optimized deep forest prediction model.

[0134] The present invention dynamically triggers the adaptive optimization of the deep forest prediction model by collecting the actual energy consumption of street lamps and environmental perception data in real time, calculating the energy consumption deviation, and judging the model performance based on the average deviation. The fine-tuning mechanism is used to incrementally update the model parameters, improving the model's response ability and prediction accuracy to environmental changes. This method can effectively compensate for the performance decay problem of the model caused by environmental changes, ensure the continuous optimization of the lighting demand prediction and energy-saving control strategy, significantly improve the energy-saving effect and adaptive ability of the street lamp intelligent scheduling system, and is applicable to complex and changeable Internet of Things urban lighting scenarios.

[0135] In this embodiment, the S64 specifically includes:

[0136] S641. In the case of triggering the optimization of the prediction model, based on the feedback data set, extract the feature subset;

[0137] S642. Input the feature subset into the currently optimized deep forest prediction model. In the currently optimized deep forest prediction model, select the sub-forest models whose contribution rate to the overall energy consumption prediction error is greater than the preset threshold σ M , and form a set of sub-models to be fine-tuned by all the sub-forest models that meet the conditions;

[0138] S643. For each sub-model in the set of sub-models to be fine-tuned, perform incremental learning based on the feedback data set, adjust the original parameter set, and superimpose the parameter update amount obtained by training with the feedback data on the original parameters to generate a fine-tuned sub-model parameter set;

[0139] S644. Replace the corresponding set of sub-models to be fine-tuned in the currently optimized deep forest prediction model with the set of fine-tuned sub-models, keep the structures and parameters of other unadjusted sub-models unchanged, and output the adaptively optimized deep forest prediction model.

[0140] This method extracts feature subsets from feedback data, specifically screens sub-forest models that contribute significantly to energy consumption prediction errors, and performs incremental learning and parameter fine-tuning on these sub-forest models to form an optimized deep forest prediction model. By fine-tuning local sub-models rather than retraining the entire system, the computational cost and time delay of model optimization are significantly reduced, while also improving the model's adaptability to environmental changes. This mechanism ensures that the lighting demand prediction model maintains high accuracy and energy-saving performance in dynamic IoT environments, significantly enhancing the stability and practicality of intelligent streetlight control systems.

[0141] Example 1:

[0142] To demonstrate the feasibility of this invention, we applied it to a smart streetlight management system in a core business district of a large city. This area, approximately 5 square kilometers, encompasses main roads, secondary roads, sidewalks, and public squares, and boasts over 1,000 smart streetlights. This area is characterized by high traffic volume and dense pedestrian traffic. Furthermore, due to the high building density and frequent weather changes, ambient lighting conditions fluctuate significantly, making it both complex and representative.

[0143] In this application scenario, an IoT sensing terminal is first installed on each streetlight. The terminal device includes a light sensor, temperature and humidity sensor, air pressure sensor, and traffic flow monitoring module to collect environmental perception data in real time. The data is collected once a minute, and uploaded to the city lighting cloud management platform via a wireless IoT communication module for centralized processing. The proposed method extracts features from the continuously collected environmental data using a sliding window mechanism, with a window size of 30 minutes and a step size of 10 minutes. This method dynamically captures local trends in light, temperature, humidity, and traffic flow, generating a multi-source local feature subsample set.

[0144] Based on the generated local feature subsamples, the present invention establishes a deep forest prediction model, employing multiple random forests and completely random forests for initial processing. Layer-by-layer feature enhancement is achieved through a cascaded deep forest architecture. During the initial training phase, an artificial immune algorithm is applied to optimize the deep forest model's hyperparameters. These hyperparameters are encoded as antibody structures. An affinity function is defined based on illumination prediction error and energy-saving effects. Clonal amplification and mutation are then performed, and iterative optimization is performed to achieve a more adaptable deep forest prediction model.

[0145] Subsequently, based on the optimized prediction results of lighting requirements, a dynamic on-off control strategy and a brightness adjustment strategy for street lights are formulated. The system sets double thresholds for dynamic turning on and off of lights. When the predicted lighting requirement is higher than the turning-on threshold, the street lights are turned on; when it is lower than the turning-off threshold, the street lights are turned off; and within the threshold range, the brightness of the street lights is dynamically adjusted proportionally. The control instructions are sent to the street light terminals through the Internet of Things network to perform dynamic on-off actions and brightness adjustment actions in real time. At the same time, the street light terminals continue to upload the energy consumption data after execution and new environmental change data, and the platform performs feedback collection and model fine-tuning learning in real time according to the energy consumption deviation to achieve continuous adaptive optimization of the deep forest model.

