Intelligent road lighting control method and system based on vehicle-road-cloud collaboration

Abstract

The present application relates to the field of data processing, and specifically provides a road intelligent lighting control method and system based on vehicle-road-cloud collaboration. The method identifies vehicle flow based on a road time-series monitoring data set containing a target vehicle flow intensity, and obtains time-series distribution information. Based on this, the target vehicle flow intensity is subjected to evolution trend detection, which can reduce data computing consumption. Feature extraction is performed on the time-series distribution information of the target vehicle flow intensity in each road monitoring data sequence to obtain a time-series distribution sub-implicit representation sequence, and unqualified time-series distribution information can be deleted. The time-series distribution information corresponding to the target vehicle flow intensity in each road monitoring data sequence is unified to obtain a time-series distribution sub-implicit representation sequence, and generalization is improved. In addition, implicit representation based on the time order between sequences increases the accuracy of feature representation.

Description

Technical Field

[0001] The present application relates to the field of data processing, and more specifically, to a road intelligent lighting control method and system based on vehicle-road-cloud collaboration. Background Art

[0002] With the acceleration of urbanization, the pressure of road traffic is increasing, and higher requirements are put forward for the intelligent management of road lighting systems. Traditional road lighting systems mostly use time switches or manual control methods, which cannot flexibly adjust the lighting brightness according to real-time traffic conditions, which not only causes energy waste, but also makes it difficult to meet the needs of nighttime driving safety. Therefore, it is particularly important to develop an intelligent road lighting system that can automatically adjust the lighting strategy according to changes in traffic flow. In the prior art, although there are some road lighting control schemes based on traffic flow monitoring, most of these schemes rely on traffic flow data from a single monitoring point and lack in-depth analysis of the overall traffic flow trend of the road. At the same time, these schemes often have problems such as large computational complexity and poor real-time performance in data processing and analysis, making it difficult to achieve efficient and accurate control in a complex and changeable traffic environment. In addition, although the existing road monitoring equipment can collect rich traffic data, these data are often stored in raw form and lack effective feature extraction and pattern recognition mechanisms. Directly processing these raw data is not only inefficient, but also difficult to dig out the deep laws behind the data. Summary of the invention

[0003] In view of this, the present application provides a road intelligent lighting control method and system based on vehicle-road-cloud collaboration.

[0004] According to one aspect of an embodiment of the present application, a road intelligent lighting control method based on vehicle-road-cloud collaboration is provided, which is applied to a cloud service system, and the method comprises: obtaining a road time-series monitoring data set containing a target vehicle flow intensity, performing vehicle flow identification on multiple road monitoring data sequences in the road time-series monitoring data set, and obtaining time-series distribution information of the target vehicle flow intensity in each road monitoring data sequence; wherein the road time-series monitoring data set is collected by road-side monitoring equipment and vehicle monitoring equipment; performing feature extraction on the time-series distribution information of the target vehicle flow intensity in each road monitoring data sequence, and obtaining each road monitoring data sequence. Obtain the time sequence between the multiple road monitoring data sequences, merge the time sequence distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence, extract the traffic flow change characteristics of the implicit representation sequence time series set, and obtain the implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set; perform evolution type detection on the implicit representation of traffic flow evolution, and obtain the traffic flow evolution trend detection result corresponding to the target traffic flow intensity; formulate a road lighting strategy for the future target time period based on the traffic flow evolution trend detection result.

[0005] According to another aspect of an embodiment of the present application, a cloud service system is provided, including: a processor; and a memory, wherein the memory stores a computer-readable code, and when the computer-readable code is executed by the processor, the processor executes the method described above.

[0006] The beneficial effects of the present application include: based on obtaining a road time-series monitoring data set containing a target traffic flow intensity, the application performs traffic flow identification on multiple road monitoring data sequences in the road time-series monitoring data set, obtains the time-series distribution information of the target traffic flow intensity in each road monitoring data sequence, and detects the evolution trend of the target traffic flow intensity through the time-series distribution information of the target traffic flow intensity in each road monitoring data sequence, which can reduce data computing consumption. Feature extraction is performed on the time-series distribution information of the target traffic flow intensity in each road monitoring data sequence to obtain the time-series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence. Based on the feature extraction of the time-series distribution information of the target traffic flow intensity in each road monitoring data sequence, unqualified time-series distribution information can be deleted, and at the same time, the time-series distribution information corresponding to the target traffic flow intensity in each road monitoring data sequence is unified to obtain the time-series distribution sub-implicit representation sequence, and the generalization is improved. The time sequence between multiple road monitoring data sequences is obtained, and the implicit representation sequences of the time distribution sub-series corresponding to each road monitoring data sequence are merged into an implicit representation sequence time series set according to the time sequence. The traffic flow change features of the implicit representation sequence time series set are extracted to obtain the implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set. Based on this, the implicit representation is performed based on the time sequence between multiple road monitoring data sequences, which increases the accuracy of the feature representation. The evolution type of the implicit representation of traffic flow evolution is detected to obtain the traffic flow evolution trend detection result corresponding to the target traffic flow intensity, so as to formulate the road lighting strategy for the future target time period based on this, improve the intelligence of road lighting, and provide assistance for energy conservation and emission reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the architecture of an application scenario provided by this application;

[0008] Figure 2 This is a flow chart of a road intelligent lighting control method based on vehicle-road-cloud collaboration provided by this application;

[0009] Figure 3 It is a structural diagram of a cloud service system provided in an embodiment of the present application.

[0010] Figure numerals: 10 - cloud service system; 1001 - processor; 1002 - communication bus; 1003 - user interface; 1004 - network interface; 1005 - memory. DETAILED DESCRIPTION

[0011] In order to facilitate a clearer understanding of the present application, the application scenarios of the road intelligent lighting control method based on vehicle-road-cloud collaboration of the present application are first introduced. Figure 1As shown, the scenario includes a cloud service system 10 and a terminal cluster, and the terminal cluster may include one or more terminals, and the number of terminals is not limited here. Figure 1 As shown, the terminal cluster may specifically include terminal 1, terminal 2, ..., terminal n; it can be understood that terminal 1, terminal 2, terminal 3, ..., terminal n can all be connected to the cloud service system 10 through a network connection, so that each terminal can exchange data with the cloud service system 10 through the network connection.

[0012] It is understandable that the cloud service system 10 may refer to a server, and the cloud service system 10 may also be used to store data sent by a terminal. Among them, the server may be an independent physical server, or a server cluster or distributed system composed of at least two physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. The terminal may specifically refer to a vehicle-mounted terminal, a roadside camera, a radar speed meter, and the like.

[0013] For further information, see Figure 2 , is a flow chart of a road intelligent lighting control method based on vehicle-road-cloud collaboration provided by an embodiment of the present application. Figure 2 As shown, this method can be Figure 1 The cloud service system 10 is executed, wherein the road intelligent lighting control method based on vehicle-road-cloud collaboration may include the following steps:

[0014] Step S100: Obtain a road time-series monitoring data set containing target traffic flow intensity, perform traffic flow identification on multiple road monitoring data sequences in the road time-series monitoring data set, and obtain the time-series distribution information of the target traffic flow intensity in each road monitoring data sequence; wherein, the road time-series monitoring data set is collected by road-side monitoring equipment and vehicle monitoring equipment.

[0015] In step S100, the cloud service system, as the execution subject, is responsible for efficiently and accurately acquiring and analyzing the road time series monitoring data from multiple sources. First, the cloud service system collects data through a variety of monitoring devices (including road-side monitoring devices and vehicle monitoring devices) integrated on the road. Road-side monitoring devices may include high-definition cameras, infrared sensors, radar speed guns, etc., which can monitor the traffic conditions on the road in real time, including information such as vehicle type, speed, and distance. At the same time, vehicle monitoring equipment such as vehicle-mounted GPS positioning systems and OBD (on-board diagnostic system) data interfaces can also provide dynamic information such as vehicle driving trajectories and speed changes. The raw data generated by these devices are transmitted to the cloud service system in real time to form a road time series monitoring data set. Taking a certain city trunk road as an example, high-definition cameras and radar speed guns are installed at certain intervals on both sides of the road. The camera is responsible for capturing vehicle images on the road and identifying vehicle types through image recognition technology; the radar speed gun is responsible for measuring the real-time speed of each vehicle. At the same time, the vehicle-mounted GPS data of passing vehicles is also regularly uploaded to the cloud service system. These data are organized in time series to form a time series monitoring data set for the road. After obtaining these data, the cloud service system enters the core link of traffic flow identification. This process may involve the application of machine learning models, especially the combination of time series analysis models and image processing models. The system will first preprocess the raw data, including data cleaning (removing noise, outliers, etc.), data normalization (ensuring that different features are compared on the same scale) and other steps. Then, the cloud service system uses the pre-trained traffic flow identification model to analyze the processed data. This model may be built based on a recurrent neural network (RNN) or a long short-term memory network (LSTM) because they are good at processing sequence data and capturing time dependencies. The model input is multiple road monitoring data sequences in the road time series monitoring data set. Each sequence contains a series of data points arranged in chronological order, and each data point contains multiple features such as vehicle speed, vehicle distance, and vehicle type.

[0016] For example, a data sequence may be recorded in the form of: [{timestamp: T1, speed: 60km / h, distance: 10m, vehicle type: small car}, {timestamp: T2, speed: 65km / h, distance: 8m, vehicle type: large car}, ...]. The model will traverse this sequence, analyze the data points at each time point, and use internal parameters and algorithm logic to calculate the traffic flow intensity at that moment. The "traffic flow intensity" here is a comprehensive indicator, which may be calculated based on a weighted calculation of multiple factors such as the number of vehicles, vehicle speed, and vehicle type.

[0017] In order to more accurately identify changes in traffic flow, the model can also consider temporal continuity. For example, the difference between the traffic flow intensity at the current time point and the previous time point is compared, and whether this difference exceeds a preset threshold, so as to judge the change trend of traffic flow (increase, decrease or stability). Finally, the cloud service system obtains the time series distribution information of the target traffic flow intensity in each road monitoring data sequence through the traffic flow identification model. This information is presented in the form of structured data, which may include the traffic flow intensity value at each time point, the change trend (increase, decrease or stability), etc. For example, the output may be a time series array: [{timestamp: T1, traffic flow intensity: 100}, {timestamp: T2, traffic flow intensity: 120, trend: increase}, ...], which clearly shows the traffic flow status of the target road at different time points.

[0018] Step S200: extracting features from the time series distribution information of the target vehicle flow intensity in each road monitoring data sequence, and obtaining a time series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence.

[0019] Step S200 involves in-depth feature extraction of the time series distribution information of the target traffic flow intensity in each road monitoring data sequence to construct a unique time series distribution sub-implicit representation sequence for each sequence. This process requires not only accurate capture of key information in the data, but also effective feature conversion technology to convert the original time series data into high-level feature representations that are easy to process for machine learning models.

[0020] Feature extraction is an important step in the field of machine learning. It transforms the original data to extract key information that is helpful for learning tasks. In the scenario of intelligent road lighting control, although the time series distribution information of the target traffic flow intensity intuitively reflects the changes in traffic flow, the original data often contains noise and redundant information, and there may be differences between different sequences. Therefore, through feature extraction, the cloud service system can remove irrelevant information and retain key features that are important indicators of the evolution trend of traffic flow, providing a more valuable data basis for subsequent analysis and decision making.

[0021] In step S200, the cloud service system may use a variety of feature extraction techniques, including but not limited to statistical feature extraction, model-based feature extraction, and deep learning feature extraction. For example, the cloud service system first organizes the time series distribution information output by step S100 to ensure that each road monitoring data sequence has been converted into time series data in a unified format. For example, each sequence may contain traffic flow intensity values ​​at multiple time points, expressed in array form: [traffic flow intensity_T1, traffic flow intensity_T2, ..., traffic flow intensity_Tn].

[0022] In order to extract effective features from time series data, the cloud service system may select or build a deep learning model suitable for processing time series data, such as a hybrid model of convolutional neural network (CNN) combined with recurrent neural network (RNN), or long short-term memory network (LSTM). These models can capture the temporal dependencies and local features in sequence data. Take LSTM as an example. It is a special RNN that solves the problem of gradient vanishing or gradient exploding that is prone to occur in traditional RNN during long sequence training by introducing a "gate" mechanism (forget gate, input gate, output gate). The cloud service system can build a neural network model containing multiple LSTM layers for feature extraction.

[0023] The time series distribution information of each road monitoring data sequence is used as input and passed into the input layer of the LSTM model. The traffic flow intensity value at each time point is converted into a feature vector (here, since the traffic flow intensity is a single value, the feature vector may contain only one element, but in more complex cases, the feature vector may contain multiple related indicators). The data is processed through the "gate" mechanism of the LSTM layer. At each step, the LSTM unit updates its internal state based on the current input and the output and state of the previous time step, and outputs the hidden state of the current time step. This hidden state is actually a comprehensive representation of the information of the current time step and all previous time steps, containing rich time series features. Although the direct output of the LSTM model is the hidden state of each time step, the cloud service system usually does not directly use these states as the final features. Instead, it can adopt a certain strategy to summarize or transform these hidden states to generate a sequence of sub-implicit representations of the time series distribution corresponding to each sequence. For example, the average of all hidden states in each sequence or the last hidden state can be selected as the feature representation of the sequence.

[0024] For example, suppose that the time series distribution information of a road monitoring data sequence is: [50, 60, 70, 80, 90,100]. After being processed by the LSTM model, a hidden state vector is output at each time step. To simplify the explanation, assume that each hidden state vector contains only one element (in actual situations, the hidden state vector usually contains multiple elements). Then, the hidden state sequence of the sequence after the LSTM model may be: [h1, h2, h3, h4, h5, h6]. The cloud service system can select the average value of these hidden states [(h1+h2+...+h6) / 6] as the sub-implicit representation of the time series distribution of the sequence, or simply use the last hidden state h6 as the representation. Through the feature extraction of step S200, the cloud service system can convert the original time series distribution information into a higher-level and more abstract feature representation. These feature representations not only remove the noise and redundant information in the original data, but also capture the key patterns of changes in traffic flow intensity.

[0025] Step S300: Obtain the time sequence between multiple road monitoring data sequences, merge the time series distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence, extract the traffic flow change characteristics of the implicit representation sequence time series set, and obtain the implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set.

[0026] In step S300, the cloud service system first obtains the time sequence between multiple road monitoring data sequences. These sequences are sorted according to the time points of data collection, reflecting the changes in traffic flow in different time periods. Then, according to these time sequences, the system merges the implicit representation sub-sequences of the time series distribution obtained in step S200 for each sequence to form a unified implicit representation sequence time series set.

[0027] For example, suppose there are three road monitoring data sequences, which are collected from three different positions (position A, position B, position C) on the same road section, and their time order is known to be position A -> position B -> position C. After step S200, each position obtains a temporal distribution sub-implicit representation sequence, such as the sequence of position A is [Fa1, Fa2, ..., Fan], the sequence of position B is [Fb1, Fb2, ..., Fbm], and the sequence of position C is [Fc1, Fc2, ..., Fcn] (where Fi represents the temporal distribution sub-implicit representation of position i at time t). The cloud service system merges these sequences into an implicit representation sequence temporal set in chronological order, such as [[Fa1, ..., Fan], [Fb1, ..., Fbm], [Fc1, ...,Fcn]].

[0028] The merged implicit representation sequence time series set contains the traffic flow feature information of multiple road monitoring points at different time points. In order to reveal the intrinsic connection and overall evolution trend between these features, the cloud service system further extracts the traffic flow change features. This process usually relies on complex machine learning models, especially those that are good at processing sequence data, such as recurrent neural networks (RNN), long short-term memory networks (LSTM) or gated recurrent units (GRU). For example, the cloud service system can choose a LSTM-based deep learning model to extract traffic flow change features. The model can accept the implicit representation sequence time series set as input, and capture the long-term dependencies and short-term change patterns in the sequence through its internal "gate" mechanism (forget gate, input gate, output gate). During the training process, the model will learn how to calculate the hidden state at the current moment based on the hidden state at the previous moment and the input at the current moment, and use this hidden state to predict or represent the change characteristics of traffic flow.

[0029] In the specific implementation of feature extraction, the LSTM model will traverse each subsequence in the implicit representation sequence time series set in chronological order. For each time step, the model will calculate the hidden state of the current time step based on the current time series distribution sub-implicit representation and the hidden state of the previous time step. This hidden state not only contains the information of the current time step, but also integrates the information of all previous time steps through the "gate" mechanism, thus forming a comprehensive representation of the current traffic flow evolution trend. After being processed by the LSTM model, the cloud service system obtains a sequence containing the hidden states of all time steps. This sequence is actually an implicit representation of the overall evolution trend of the target traffic flow intensity in the road time series monitoring data set. The cloud service system can use this sequence as the basis for subsequent analysis to detect the evolution type of traffic flow, predict future traffic flow changes, etc.

[0030] Step S400: performing evolution type detection on the implicit representation of traffic flow evolution to obtain a traffic flow evolution trend detection result corresponding to the target traffic flow intensity.

[0031] In step S400, the cloud service system uses its powerful computing power and pre-trained machine learning model to deeply explore the implicit representation of traffic flow evolution to reveal the evolution trend of traffic flow.

[0032] Before performing the evolution type detection, the cloud service system performs necessary preprocessing on the implicit representation of the traffic flow evolution generated in step S300. For example, it includes steps such as normalization, standardization or smoothing of the data to ensure the consistency and comparability of the data. For example, if the implicit representation of the traffic flow evolution exists in the form of a time series array, the cloud service system can smooth the data by methods such as sliding average to reduce noise interference.

[0033] In order to accurately identify the evolution type of traffic flow, the cloud service system selects or builds a suitable machine learning model. Given that traffic flow evolution has the characteristics of time series data, models such as recurrent neural networks (RNN), long short-term memory networks (LSTM), and gated recurrent units (GRU) are ideal choices because they are good at processing sequence data. The cloud service system can decide which model to use based on the specific application scenario and data characteristics, or combine the advantages of multiple models through methods such as ensemble learning. After determining the machine learning model, the cloud service system uses historical traffic flow data for model training. In this process, the cloud service system will use the known traffic flow evolution trend as a label and input it into the model together with the corresponding traffic flow evolution implicit representation data. Through multiple iterative training, the model will gradually learn the mapping relationship between the traffic flow evolution trend and the traffic flow evolution implicit representation. In order to further improve the accuracy and generalization ability of the model, the cloud service system can also tune the model, including adjusting model parameters, adding regularization terms, and using dropout and other strategies to prevent overfitting.

[0034] After completing the model training, the cloud service system inputs the newly generated implicit representation of traffic flow evolution into the trained model to detect the evolution type. This process is essentially the process of the model predicting the corresponding traffic flow evolution trend based on the input data. The cloud service system will determine the evolution type of traffic flow based on the output results of the model, such as growth type, decay type, fluctuation type, etc.

[0035] For example, suppose the cloud service system selects an LSTM-based machine learning model to detect the type of traffic flow evolution. In the training phase, the cloud service system uses one year of historical traffic flow data as the training set. These data have been processed through steps S100 to S300 to obtain the corresponding implicit representation of traffic flow evolution. At the same time, these data are also accompanied by labels of traffic flow evolution trends annotated by experts (such as "continuous growth", "increase first and then decrease", etc.). The cloud service system inputs these data and labels into the LSTM model for training until the model's performance on the validation set reaches a stable and satisfactory level.