[0146] During the three-month trial operation period, a comparative test was conducted between the intelligent street light system controlled by the method of the present invention and the intelligent street light system controlled by the traditional fixed-time on-off control. In the test area, the traditional control strategy performs on-off actions according to a fixed time schedule and static lighting thresholds, and cannot be adjusted according to actual weather changes or traffic density fluctuations, resulting in excessive energy consumption and insufficient lighting in some periods. However, the system of the present invention can dynamically adjust the turning-on time and brightness according to real-time perception data and prediction models, and significantly reduces the overall energy consumption on the premise of ensuring lighting safety.

[0147] According to the statistical data, during the periods when multiple typical rainy and cloudy weather conditions continuously occur, the system of the present invention can accurately predict the lighting requirements, reasonably delay or advance the on-off time periods of the lights, and reduce the daily average energy consumption by about 18% to 20% compared with the traditional system. During the period of continuous high incidence of haze, due to the lack of real-time adjustment ability, the traditional system has excessive lighting. The present invention reduces the brightness output in a timely manner through the prediction model, reducing the street light brightness adjustment by about 21% compared with the traditional system, while the visible distance of pedestrians and the vehicle passing safety index do not decrease significantly, ensuring traffic and pedestrian safety.

[0148] During the peak traffic periods of large holidays, the system of the present invention dynamically adjusts the brightness of street lights according to the real-time perception of the density of pedestrian and vehicle flows and the predicted demand, so that the main traffic arteries and crowded areas maintain a relatively high brightness, while the brightness of the side roads and areas with few people is appropriately reduced, and the daily average energy saving rate is increased by more than about 17%. Through comparative analysis, the traditional fixed control method generally has excessive lighting during peak periods, resulting in a large amount of redundant energy consumption, while the system of the present invention effectively suppresses the growth of unnecessary energy consumption on the basis of meeting the usage requirements.

[0149] The following table gives the comparative statistics of the energy consumption between the traditional control system and the application system of the present invention during the test period:

[0150] Table 1 Comparison of Energy Consumption between Traditional Control of Intelligent Street Lights and the Method of the Present Invention

[0151]

[0152] From the comparison data in Table 1 above, it can be seen that in the application scenario of urban intelligent lighting, the present invention is superior to the traditional control method in terms of energy consumption control and energy-saving efficiency. Specifically, during the continuous sunny days, the daily average energy consumption of the traditional system reaches 2000 kWh, while after the application of the method of the present invention, the daily average energy consumption drops to 1650 kWh, and the energy-saving rate reaches 17.5%. This shows that even under good lighting conditions where the traditional system seemingly works normally, the present invention can still significantly reduce energy consumption through a more refined dynamic adjustment strategy. Especially for the response ability to minute environmental changes, it effectively avoids the energy waste caused by the traditional system turning on lights too early or having too high brightness.

[0153] During the continuous rainy days, due to the unstable environmental light, the traditional system based on static threshold control causes the street lights to remain at high brightness for a long time, and the daily average energy consumption rises to 2105 kWh. While the system of the present invention can predict the actual lighting demand based on multi-source environmental perception data, and through dynamic regulation, the daily average energy consumption is controlled at about 1680 kWh, and the energy-saving rate is increased to 20.1%. Compared with the traditional scheme, the present invention not only achieves obvious energy consumption reduction, but also improves the adaptability of the system to abnormal weather conditions through predictive control means.

[0154] During the severe haze stage, there is usually a sharp attenuation of light, and it is difficult for the traditional method to adjust the lighting strategy in real time, resulting in a further increase in the daily average energy consumption to 2198 kWh. After adopting the method of the present invention, through the early perception of the visibility change trend and dynamic brightness control, the daily average energy consumption of the system drops to 1755 kWh, and the energy-saving rate reaches 20.2%. It can be seen that in a complex aerosol environment, the present invention can effectively avoid the problem of over-illumination caused by lagging environmental monitoring and improve the environmental adaptability of the overall system.