[0036] In the detection phase, the cloud service system receives a set of new traffic flow monitoring data, and after processing in steps S100 to S300, obtains a new implicit representation of traffic flow evolution. The cloud service system inputs this set of data into the trained LSTM model, and the model outputs a prediction result after internal calculation, indicating that the traffic flow evolution trend corresponding to this set of data is "continuous growth". The cloud service system formulates corresponding lighting strategies based on this prediction result, such as increasing the brightness of street lights to cope with the upcoming traffic peak.

[0037] Step S500: Formulate a road lighting strategy for a future target time period based on the traffic flow evolution trend detection result.

[0038] In step S500, the cloud service system formulates the most appropriate road lighting strategy in the future target time period according to the traffic flow evolution trend detection result obtained in step S400. In step S500, the cloud service system converts the traffic flow evolution trend into a specific lighting adjustment instruction and sends it to the road lighting system for execution.

[0039] Specifically, the cloud service system first reviews the detection result of the traffic flow evolution trend obtained in step S400. This result may indicate that the traffic flow will show different change patterns such as growth, decrease or fluctuation in a certain time period in the future. For example, the detection result may indicate that the traffic flow of a certain road section is expected to increase by 30% in the next two hours. Based on the traffic flow evolution trend, the cloud service system further evaluates the lighting demand in the time period. This includes considering the impact of road type (such as main roads, secondary roads, branch roads, etc.), traffic flow, vehicle speed, and weather conditions (such as sunny days, rainy days, foggy days, etc.) on lighting. For example, on main roads and when the traffic flow is expected to increase significantly, higher illumination is required to ensure driving safety; on rainy or foggy days, in order to improve visibility, the lighting brightness also needs to be increased. According to the lighting demand, the cloud service system determines specific lighting parameters, such as light intensity, color temperature, lighting range, etc. The setting of these parameters should be aimed at improving road safety, driving comfort and energy efficiency. For example, in order to improve visibility during nighttime driving, the cloud service system selects a higher light intensity and a wider lighting range; in the late night when there is less traffic, the light intensity can be appropriately reduced to save energy. (Specific parameter settings can be adaptively set according to actual needs and specific environments, and this application does not limit this)

[0040] When developing lighting strategies, the cloud service system can comprehensively consider energy efficiency and operating costs. By intelligently scheduling the working hours and power output of lighting equipment, it is possible to reduce energy consumption and maintenance costs while meeting lighting needs. For example, the cloud service system can dynamically adjust the on / off time and brightness level of street lights based on historical data and real-time traffic flow forecasts to avoid unnecessary energy waste.

[0041] Based on the above analysis, the cloud service system generates a set of road lighting strategies for future target time periods. This strategy specifies the lighting parameter settings for each road section and each time period, and sends instructions to the road lighting system for execution through network communication technology. For example, the system can send instructions to the lighting controller of a specific road section, requiring it to gradually increase the light intensity to a preset level within the next two hours.

[0042] For example, suppose that the traffic volume on a city's main road is expected to increase by 30% in the next two hours. The cloud service system first analyzes this trend and identifies the road section as a busy traffic area. Next, the system evaluates the lighting needs and decides to increase the lighting brightness of the road section in consideration of nighttime driving safety and driving comfort. When determining the lighting parameters, the cloud service system sets a 20% increase in light intensity and an unchanged color temperature to maintain a comfortable visual environment. At the same time, the system also expands the lighting range by optimizing the lighting layout and light distribution to ensure clear driving vision. Finally, based on energy efficiency and cost considerations, the cloud service system develops a detailed lighting strategy and sends instructions to the lighting controller of the road section through the network for execution.

[0043] As an implementation manner, the plurality of road monitoring data sequences include a road monitoring data sequence S m , m is not greater than the number G of multiple road monitoring data sequences, 1≤m≤G. Based on this, step S100, traffic flow recognition is performed on multiple road monitoring data sequences in the road time series monitoring data set to obtain the time series distribution information of the target traffic flow intensity in each road monitoring data sequence, including:

[0044] Step S110: road monitoring data sequence S m Perform arbitrary segmentation to obtain K time series intervals to be selected, where K ≥ 1;

[0045] Step S120: extracting data items from the K time series intervals to be selected respectively, and obtaining interval data items corresponding to the K time series intervals to be selected respectively;

[0046] Step S130: According to the data item value range related to the target traffic flow intensity, the interval data items contained in the K candidate time series intervals are divided into data items to obtain the target data items in each candidate time series interval, and the road monitoring data sequence S is divided into the target data items based on the target data items. m Sequence segmentation is performed to obtain a binary tag set corresponding to the target traffic flow intensity; the data item value corresponding to the target data item belongs to the data item value range;

[0047] Step S140: Determine the target vehicle flow intensity in the road monitoring data sequence S based on the binary tag set. m The timing distribution information in .

[0048] In step S110, the cloud service system processes the road monitoring data sequence S mPreliminary processing is performed to facilitate a more detailed analysis of traffic flow changes. Since road monitoring data is usually collected continuously and contains a large number of time point records, directly processing the entire sequence can reduce efficiency due to the large amount of data. Therefore, the cloud service system adopts a segmentation strategy to split the long sequence into multiple shorter candidate time series intervals for parallel processing or step-by-step analysis.

[0049] For example, suppose the road monitoring data sequence S m The traffic flow data of a certain road section is recorded every one minute in a day. In order to simplify the processing, the cloud service system decides that each selected time series interval contains 10 minutes of data (i.e., records within 600 seconds). m The total duration is 24 hours (i.e. 1440 minutes), so theoretically it can be divided into 144 candidate time series intervals. However, the actual segmentation may be more flexible. For example, considering the difference in data density during peak and non-peak traffic hours, the cloud service system may use a shorter interval length during peak hours and a longer interval during non-peak hours. The segmentation process can be to index the original sequence according to the preset time interval. The cloud service system maintains a timestamp array or list to record the start and end time of each candidate time series interval.

[0050] After dividing the long sequence into multiple candidate time series intervals, the cloud service system performs a detailed analysis of the data in each interval. This step mainly extracts data items related to traffic flow identification from each interval. These data items may include but are not limited to vehicle counts, average speeds, vehicle type distribution, etc. For example, assume that each candidate time series interval contains multiple data points, and each data point records the number of vehicles passing through, the average speed, and the type ratio of vehicles passing through at that time point. The cloud service system will traverse the data points in each interval and extract these key data items. For example, in the first 10-minute candidate time series interval, the system may extract the following data items: the vehicle count is 300, the average speed is 50km / h, of which small cars account for 70% and large cars account for 30%.

[0051] The data item extraction process depends on the structure and format of the original data. The cloud service system may need to customize the data parsing logic according to different data sources and data collection devices. For example, if the data is stored in the form of CSV files, the system may need to use regular expressions or pandas libraries to read and parse the files; if the data is stored in a database, the system may need to execute SQL queries to obtain the required data items.

[0052] In step S130, the cloud service system selects data items that are highly relevant to the target traffic flow intensity according to the predefined data item value range. These data items can often directly reflect the changes in traffic flow, such as vehicle counts, vehicle passing rate in a specific time period, etc. By dividing these key data items, the system can further determine at which time points or time periods the traffic flow intensity reaches a specific threshold.

[0053] For example, suppose that the cloud service system sets a threshold for traffic intensity (such as more than 50 vehicles passing through per minute is considered high traffic intensity). Based on the interval data items extracted in step S120, the system will check the vehicle count at each time point. If the vehicle count at a certain time point exceeds 50, then that time point is considered to be a moment of high traffic intensity. In order to simplify subsequent processing, the system will mark these moments of high traffic intensity as 1 (indicating "yes") and other moments as 0 (indicating "no"). In this way, for each selected time series interval, the system will generate a corresponding binary label set (or mask) to represent the distribution of traffic intensity in the interval.

[0054] The data item segmentation process usually involves conditional judgment logic. The cloud service system will traverse the interval data item array according to the preset threshold and conditional expression, perform judgment operations on each element, and generate a binary label based on the judgment result. In order to improve efficiency, the system can use vectorized operations or parallel processing technology to accelerate this process. In addition, it is worth noting that the setting of the threshold is crucial for the identification of traffic flow intensity. In practical applications, this threshold may need to be dynamically adjusted according to the road type, traffic flow characteristics, and lighting control requirements. The cloud service system can use historical data or expert knowledge to set the initial threshold, and iteratively optimize based on feedback in practical applications.

[0055] After the previous steps, the cloud service system has generated a corresponding binary tag set for each candidate time series interval. These sets intuitively show the distribution of traffic flow intensity in each interval. However, in order to obtain the entire road monitoring data sequence S m In order to obtain the complete time series distribution information of the target traffic flow intensity, the system also needs to integrate these local binary label sets.

[0056] For example, suppose the cloud service system has mThe sequence is segmented and a set of binary tags is generated for each candidate time interval. Now, the system reassembles these sets according to their time order in the original sequence to form a continuous time series distribution graph or array. In this array, the value of each element represents the traffic flow intensity state at the corresponding time point (1 represents high traffic flow intensity, and 0 represents non-high traffic flow intensity). In this way, the system can clearly see the changes in traffic flow intensity during the entire monitoring period, providing an important basis for subsequent feature extraction and lighting strategy formulation.

[0057] The integration process of time series distribution information mainly involves the merging operation of arrays or lists. The cloud service system determines the position of each candidate time series interval in the original sequence based on the start and end timestamps of each interval, and inserts the corresponding binary markers into the correct position. If there are time intervals between adjacent intervals (i.e., due to missing data or inconsistent sampling rates), the system may need to use interpolation or filling strategies to fill these intervals to ensure the continuity and integrity of the time series distribution information.

[0058] In addition, in order to more intuitively display the time series distribution characteristics of traffic flow intensity, the cloud service system can also use visualization technology to convert the time series distribution array into a chart form (such as a line chart, bar chart, etc.). These charts can not only help operators quickly understand the trend of data changes, but also serve as reports or presentation materials to show the analysis results to relevant personnel.

[0059] Step S100 uses a series of refined processing flows (S110 to S140) to enable the cloud service system to effectively extract the time series distribution information related to the target traffic flow intensity from the road time series monitoring data set. This information provides a solid foundation for subsequent feature extraction, traffic flow evolution trend detection, and lighting strategy formulation.

[0060] As an implementation method, step S140 determines the target vehicle flow intensity in the road monitoring data sequence S based on the binary tag set. m The timing distribution information in the , may specifically include:

[0061] Step S141: Determine the target vehicle flow intensity in the road monitoring data sequence S based on the binary tag set. m The distribution coordinates in ;

[0062] Step S142: performing curve fitting on the distribution coordinates to determine A curve segments, where A≥1;

[0063] Step S143: Obtain the connection point between any two adjacent curve segments in the A curve segments, obtain the time series distribution coordinates corresponding to the connection point, and generate the target vehicle flow intensity in the road monitoring data sequence S based on the time series distribution coordinates corresponding to the connection point. mThe timing distribution information in .

[0064] In step S141, the main task of the cloud service system is to map each label (such as 0 or 1) in the binary label set back to the original road monitoring data sequence S m The corresponding time points in the sequence are used to determine the specific distribution coordinates of the target traffic flow intensity (i.e., the time points with binary marks of 1) in the sequence. These distribution coordinates are the basis for subsequent curve fitting and time series distribution information generation.

[0065] For example, suppose the road monitoring data sequence S m The traffic flow monitoring data is recorded every hour from 0:00 to 24:00 (i.e., the time range of one day). The cloud service system has generated a set of binary tags based on this data, where the point marked as 1 indicates that the traffic flow intensity within that hour exceeds the preset threshold (such as the number of vehicles passing through per hour exceeds a certain number). For example, the binary tag set may be [0, 0, 1, 1, 0, 0, 1, 0, ..., 0], where the traffic flow intensity of the 3rd and 4th hours (counting from 0) and the 7th hour exceeds the threshold. Based on this set of binary tags, the cloud service system determines these "high traffic flow intensity" time points in the original sequence S n The specific positions in the distribution are their distribution coordinates. In this example, the distribution coordinates will be the timestamps or index values ​​of the 3rd, 4th, and 7th hours.

[0066] For example, the cloud service system maintains a table of correspondence between timestamps and data indexes. For each tag in the binary tag set, the system determines its corresponding timestamp or index value by looking up the table. During the mapping process, the cloud service system performs a consistency check to ensure that each binary tag can be accurately mapped to the corresponding position in the original sequence.

[0067] After determining the distribution coordinates of the target traffic flow intensity, the cloud service system needs to perform curve fitting on these coordinates. Curve fitting is a mathematical method used to construct a smooth curve through a series of data points (here, distribution coordinates) to approximate the trend of the original data. In this scenario, the purpose of curve fitting is to more intuitively display the changing trend of traffic flow intensity and facilitate subsequent analysis.

[0068] For example, continuing with the previous example, assume that the cloud service system has determined three distribution coordinates of high traffic intensity, corresponding to the 3rd hour, the 4th hour, and the 7th hour. The system now needs to fit these discrete coordinate points into one or more smooth curves. Since these coordinate points may not completely follow a simple linear relationship (for example, they may represent the alternation of peak and non-peak traffic hours), the cloud service system can choose a more flexible curve model, such as a polynomial curve, a spline curve, or a Gaussian process regression. In actual operation, the cloud service system can first try to use a simple linear model for fitting, but it will soon be found that this model cannot capture the nonlinear changes in the data well. Therefore, the system can switch to a higher-order polynomial model or a spline curve model. After multiple iterations and optimizations, the system finally determines an optimal curve fitting scheme that connects the three distribution coordinate points into a smooth curve, and may also include other "low traffic intensity" areas that are not directly represented by binary labels but exist logically outside this curve. However, it should be noted that in this specific scenario, since we only care about the time period of "high traffic intensity" (i.e., the points with binary value 1), the main purpose of curve fitting is not to accurately predict the traffic flow value at each time point, but to grasp the changing trend of traffic flow intensity from a macro perspective. Therefore, in the fitting process, the cloud service system can pay more attention to the overall shape and trend of the curve, rather than the precise matching of a single data point.

[0069] The cloud service system can select a suitable curve model based on the characteristics of the data and the analysis requirements. In the selection process, the system can consider multiple factors such as the complexity of the model, the goodness of fit (such as R² value), residual analysis, etc. Once the curve model is determined, the cloud service system estimates the parameters of the model. For example, by solving an optimization problem to find the parameter value that minimizes the model fitting error. For example, the optimization algorithm may include the least squares method, the gradient descent method, the Newton method, etc. After the fitting is completed, the cloud service system verifies the model to ensure that it can not only fit the known data points (i.e., the distribution coordinates) well, but also predict unknown data points or trends to a certain extent.

[0070] It is worth noting that since we only care about the time period of “high traffic intensity”, the fitted curve may not completely cover the entire original sequence S n time range. Instead, it may only contain those time periods that are binary-marked as 1 and the transition areas between them. This processing method helps to more clearly identify the changing characteristics of traffic flow intensity and provide more targeted guidance for subsequent lighting strategy formulation.

[0071] In step S142, the cloud service system obtains one or more smooth curves through curve fitting to represent the change trend of the target traffic flow intensity. However, these curves are continuous, and in practical applications, it is often necessary to discretize this continuous change trend into a series of specific time points or time periods (i.e., time series distribution information) to match the operation cycle of the road lighting system.

[0072] Therefore, in step S143, the task of the cloud service system is to find the connection points between adjacent curve segments in the fitting curve (i.e., the points where the direction of the curve changes significantly) and map these connection points back to the original sequence S n These connection points represent the turning points or key points of the traffic intensity change trend, which are of great significance in the formulation of lighting strategies.

[0073] For example, assume that the cloud service system fits a smooth curve containing multiple curve segments in step S142 to represent the changing trend of the target traffic flow intensity. These curve segments may correspond to different traffic conditions (such as peak traffic, flat traffic, and valley traffic). The cloud service system now needs to find the connection points between these curve segments, that is, the points where the slope of the curve changes significantly.

[0074] In actual operation, the cloud service system can use methods such as derivative analysis, extreme point detection or edge detection to identify these connection points. For example, the system can calculate the derivative of the fitting curve (or its first-order difference approximation), and then find the points whose absolute value of the derivative exceeds a certain threshold as potential connection point candidates. Then, the system will further screen and verify these candidate points to ensure that they truly represent the turning point of the trend of traffic flow intensity changes. Once all the connection points are found, the cloud service system maps these points back to the original sequence S n These time points or time periods will serve as key references for adjusting lighting strategies. For example, if a connection point represents the time point when the transition from low traffic intensity to high traffic intensity begins, the cloud service system can gradually increase the road lighting brightness before that time point to cope with the upcoming traffic peak.

[0075] The cloud service system detects the connection points based on the mathematical properties of the fitted curve, for example, by taking derivatives or differences of the curve to find points where the slope changes significantly. After finding the connection points, the cloud service system maps these points back to the original sequence S n The specific time point or time period in the connection point. This may need to be completed in combination with the timestamp information of the original data. Based on the time series distribution coordinates corresponding to the connection point and the original sequence S nThe cloud service system will generate specific lighting strategy recommendations based on other relevant information (such as weather conditions, holiday arrangements, etc.). These recommendations will guide the road lighting system on when and how to adjust the lighting brightness to meet traffic lighting needs in different time periods. It is worth noting that in actual applications, since changes in traffic flow are often affected by multiple factors (such as weather, holidays, special events, etc.), the fitted curves and detected connection points may only provide a rough reference framework. When formulating lighting strategies, the cloud service system also needs to comprehensively consider these factors and feedback from real-time traffic data to continuously adjust and optimize the strategies.

[0076] Step S140 uses a series of refined processing flows (S141 to S143) to enable the cloud service system to accurately determine the target vehicle flow intensity in the road monitoring data sequence S based on the binary tag set. m This information not only reveals the changing trend and key points of traffic flow intensity, but also provides strong data support for the subsequent lighting strategy formulation. In this way, the cloud service system can more intelligently control the road lighting system to adapt to the traffic lighting needs in different time periods, thereby improving road safety and energy efficiency.

[0077] As an implementation mode, step S200 performs feature extraction on the time series distribution information of the target vehicle flow intensity in each road monitoring data sequence to obtain a time series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence, which may specifically include:

[0078] Step S210: merging the time series distribution information of the target traffic flow intensity in multiple road monitoring data sequences to obtain a time series distribution array corresponding to the target traffic flow intensity, inputting the time series distribution array into a feature extraction component in the traffic flow evolution trend detection network, filtering the time series distribution array, and obtaining time series filtering information corresponding to the time series distribution array;

[0079] Step S220: performing a normalization operation on the time series filtering information to obtain the normalized time series filtering information;

[0080] Step S230: Based on the nonlinear transformation operator in the feature extraction component, the time series filtering information after the normalization operation is activated to generate a time series distribution implicit representation sequence corresponding to the time series distribution information of the target vehicle flow intensity in multiple road monitoring data sequences;

[0081] Step S240: Based on the time series distribution implicit representation sequence, determine the time series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence.

[0082] In step S210, the cloud service system merges the time series distribution information of the target traffic flow intensity in multiple road monitoring data sequences to form a unified time series distribution array. This array will become the input of the subsequent feature extraction process. Subsequently, the system inputs the array into the feature extraction component in the traffic flow evolution trend detection network to extract the key features in the array through filtering (or convolution) operations.