[0155] In special scenarios such as traffic peaks and holidays, due to the dense flow of people and vehicles, the traditional system usually sets the street lights to run at high brightness throughout the day to ensure traffic safety, resulting in a daily average energy consumption as high as 2075 kWh. The system of the present invention dynamically adjusts the brightness of street lights in different areas according to the real-time traffic flow changes. While ensuring high-brightness lighting on the main roads and areas with high population density, it reasonably reduces the lighting level on the side roads and areas with sparse people flow, reducing the daily average energy consumption to 1725 kWh and achieving an energy-saving effect of 16.9%. This regional refined control based on multi-source perception fully reflects the comprehensive optimization ability of the present invention in balancing safety and energy conservation.

[0156] On the night of an ordinary working day with low traffic, since the traditional control strategy fails to identify low pedestrian flow sections, it still maintains high-power lighting, and the average daily energy consumption remains at 1950 kWh. While the system of the present invention can timely reduce the brightness through the traffic density dynamic perception mechanism, and the average daily energy consumption is further reduced to 1615 kWh, with an energy saving rate reaching 17.2%. Especially during the late night to early morning period, the energy consumption reduction is particularly significant, effectively improving the overall energy utilization efficiency without having a negative impact on traffic safety.

[0157] In summary, through the comparison of a large amount of experimental data under continuous multi-stage, different weather and traffic conditions, it can be clearly seen that the intelligent street lamp energy-saving control method of the present invention, which combines Internet of Things perception and deep forest lighting demand prediction with artificial immune optimization and fine-tuning learning, can achieve a more significant energy-saving effect than the traditional fixed threshold control method in various typical urban application scenarios, while maintaining good lighting safety guarantee capabilities. Especially under complex conditions where the environment changes drastically (such as rainy, foggy, or hazy) or the density of pedestrians and vehicles fluctuates greatly, the system of the present invention demonstrates stronger environmental adaptability and energy consumption control capabilities, fully verifying the practical application value and broad promotion potential of the present invention in the field of intelligent lighting energy-saving optimization.

[0158] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. An intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things, characterized in that It includes the following steps: S1. Collect environmental data through an Internet of Things terminal, construct an original environmental perception dataset, and use a sliding window mechanism to perform multi-granularity feature extraction on the original environmental perception dataset to generate a multi-source local feature subsample set; S2. Construct a deep forest prediction model, input the multi-source local feature subsample set into multiple groups of random forests and completely random forests for preliminary processing, and combine the cascade deep forest structure to perform layer-by-layer feature enhancement and lighting demand prediction; S3. Use an artificial immune algorithm to optimize the deep forest prediction model, encode the hyperparameters of the deep forest prediction model into an antibody structure, define an affinity function, perform cloning amplification and mutation on high-affinity antibodies, and select the antibody with the highest affinity to update the hyperparameters of the deep forest prediction model to generate an optimized deep forest prediction model; S4. Based on the lighting demand prediction results output by the optimized deep forest prediction model, formulate a street lamp switch control strategy and a brightness adjustment strategy to obtain a control strategy instruction set; S5. Send the control strategy instruction set to the street lamp intelligent terminal to execute dynamic switch actions and brightness adjustment actions; S6. Collect the energy consumption feedback data and new environmental change data after the street lamp intelligent terminal executes, update the parameters of the deep forest prediction model through a fine-tuning mechanism, and output an adaptively optimized deep forest prediction model.

2. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 1, characterized in that The original environmental perception dataset includes environmental light intensity data, weather condition data, traffic flow data, time information data, and historical energy consumption data.

3. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 1, characterized in that, The generation of the multi-source local feature subsample set includes applying a sliding window mechanism to the original environmental perception dataset, setting the window size to w and the step size to s. At each sliding position k, extract w consecutive sensed data vectors within the window to generate a local feature subsample; number the local feature subsamples generated at all sliding positions k to form a multi-source local feature subsample set.

4. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 1, characterized in that Specifically, S2 includes: S21. Input the multi-source local feature subsample set into the primary random forest set for training to obtain a primary random forest enhanced feature set; S22. Input the multi-source local feature subsample set into the completely random forest set for training to obtain a primary completely random forest enhanced feature set; S23. Perform feature-level splicing on the primary random forest enhanced feature set and the primary completely random forest enhanced feature set to form a primary enhanced feature set; S24. Use the primary enhanced feature set as the input of the first-level cascade forest, train the first-level deep forest, process the primary enhanced feature set through multiple random forest classifiers and multiple completely random forest classifiers respectively, extract the category probability distribution features, and splice the primary enhanced feature set with the category probability distribution features output by each forest to generate a first-level enhanced feature set, and generate a first-level lighting demand prediction distribution based on the first-level enhanced feature set; S25. Use the first-level enhanced feature set as the input, iteratively construct the l-level cascade forest, and sequentially generate the l-level enhanced feature set and the corresponding l-level lighting demand prediction distribution: where Y l is the predicted distribution of the l-th level of light demand, K is the number of random forest classifiers used in the current l-th layer, M is the number of completely random forest classifiers used in the current l-th layer, k is the random forest classifier index, is the weighted coefficient assigned to the k-th random forest classifier in the comprehensive prediction, is the probability distribution of the light demand output by the k-th random forest classifier for the enhanced feature set of the previous layer, F l-1 is the enhanced feature set output by the depth forest of the (l - 1)-th layer, m is the completely random forest classifier index, is the weighted coefficient assigned to the m-th completely random forest classifier in the comprehensive prediction, is the probability distribution of the light demand output by the m-th completely random forest classifier for the enhanced feature set of the previous layer; S26. Calculate the performance improvement rate based on the predicted light demand distribution at the (l - 1) - th level and the predicted light demand distribution at the l - th level. If the performance improvement rate is less than or equal to the set threshold ∈, terminate the cascade and output the deep forest prediction model obtained from the current training; otherwise, continue to execute step S25.

5. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. Encode the hyperparameter set of the deep forest prediction model into an antibody population P = {a1, a2, …, a N}, where each antibody a i represents a set of hyperparameter combinations, and the hyperparameter combinations include the number of random forests, the number of completely random forests, the number of decision trees inside each forest, and the maximum depth of each decision tree; S32. Based on each antibody a i corresponding hyperparameter configuration, train a deep forest prediction model, evaluate the prediction performance on the validation set, calculate the prediction mean squared error and the energy saving rate respectively, and define the affinity: Affinity(a i ) = w1 × (1 - MSE(a i )) + w2 × ER(a i ); Among them, Affinity(a i ) is the affinity of the deep forest prediction model corresponding to antibody a i under the comprehensive evaluation of light demand prediction and energy-saving effect, a i is a set of hyperparameter configurations of the deep forest model, i is the index number of the antibody, w1 is the prediction accuracy weight coefficient, MSE(a i ) is the mean square error of the light demand prediction of the deep forest model trained with the configuration of antibody a i on the validation set, w2 is the energy-saving effect weight coefficient, ER(a i ) is the deep forest model trained with the configuration of antibody a i ; S33. Clone the antibody population according to the affinity, where the number of clones is proportional to the affinity, to generate a set of cloned antibodies. S34. Perform a mutation operation on each antibody in the set of cloned antibodies. The mutation method is to apply a small perturbation to the hyperparameters to generate a set of mutant antibodies. S35. Retrain the deep forest prediction model based on the set of mutant antibodies and calculate the affinity. Combine the original population and the mutant population, and select the N antibodies with the highest affinity from them to update to a new generation of population. S36. Iteratively execute steps S32 to S35 until the average affinity change rate of the antibody population for consecutive G generations meets the convergence condition, and output the optimized deep forest prediction model. wherein, is the average affinity of the g-th generation of the artificial immune population, is the average affinity of the (g-1)-th generation of the artificial immune population, g is the generation number of the current artificial immune evolution iteration, and δ is the convergence error threshold.

6. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 5, characterized in that The specific steps of S34 are as follows: S341. Select the set of cloned antibodies, and apply a perturbation mechanism to the hyperparameter set of each antibody separately. Apply small perturbations to the number of random forests, the number of completely random forests, the number of internal decision trees, and the maximum depth of decision trees according to the set mutation rules to generate preliminary mutant antibodies. S342. Uniformly integrate the mutation results of the preliminary mutant antibodies to form standardized mutant antibodies. S343. Calculate the mutation amplitude of each mutant antibody. The mutation amplitude is defined as the average of the normalized change rates of the original antibody and the mutant antibody on each hyperparameter. S344. Determine whether the mutation amplitude of the mutant antibody exceeds the maximum allowable mutation amplitude threshold. If the mutation amplitude of the mutant antibody is greater than the maximum allowable mutation amplitude threshold, discard the mutant antibody; if the mutation amplitude of the mutant antibody is less than or equal to the maximum allowable mutation amplitude threshold, retain the mutant antibody to form a set of mutant antibodies.

7. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Using the optimized deep forest prediction model as the input, based on the input environmental perception data set, generate a lighting demand prediction result Y, where the lighting demand prediction result Y = {y1, y2, …, y T}, where y t represents the predicted value of the road lamp lighting demand at time step t; S42. Based on the predicted result Y of the lighting demand, set the turning-on threshold and the turning-off threshold to generate the street lamp switch control instruction C corresponding to the time step t t ; S43. For street lights in the on state, generate a brightness adjustment instruction according to the predicted light demand at the corresponding time step t. Among them, L t is the street lamp brightness adjustment instruction at time step t, where t is the time index, clip(·) is the limiting function, and y t is the predicted value of the lighting demand output by the optimized deep forest prediction model at time step t, θ off is the preset turning-on threshold, and θ off is the preset turning-off threshold; S44. Form a complete street lamp control instruction set U by combining the switch control instruction C corresponding to the time step t t with the brightness adjustment instruction L t .

8. The intelligent perception and energy-saving control optimization method for street lights based on the Internet of Things according to claim 7, characterized in that The street - light switch control instruction includes, at each time step t, making a judgment based on the predicted light demand and the set turn - on threshold and turn - off threshold. When the predicted light demand is greater than or equal to the turn - on threshold, set the street - light switch control instruction to the on state; when the predicted light demand is less than or equal to the turn - off threshold, set the street - light switch control instruction to the off state.

9. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 1, characterized in that, The specific steps of S6 are as follows: S61. During the execution of the street - light control instruction set U, collect the actual energy consumption data and actual environmental perception data corresponding to each time step t to generate a feedback data set. S62. Based on the feedback data set, calculate the energy consumption deviation set for each time step according to the theoretically calculated energy consumption data corresponding to the executed control instruction set. Among them, ΔE t is the energy consumption deviation at time step t, where t is the time index, is the actual street lamp energy consumption value collected in real time by the Internet of Things environmental perception device at time step t, P base is the fixed standby power consumption of the street lamp device in the off state or the basic on state, C t is the switch control instruction, P max is the maximum power consumption of the street lamp when working at the maximum brightness, L t is the brightness adjustment instruction; S63. Calculate the average energy consumption deviation based on the energy consumption deviation set, and set the energy consumption deviation correction threshold to ∈ E . If the average energy consumption deviation is greater than the energy consumption deviation correction threshold, trigger the optimization of the prediction model; S64. In the case of triggering the optimization of the prediction model, retrain the deep forest prediction model based on the feedback data set, update the parameters of the deep forest prediction model through a fine - tuning mechanism, and output the adaptively optimized deep forest prediction model.

10. The intelligent perception and energy-saving control optimization method for street lamps based on the Internet of Things according to claim 9, characterized in that, The specific steps of S64 are as follows: S641. When triggering the optimization of the prediction model, extract a feature subset based on the feedback data set; S642. Input the feature subset into the currently optimized deep forest prediction model, and in the currently optimized deep forest prediction model, select the sub-forest models with the contribution rate of the overall energy consumption prediction error greater than the preset threshold σ M to form a set of sub-models to be fine-tuned by combining all the sub-forest models that meet the conditions; S643. For each sub-model in the sub-model set to be fine-tuned, perform incremental learning based on the feedback data set, adjust the original parameter set, and superimpose the parameter update amount obtained by training with the feedback data on the original parameters to generate a fine-tuned sub-model parameter set; S644. Replace the corresponding sub-model set to be fine-tuned in the currently optimized deep forest prediction model with the fine-tuned sub-model set, keep the structures and parameters of other unadjusted sub-models unchanged, and output the adaptively optimized deep forest prediction model.

Images