[0083] For example, suppose there are three road monitoring data sequences, denoted as Seq1, Seq2, and Seq3, each of which contains the time series distribution information of the target traffic flow intensity over a period of time. The cloud service system first aligns the time series distribution information in these sequences according to the timestamps and merges them into a large time series distribution array. For example, if each sequence records the traffic flow intensity data from 0 to 24 o'clock (in hours), then the merged array will contain 72 data points (because there are 24 hours * 3 sequences = 72 data points). Next, the cloud service system inputs this time series distribution array into the feature extraction component. The feature extraction component may be a CNN-based model that contains multiple convolutional layers. In the convolutional layer, the system uses a series of convolution kernels (also called filters) to perform a sliding window operation on the array to extract local features. Each convolution kernel corresponds to a set of weight parameters, which are optimized by the backpropagation algorithm during the training process.

[0084] Assuming that the size of the convolution kernel is k×1 (because one-dimensional time series data is processed here), the step size is s, and the padding is p, the convolution operation can be expressed as:

[0085] ;

[0086] Where x is the input time series distribution array, w is the weight of the convolution kernel, y is the output time series filter information, and t is the current time index. i is an index variable used to traverse the weight w[i] of the convolution kernel and the corresponding fragment of the input signal x[t]. i starts from 0 and ends at k-1, where k is the length of the convolution kernel (or the window size or filter size of the convolution kernel). For each value of i, the input signal x[t] at position The product of the value at position i and the weight w[i] of the convolution kernel at position i. The weight parameters of the convolution kernel (i.e., the value in w) are updated during the training process by minimizing the loss function (such as mean square error, cross entropy, etc.). The loss function calculates the difference between the model's predicted value and the actual value.

[0087] After obtaining the time series filtering information, the cloud service system standardizes the information to ensure that they are on the same scale, thereby improving the efficiency and accuracy of subsequent processing. The standardization operation usually includes two steps: de-meaning and normalization. For example, assuming that the time series filtering information is an array containing multiple time points, the cloud service system will first calculate the expectation (also called the mean) and dispersion (also called the standard deviation) of the array. Then, for each element in the array, the mean is subtracted and divided by the standard deviation to obtain a standardized array. This process is similar to the Z score calculation in statistics.

[0088] Specifically, the mean μ of the time series filtering information is calculated and subtracted from each element. The standard deviation σ (or the square root of the variance σ2) of the time series filtering information is calculated and each element after the mean is removed is divided by the standard deviation.

[0089] Although the time series filter information after normalization removes the scale difference, it is still linear. In order to capture the nonlinear features in the data, the cloud service system applies nonlinear transformation operators (such as activation functions) to further process the data. For example, in the feature extraction component, an activation layer is usually added after the convolution layer to perform a nonlinear transformation on the convolution result. Common activation functions include ReLU (Rectified Linear Unit), Sigmoid, and Tanh. Taking the ReLU function as an example, it sets the part of the input value less than 0 to 0 and keeps the part greater than 0 unchanged. This operation helps to introduce nonlinear factors, allowing the model to learn more complex patterns. For each element x in the normalized time series filter information, the ReLU activation function maps it to:

[0090] f(x)=max(0,x);

[0091] This way, all negative values ​​are set to 0, while positive values ​​remain unchanged.

[0092] In addition to ReLU, activation functions such as Sigmoid and Tanh can also be used for this purpose. The Sigmoid function maps the input value to the interval (0,1), while the Tanh function maps it to the interval (-1,1). These functions have their own advantages and disadvantages, and they need to be selected according to specific circumstances in practical applications.

[0093] After nonlinear transformation, the cloud service system obtains the implicit representation sequence of the time series distribution corresponding to the time series distribution information. This sequence captures the key features and nonlinear relationships in the original time series data, providing strong support for subsequent analysis and decision-making.

[0094] After obtaining the overall implicit representation sequence of the time series distribution, the cloud service system divides the implicit representation sequence according to the original road monitoring data sequence, thereby obtaining the sub-implicit representation sequence of the time series distribution corresponding to each sequence. For example, assume that the merged time series distribution array is merged based on three road monitoring data sequences (Seq1, Seq2, Seq3). After obtaining the overall implicit representation sequence of the time series distribution, the cloud service system divides the implicit representation sequence according to the starting and ending positions of each sequence in the original data. For example, if Seq1 occupies the first 24 data points of the merged array (corresponding to the time range from 0 to 24 o'clock), the cloud service system will assign the first 24 elements of the implicit representation sequence to Seq1 as its sub-implicit representation sequence of the time series distribution. The cloud service system can maintain an index table that records the starting and ending positions of each road monitoring data sequence. When dividing the implicit representation sequence, the system will locate the corresponding part of each sequence in the implicit representation sequence according to this index table, and divide it out as the sub-implicit representation sequence of the time series distribution of the sequence.

[0095] During the segmentation process, the cloud service system ensures that the implicit representation sequence of the time series distribution of each sequence is consistent with the original data in time. This means that if a sequence in the original data has missing or overlapping parts in time, the corresponding implicit representation sequence of the time series distribution sub-series will also be missing or overlapping accordingly.

[0096] Through the above steps, the cloud service system successfully extracted the time series distribution characteristics of the target traffic flow intensity from multiple road monitoring data sequences, and generated the time series distribution sub-implicit representation sequences corresponding to each sequence. These representation sequences not only retain the key information and nonlinear relationships in the original data, but also provide rich data support for the subsequent detection of traffic flow evolution trends and the formulation of lighting strategies.

[0097] As an implementation mode, step S230, based on the nonlinear transformation operator in the feature extraction component, activates the time series filtering information after the normalization operation to generate a time series distribution implicit representation sequence corresponding to the time series distribution information of the target vehicle flow intensity in multiple road monitoring data sequences, which may include:

[0098] Step S231: based on the nonlinear transformation operator in the feature extraction component, the time series filtering information after the normalization operation is activated to obtain a candidate time series distribution implicit representation sequence corresponding to the time series distribution information in the plurality of road monitoring data sequences;

[0099] Step S232: Obtain the output dimension and output scale in the feature extraction component, and optimize the candidate time series distribution implicit representation sequence based on the output dimension and output scale to obtain the time series distribution implicit representation sequence corresponding to the time series distribution information of the target traffic flow intensity in multiple road monitoring data sequences; the scale corresponding to the time series distribution implicit representation sequence is the multiplication result of the output scale and the output dimension.

[0100] In step S231, the cloud service system uses the nonlinear transformation operator (such as ReLU, Sigmoid, Tanh and other activation functions) in the feature extraction component to activate the time series filter information after the normalization operation. The introduction of the activation function is to introduce nonlinear factors so that the model can learn complex patterns in the data rather than simple linear combinations. For example, suppose the time series filter information after normalization is a one-dimensional array, denoted as X norm =[x1,x2,…,x n ], where n is the number of time points. The cloud service system selects ReLU as the activation function for X norm The mathematical expression of the ReLU function is f(x)=max(0,x), which means that it sets all negative values ​​to 0 and keeps the positive values ​​unchanged.

[0101] After applying the ReLU activation function, the cloud service system obtains a new array X act =[f(x1),f(x2),…,f(x n )], that is, the candidate time series distribution implicitly represents the sequence. Each element in this new array is the result of nonlinear transformation of the corresponding element in the original array, reflecting the nonlinear characteristics in the data.

[0102] In practical applications, the choice of activation function depends on the specific task and data characteristics. In addition to ReLU, Sigmoid and Tanh are also commonly used activation functions. The Sigmoid function maps the input to the (0,1) interval, which is suitable for the output layer of binary classification problems; the Tanh function maps the input to the (-1,1) interval, which has the advantage of zero centering, but may cause gradient vanishing problems in deep networks. Since the activation function is applied element by element, this process can be highly parallelized. Cloud service systems can use GPUs or dedicated hardware accelerators to accelerate activation calculations and increase processing speed.

[0103] After generating the candidate time series distribution implicit representation sequence, the cloud service system further considers the output dimension (such as the number of channels) and output scale of the feature extraction component in order to optimize the candidate sequence and obtain the final time series distribution implicit representation sequence. This process may involve operations such as dimension transformation and scale adjustment to ensure that the output sequence matches the requirements of the target task. For example, assume that the output dimension of the feature extraction component is C (the number of channels), and the length of the candidate time series distribution implicit representation sequence corresponding to each channel is n (the same as the length of the original time series filtering information). Then, without considering any additional operations, the candidate time series distribution implicit representation sequence can be regarded as a two-dimensional array (or tensor) with a shape of (n, C). However, in some cases, it may be necessary to adjust the shape or scale of this tensor to meet the needs of subsequent processing or analysis. For example, if the subsequent traffic flow evolution trend detection network expects the input to be a one-dimensional vector or a tensor with a specific shape, the cloud service system needs to optimize the candidate time series distribution implicit representation sequence accordingly. This processing may include operations such as dimension compression, splicing, and rearrangement.

[0104] Suppose you want to integrate the information of C channels into one channel in order to obtain a one-dimensional sequence of implicit representations of time series distributions. The cloud service system can use methods such as average pooling or max pooling to aggregate the information of each channel. Taking average pooling as an example, the cloud service system averages the values ​​of C channels at each time point to obtain a one-dimensional array of length n as the final sequence of implicit representations of time series distributions. However, in this particular scenario, since you are concerned with the feature representation at each time point in the time series data rather than simple aggregate statistics, the more likely choice is to retain the information of all channels but adjust the shape of the tensor. For example, the cloud service system can reshape a two-dimensional tensor of shape (n, C) into a one-dimensional tensor of shape (n×C, 1) (if subsequent processing does require a one-dimensional input). But it is more common to keep the shape of the two-dimensional tensor unchanged or perform other forms of transformation according to specific needs.

[0105] However, in order to be closer to actual application scenarios and maintain clarity of explanation, it is assumed here that the cloud service system chooses to retain the information of all channels but may need to adjust the scale of the sequence to match the input requirements of the subsequent network. For example, if the subsequent network expects the input scale to be twice the original scale but the number of channels remains unchanged, the cloud service system may need to increase the length of the sequence without changing the number of channels through methods such as upsampling or interpolation. But please note that this is not common in practical applications because upsampling is usually used for processing spatial data such as images rather than one-dimensional time series data; for time series data, a more common approach may be to keep the scale unchanged or downsample to reduce the amount of computation. However, in this particular step, it is assumed that the cloud service system simply adjusts the shape of the candidate time series distribution implicit representation sequence according to the output dimension and scale (if necessary) without changing its actual content or performing other complex transformation operations.

[0106] The output dimension and scale are usually determined by the design of the feature extraction component. In deep learning, these parameters are often determined in the network structure design stage and will be adjusted and optimized according to the specific task to achieve the best performance. According to specific needs, the cloud service system can use a variety of shape adjustment operations including but not limited to reshape, transpose, concatenate, etc. These operations usually do not change the content of the data itself but change the organization of the data for subsequent processing or analysis. When performing any shape adjustment or scale transformation, the cloud service system needs to ensure that key information is retained and will not be lost or distorted due to the transformation. This requires the cloud service system to fully consider the characteristics of the data and the requirements of the task when designing feature extraction components and formulating optimization strategies.

[0107] Step S230 introduces nonlinear factors through activation functions and optimizes the candidate time series distribution implicit representation sequence using output dimension and scale information to finally generate the time series distribution implicit representation sequence of the target traffic flow intensity. This process not only retains the key features in the original data, but also makes the output sequence more in line with the needs of subsequent processing or analysis through nonlinear transformation and shape adjustment, providing strong data support for subsequent traffic flow evolution trend detection and lighting strategy formulation.

[0108] As an implementation manner, step S220, performing a normalization operation on the time series filtering information to obtain the time series filtering information after the normalization operation, may include:

[0109] Step S221: clustering the time series filtering information based on the standard operator in the feature extraction component to obtain B time series filtering clusters, where B≥1;

[0110] Step S222: obtaining the expected values ​​and dispersions respectively corresponding to the B time series filter clusters, and based on the expected values ​​and dispersions respectively corresponding to the B time series filter clusters, performing a normalization operation on the time series filter information in each time series filter cluster to obtain the normalized time series filter information.

[0111] In step S221, the cloud service system uses the canonical operator (clustering algorithm) in the feature extraction component to perform clustering processing on the time series filter information. Clustering is an unsupervised learning method that aims to divide the samples in the data set into several subsets (clusters) so that the similarity of samples in the same cluster is high, while the similarity of samples between different clusters is low. In the scenario of time series filter information, clustering can help the cloud service system group filter information with similar features or trends into a group for subsequent unified processing. For example, suppose the time series filter information is a one-dimensional array X filter =[x1,x2,…,x n ], where n is the number of time points. The cloud service system selects K-means clustering algorithm as the canonical operator for X filter K-means algorithm is a widely used clustering algorithm, which iteratively finds K cluster centers in the data and assigns each sample point to the cluster to which the nearest cluster center belongs.

[0112] The cloud service system first randomly selects K sample points as the initial cluster centers. In practical applications, the value of K can be determined based on prior knowledge or cross-validation. For example, if it is known that there are obvious morning and evening peak and off-peak periods in the time series filtering information, K=3 can be selected as the number of initial clusters.

[0113] For X filter Each element x in i , the cloud service system calculates its distance to each cluster center (usually Euclidean distance is used as a metric) and converts x i Assign to the cluster to which the nearest cluster center belongs. Based on the current cluster assignment results, the cloud service system recalculates the mean of each cluster as the new cluster center.

[0114] Repeat the above steps until the change in cluster center is less than a preset threshold or the maximum number of iterations is reached.

[0115] After K-means clustering, the cloud service system obtains B time series filter clusters (B=K), each of which contains a group of time series filter information with similar characteristics or trends.

[0116] In addition to Euclidean distance, other distance metrics such as Manhattan distance and cosine similarity can also be selected, depending on the characteristics of the data and the requirements of the task.

[0117] The K-means algorithm is sensitive to the selection of the initial cluster center. Different initial cluster centers may lead to different clustering results. Therefore, in practical applications, improved algorithms such as K-means++ can be used to optimize the selection of the initial cluster center. In addition to the change in the cluster center, other convergence conditions can be set to terminate the iteration process, such as minimizing the intra-cluster variance.

[0118] After obtaining B time series filter clusters, the cloud service system standardizes the time series filter information within each cluster. Standardization is a common data preprocessing method that aims to reduce the differences between different samples by removing the mean of the data and scaling it to a uniform scale. In this step, the cloud service system uses the expected value and dispersion of each cluster to calculate the standardization factor and transform the time series filter information within the cluster accordingly. For example, for each time series filter cluster C j (j=1,2,…,B), the cloud service system performs the following operations: The cloud service system calculates the mean of all elements in the cluster as the expected value of the cluster. The variance or standard deviation of the cluster is further calculated as a measure of the dispersion of the cluster. Each time series filter information x in the cluster is standardized using the expected value and standard deviation of the cluster. After this processing, the time series filter information in the cluster will be scaled to a distribution with a mean of 0 and a standard deviation of 1. When calculating the standard deviation, if all elements in the cluster are equal (that is, the variance is zero), the standard deviation cannot be calculated directly. At this time, special treatment can be taken, such as standardizing all elements to 0, or keeping the original value unchanged. In actual calculations, in order to avoid the accuracy problem of floating-point operations, the cloud service system can use numerically stable methods to calculate the variance and standard deviation. For example, the sum of the squares of all elements and the square of the sum can be calculated first, and then the variance can be calculated using these two values.

[0119] Through the clustering and standardization processing of step S220, the cloud service system successfully converts the original time series filtering information into a set of standardized time series data. These data not only remove the influence of dimension and distribution differences, but also retain the important features of the original data, providing a more stable and reliable foundation for subsequent feature extraction and traffic flow evolution trend detection.

[0120] As an implementation mode, step S300, obtaining the time sequence between multiple road monitoring data sequences, merging the time series distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence, extracting the traffic flow change feature of the implicit representation sequence time series set, and obtaining the implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set, may include:

[0121] Step S310: obtaining a time sequence between multiple road monitoring data sequences, and merging the time series distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence;

[0122] Step S320: Based on the time recursive component of the traffic flow evolution trend detection network, the implicit representation sequence time series set is recursively embedded to obtain an embedding array corresponding to the implicit representation sequence time series set;

[0123] Step S330: Based on the nonlinear transformation operator in the time recursive component, the embedded array is activated to obtain an implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set.

[0124] In step S310, the primary task of the cloud service system is to clarify the time order between multiple road monitoring data sequences, and merge the time series distribution sub-implicit representation sequences corresponding to each sequence into a unified implicit representation sequence time series set in this order. This set will serve as input data for subsequent vehicle flow change feature extraction. For example, suppose that the cloud service system monitors three adjacent road sections (section A, section B, section C), and each section has a corresponding road monitoring data sequence (Seq_A, Seq_B, Seq_C). These sequences record the vehicle flow information of each section at different time points, and have been processed through the previous steps to obtain their respective time series distribution sub-implicit representation sequences (Rep_A, Rep_B, Rep_C).

[0125] The cloud service system first obtains the timestamp information of the three sequences and confirms that their time order is Seq_A ->Seq_B ->Seq_C, that is, the data of segment A precedes segment B, and the data of segment B precedes segment C. Then, the cloud service system merges Rep_A, Rep_B, and Rep_C into a new implicit representation sequence time series set according to this time order. The elements in this set will be arranged in the order of segment A, segment B, and segment C, and each element is a time series distribution sub-implicit representation sequence.

[0126] Before merging the sequences, the cloud service system ensures that the timestamps of all sequences are aligned. If some time points are missing in some sequences, the system may need to fill in these missing points through interpolation or filling strategies.

[0127] The merged implicit representation sequence time series collection may be a multidimensional array or tensor that contains the time series distribution sub-implicit representations of all sequences. For example, if each subsequence is of length T and there are three sequences, the merged collection may be a three-dimensional tensor of shape (3, T, D), where 3 is the number of sequences, T is the length of the sequence, and D is the feature dimension at each time point.

[0128] In step S320, after obtaining the implicit representation sequence time series set, the cloud service system uses the time recursive component (such as RNN, LSTM or GRU, etc.) in the traffic flow evolution trend detection network to recursively embed the set. The purpose of recursive embedding is to capture the time dependency in the sequence and link each time point in the sequence with its previous and next time points, thereby generating an embedding array containing the dynamic information of the time series.

[0129] For example, the cloud service system chose an LSTM-based RNN model for recursive embedding. LSTM is a special RNN variant that solves the shortcomings of traditional RNN in long-term dependency issues by introducing a "gate" mechanism (forget gate, input gate, output gate).

[0130] Before inputting the implicit representation sequence time series collection into the LSTM model, the cloud service system may need to perform some preprocessing on the data, such as dimension transformation or normalization. The model then processes the elements in the collection one by one in the order of the sequence. For the input at each time point t (that is, the sub-implicit representation of all sequences at time point t), the LSTM unit calculates the hidden state and output of the current time step based on the hidden state of the previous time step and the input of the current time step. This process is repeated until the entire sequence collection is processed.

[0131] The operation of the LSTM unit at each time step t can be summarized as the following steps: the forget gate determines which information is discarded from the unit state, the input gate determines which new information is stored in the unit state, and then the unit state is updated, specifically combining the output of the forget gate and the input gate to update the unit state, and based on the updated unit state, decide what value to output. In RNN, parameters (such as weight matrices and bias terms) are shared across all time steps, which helps the model learn common patterns in time series.

[0132] In step S330, after recursive embedding, the cloud service system obtains an embedding array containing dynamic information of the time series. However, the values ​​in this array are usually linear or not nonlinearly transformed, so the cloud service system uses the nonlinear transformation operator (such as activation function) in the time recursive component to activate the embedded array. The introduction of the activation function helps the model learn complex patterns and nonlinear relationships in the data. For example, after the output layer of the LSTM model, the cloud service system adds a fully connected layer (also called a dense layer or linear layer) and applies a nonlinear activation function after this layer. Common activation functions include ReLU, Sigmoid, and Tanh. In this scenario, the cloud service system may choose the ReLU function as the activation function because it can introduce nonlinear factors while maintaining the stability of the gradient (the gradient is always 1 in the positive interval).

[0133] The ReLU activation function operates on each element in the embedded array element by element: if the element value is less than 0, it is set to 0; if the element value is greater than or equal to 0, it remains unchanged. This process generates a new array as the final implicit representation of the evolution of traffic flow.

[0134] The choice of activation function depends on the specific task and data characteristics. The ReLU function is widely used in deep learning models due to its simplicity and effectiveness, but in some cases other activation functions (such as Leaky ReLU, PReLU, etc.) may need to be considered to avoid the problem of neuron death. In the fully connected layer, each input node is connected to the output node and linearly transformed through the weight matrix and bias term. The activation function is then applied to the result of the linear transformation to introduce nonlinear factors. The dimension of the final implicit representation of the traffic flow evolution depends on the output dimension of the fully connected layer. This dimension can be set according to specific needs. For example, it can be set to the same dimension as the input sequence for subsequent comparative analysis or dimensionality reduction for visualization.

[0135] Step S300 successfully integrates and analyzes the implicit representation sequences of the time series distribution subsequences of multiple road monitoring data sequences through a series of operations such as clear time order, recursive embedding and nonlinear activation to obtain the implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set. By introducing the RNN model and nonlinear activation function, the cloud service system can capture the complex dynamic changes in the time series data, thereby realizing the accurate prediction and control of traffic flow changes.

[0136] As an implementation manner, the plurality of road monitoring data sequences include a road monitoring data sequence H m and road monitoring data sequence H m+1 , road monitoring data sequence H m With the road monitoring data sequence H m+1 Adjacent, m+1 is not greater than the number G of multiple road monitoring data sequences, 1<m+1≤G. Based on this, step S320, based on the time recursive component of the traffic flow evolution trend detection network, recursively embeds the implicit representation sequence time series set to obtain the embedding array corresponding to the implicit representation sequence time series set, which may include:

[0137] Step S321: Obtain the road monitoring data sequence H in the implicit representation sequence time series set m The corresponding first memory state array stores the road monitoring data sequence H m The corresponding time series distribution sub-implicit representation sequence and the first memory state array are loaded into the time recursive component of the traffic flow evolution trend detection network to detect the road monitoring data sequence H mThe corresponding time series distribution implicit representation sequence is embedded to obtain the road monitoring data sequence H m The corresponding representation array;

[0138] Step S322: Based on the road monitoring data sequence H m The corresponding representation array determines the road monitoring data sequence H in the implicit representation sequence time series set m+1 The corresponding second memory state array stores the road monitoring data sequence H m+1 The corresponding time series distribution sub-implicit representation sequence and the second memory state array are loaded into the time recursive component of the traffic flow evolution trend detection network to detect the road monitoring data sequence H m+1 The corresponding time series distribution implicit representation sequence is embedded to obtain the road monitoring data sequence H m+1 The corresponding representation array;

[0139] Step S323: If the road monitoring data sequence H m+1 When it is the last road monitoring data sequence among multiple road monitoring data sequences, an embedding array corresponding to the implicit representation sequence time series set is generated based on the representation arrays corresponding to each road monitoring data sequence in the multiple road monitoring data sequences.

[0140] In step S321, the cloud service system first focuses on the first (or any one, but the first one is used as an example here) road monitoring data sequence H in the implicit representation sequence time series set. m The system uses the time recursive component in the traffic flow evolution trend detection network to detect H m The corresponding temporal distribution sub-implicit representation sequence is recursively embedded to capture its internal temporal dependencies and generate the corresponding representation array.

[0141] For example, suppose the road monitoring data sequence H m is a sequence of data at T time points. After the data at each time point is processed in the previous steps, a D-dimensional sub-implicit representation vector of the time series distribution is obtained. Therefore, H m The corresponding time series distribution implicit representation sequence can be viewed as a two-dimensional array with a shape of (T, D).

[0142] The cloud service system first obtains m The first memory state array associated with H. This memory state array may be generated by the time recursive component when processing the previous sequence (if any), and it contains the state information after the previous sequence is processed. m For example, if it is the first sequence in a sequence set, you may need to use an initial state (such as an all-zero vector) as the first memory state array.

[0143] Next, the cloud service system will m The time series distribution implicit representation sequence and the first memory state array are loaded into the time recursive component. The component processes the data at each time point in the sequence one by one in chronological order, and uses the internal loop mechanism (such as the forget gate, input gate, and output gate of LSTM) to update the memory state and generate the output at the current time point. This process is repeated T times until the entire sequence is processed. Finally, the cloud service system obtains H m The corresponding representation array. This array is also a two-dimensional array, and its shape may be the same as the input temporal distribution implicit representation sequence (but in some cases it may be different due to dimensionality reduction or dimensionality increase operations). Each element in the representation array is a high-level abstract representation of the original input data, which not only contains the information of the original data but also incorporates the temporal dependencies learned by the time recursive component.

[0144] In a time recursive component, the memory state is transferred across time points. When processing the current time point, the component will refer to the memory state of the previous time point and update this state according to the current input. In this way, even if the sequence is very long, the component can capture information at earlier time points. In models such as RNN or LSTM, in addition to the output, there is usually a hidden state that represents the internal state at the current time point. In some cases, the hidden state itself can be used as a representation array, while in other cases, the hidden state may need to be further processed (such as through a fully connected layer) to obtain the final representation array.

[0145] In step S322, after successfully processing H m The cloud service system then focuses on the adjacent road monitoring data sequence H m+1 . Due to H m and H m+1 are adjacent and the road traffic conditions are often continuous, so the cloud service system can use the m The learned information is used to assist in processing H m+1 .

[0146] For example, the cloud service system is first based on H m The corresponding characterization array is used to determine H m+1 The corresponding second memory state array. This process may involve H m Some aggregation or transformation operation of the representation array to obtain a single value that can represent H m The vector of the overall state is H m+1 However, in practical applications, it is more common to directly use the H m The last hidden state output by the post-time recursive component is taken as H m+1The initial memory state of H (because the hidden state itself already contains H m With the initial memory state, the cloud service system can process H m The same approach is used to handle H m+1 It will m+1 The implicit representation sequence of the time series distribution and the second memory state array are loaded into the time recursive component and recursive embedding is performed on each time point in the sequence. After the processing is completed, the cloud service system will obtain H m+1 The corresponding representation array.

[0147] For the first sequence in a sequence set, its initial memory state may be preset (such as a vector of all zeros), while for subsequent sequences, the state after the previous sequence is processed can be used as the initial state. This approach helps the model capture the continuity between sequences. m+1 When the time recursive component will be based on H m+1 The input and m The initial memory state passed is used to update the memory state and generate output. This process is similar to processing H m Similar but different initial memory states can lead to different processing results.

[0148] In step S323, after processing all the road monitoring data sequences in sequence, the cloud service system integrates the representation arrays corresponding to these sequences into a unified embedding array as input for subsequent processing.

[0149] For example, if the number of road monitoring data sequences is G and the cloud service system has processed all sequences in chronological order (from H1 to H G ) then it will have G representation arrays. These arrays may have the same shape (although their exact contents will vary depending on the original sequence they correspond to).

[0150] The cloud service system can use a simple concatenation operation to concatenate these representation arrays along a certain dimension (such as the sequence dimension) to form a larger two-dimensional array or three-dimensional tensor (if each representation array itself is also two-dimensional). This concatenated array or tensor is the embedding array corresponding to the implicit representation sequence time series set. However, in practical applications, since the lengths of different sequences may be different (although the cloud service system may have made them consistent in length by interpolation or padding in step S310), the concatenation process may be more complicated. In addition, if the subsequent processing requires a fixed-size input and the size of the embedded array may vary, the cloud service system may also need to further process the embedded array (such as pooling, cropping or padding) to meet the requirements of subsequent processing. However, in this scenario, it is assumed that the lengths of all sequences are the same and the subsequent processing has no special requirements for the input size, so the cloud service system can simply concatenate all the representation arrays to form an embedded array.

[0151] Before concatenating the representation arrays, the cloud service system ensures that their dimensions are consistent. This includes consistency in terms of sequence length, feature dimension, etc. If some sequences are shorter, they may need to be made the same length as other sequences through interpolation or padding. The concatenation operation itself is relatively simple and usually just stacks multiple arrays along a specified dimension to form a larger array. However, in practical applications, due to memory and computing resource constraints, the cloud service system may need to process these arrays in batches instead of loading them all into memory at once.

[0152] Step S320 recursively embeds the implicit sequence time series set through the time recursive component and successfully converts the original time series data into an embedded array containing time dependency information. This process not only retains the key features in the original data, but also captures the continuity between sequences through the recursive mechanism, providing strong data support for the subsequent detection of traffic flow evolution trends.

[0153] As an implementation method, the method may further include:

[0154] Step S600: If the detection result of the traffic flow evolution trend corresponding to the target traffic flow intensity corresponds to the test traffic flow evolution trend, then the road time series monitoring data set corresponding to the target traffic flow intensity is determined to be a valid road time series monitoring data set; and the test traffic flow evolution trend is a correct traffic flow evolution trend;

[0155] Step S700: If the traffic flow evolution trend detection result corresponding to the target traffic flow intensity does not correspond to the test traffic flow evolution trend, the road time series monitoring data set corresponding to the target traffic flow intensity is determined to be an invalid road time series monitoring data set.

[0156] In step S600, the cloud service system compares the traffic flow evolution trend detection result corresponding to the target traffic flow intensity with the known test traffic flow evolution trend. The test traffic flow evolution trend is a correct trend based on historical data, expert experience or actual observation results, which represents the actual change of traffic flow on a specific road in a specific time period.

[0157] For example, suppose the cloud service system has analyzed a set of time-series monitoring data for a certain road and generated a traffic flow evolution trend detection result corresponding to the target traffic flow intensity. This detection result may indicate that the traffic flow on this road will gradually increase in the next few hours. To verify the accuracy of this detection result, the cloud service system will compare it with the known test traffic flow evolution trend.

[0158] The test traffic flow evolution trend may come from multiple aspects: on the one hand, it can be derived from statistical analysis based on historical data of the same period. For example, in the past few years, the traffic flow of the road in similar time periods has shown a similar upward trend; on the other hand, it can also be a prediction result obtained by experts based on comprehensive evaluation of factors such as road planning and surrounding traffic conditions; or, in some cases, the test traffic flow evolution trend may also come directly from the observation data of the real-time traffic monitoring system. During the comparison process, the cloud service system will use a series of quantitative indicators to evaluate the similarity or difference between the two trends. These indicators may include the slope of the trend line, the correlation coefficient, the mean square error (MSE), etc. If the cloud service system finds through comparison that the traffic flow evolution trend detection result corresponding to the target traffic flow intensity is highly consistent with the test traffic flow evolution trend (for example, the MSE value is lower than a preset threshold), it will determine that the road time series monitoring data set is a valid data set. This means that the data in the set is of high quality and can accurately reflect the actual changes in road traffic flow, providing a reliable basis for the formulation of subsequent lighting strategies.

[0159] If the comparison result in step S600 shows that the traffic flow evolution trend detection result corresponding to the target traffic flow intensity is significantly different from the test traffic flow evolution trend (i.e., it does not correspond), the cloud service system will enter step S700 to further evaluate the validity of the road time series monitoring data set and ultimately determine whether it is an invalid data set.

[0160] For example, continuing with the previous example, suppose the cloud service system finds during the comparison process that although the traffic flow evolution trend detection result corresponding to the target traffic flow intensity predicts an upward trend in traffic flow, the test traffic flow evolution trend indicates that traffic flow will remain stable or slightly decrease. In this case, the difference between the two trends is large, which may mean that there is a quality problem in the time series monitoring data set.

[0161] The cloud service system will further check whether there are any problems with the source, collection method, processing process, etc. of these data sets. For example, the data collection equipment may have failed or errors have accumulated; the data processing algorithm may have defects that lead to distorted results; or external factors (such as weather changes, special events, etc.) may have had an abnormal impact on traffic flow.

[0162] After excluding all possible external interference factors, if the cloud service system still cannot explain the difference between the detection results and the test trend, it will tend to consider the road time series monitoring data set as an invalid data set. This means that these data may not accurately reflect the actual changes in road traffic flow, and formulating lighting strategies based on these data may lead to misjudgment or waste of resources.

[0163] Therefore, in step S700, the cloud service system will record the invalid state of the data set and avoid using the invalid data in subsequent processing. At the same time, it will also send warning information to relevant departments or personnel, prompting them to check the possible problems in the data collection and processing process and take measures to repair and improve them as soon as possible.

[0164] Steps S600 and S700 verify the validity of the road time series monitoring data set by comparing the detection results of the traffic flow evolution trend with the test trend. These two steps not only help to ensure the accuracy of the subsequent lighting strategy formulation, but also promote the continuous improvement and optimization of the data collection and processing process, providing a solid data foundation for intelligent road lighting control based on vehicle-road-cloud collaboration.

[0165] As an implementation manner, the present application embodiment also provides a training process of the above-mentioned traffic flow evolution trend detection network, which may specifically include the following steps:

[0166] Step S10: obtaining a first basic traffic flow evolution trend detection network and a first template road time series monitoring data set corresponding to the template traffic flow intensity, and a first a priori evolution type detection result corresponding to the template traffic flow intensity in the first template road time series monitoring data set;

[0167] Step S20: extracting features of the time series distribution information corresponding to a plurality of first template road monitoring data sequences in the first template road time series monitoring data set through the first basic traffic flow evolution trend detection network, and obtaining a time series distribution sub-implicit representation sequence corresponding to each first template road monitoring data sequence;

[0168] Step S30: obtaining a template time sequence between a plurality of first template road monitoring data sequences, merging the time series distribution sub-implicit representation sequences corresponding to each first template road monitoring data sequence into a template implicit representation sequence time series set according to the template time sequence, extracting traffic flow variation features from the template implicit representation sequence time series set, and obtaining a first inference template traffic flow evolution implicit representation of the template traffic flow intensity in the first template road time series monitoring data set;

[0169] Step S40: performing evolution type detection on the implicit representation of the first inference template traffic flow evolution to obtain a first inference evolution type detection result corresponding to the template traffic flow intensity;

[0170] Step S50: determining a first training error corresponding to a first basic traffic flow evolution trend detection network based on a first a priori evolution type detection result and a first inference evolution type detection result corresponding to the template traffic flow intensity;

[0171] Step S60: In combination with the first training error, the first basic traffic flow evolution trend detection network is debugged in a loop, and the debugged first basic traffic flow evolution trend detection network is determined as the traffic flow evolution trend detection network; the traffic flow evolution trend detection network is used to obtain the traffic flow evolution trend detection result corresponding to the target traffic flow intensity in the road time series monitoring data set.

[0172] In step S10, the cloud service system prepares various resources required for training, including basic network models, training data, and prior knowledge. Specifically, the system obtains the following three parts:

[0173] First basic traffic flow evolution trend detection network: This is a preliminary constructed network model, usually based on time recursive neural network structures such as recurrent neural network (RNN), long short-term memory network (LSTM), gated recurrent unit (GRU), etc. The network has not been fully trained, and its parameters (such as weights and biases) are in a randomly initialized state or initialized based on some preset rules.

[0174] The first template road time series monitoring data set: This is a set of template traffic flow intensity and its corresponding time series monitoring data. These data usually come from actual road monitoring equipment or simulation environments, and are used to simulate traffic flow changes on different roads and in different time periods. Each data sequence contains traffic flow information at multiple time points, such as vehicle counts, average speeds, etc. This information is preprocessed and organized into a format suitable for network input.

[0175] First a priori evolution type detection result: This is the evolution type label of the template traffic flow intensity based on expert knowledge, historical data or other reliable sources. These labels directly indicate the changing trend of the template traffic flow intensity in different time periods, such as continuous growth, gradual decline, fluctuating changes, etc. These prior knowledge will serve as supervision signals in the training process to guide the learning direction of the network model. For example, suppose the cloud service system obtains a first basic traffic flow evolution trend detection network based on LSTM, and a first template road time series monitoring data set containing traffic flow monitoring data of a busy urban road within a week. The set contains traffic flow statistics for different time periods (such as every hour) every day. At the same time, the system also obtains the first a priori evolution type detection results obtained by traffic experts based on historical data and field observations, such as "the traffic flow continues to increase during the morning peak from Monday to Friday, and gradually decreases after the evening peak; the traffic flow is relatively stable throughout the day on weekends, but it increases slightly in the afternoon."

[0176] After preparing the training data and prior knowledge, in step S20, the cloud service system will use the first basic traffic flow evolution trend detection network to perform feature extraction on the data in the first template road time series monitoring data set. The purpose of this step is to extract high-level feature representations useful for traffic flow evolution trend identification from the original time series data.

[0177] The cloud service system inputs each data sequence in the first template road time series monitoring data set into the network in chronological order. Each data sequence may contain multiple feature dimensions (such as vehicle count, average speed, etc.), which together describe the changes in traffic flow.

[0178] The network processes the input data layer by layer through its internal multi-layer structure (such as LSTM layer, fully connected layer, etc.). In each layer, the data will undergo linear transformation (such as weight matrix multiplication) and nonlinear activation (such as ReLU, Sigmoid or Tanh function) and other operations, so as to gradually abstract higher-level feature representations. For time recursive networks such as LSTM, this process also takes into account the time dependency in the data, that is, the output at the current time point depends not only on the current input but also on the state at the previous time point.

[0179] After multiple layers of processing, the network generates a corresponding time series distribution sub-implicit representation sequence for each data sequence. This sequence contains high-level feature representations at each time point in the data sequence, which capture the inherent laws and trends of traffic flow over time.

[0180] For example, taking the LSTM network as an example, the cloud service system inputs the hourly traffic flow data (including vehicle counts and average speed) in the first template road time series monitoring data set as the input sequence into the network. The network recursively processes the sequence through its internal LSTM layer, and the output of each time step not only depends on the input at the current time point but also integrates the state information of all previous time points. After multiple layers of processing, the network generates a fixed-dimensional feature vector for each hour as an implicit representation of the time point. In the end, each data sequence corresponds to a time series distribution sub-implicit representation sequence.

[0181] After obtaining the time series distribution sub-implicit representation sequence corresponding to each first template road monitoring data sequence, in step S30, the cloud service system merges these sequences into a unified template implicit representation sequence time series set according to the template time order between these sequences. Subsequently, the system further processes this set using the feature extraction module in the network to extract the overall evolution characteristics of the template traffic flow intensity.

[0182] The cloud service system sorts and merges the data sequences according to their actual chronological order based on their timestamp information. The merged sequence set maintains the temporal continuity and relevance of the original data, providing a basis for subsequent feature extraction.

[0183] The merged sequence set is input into the feature extraction module in the network (which may be another LSTM layer, attention mechanism layer, or convolution layer, etc.) for processing. This module uses more complex network structures and nonlinear transformations to capture global features and long-term dependencies in the sequence set. Through this process, the system is able to generate an implicit representation vector or matrix that can fully reflect the evolution trend of the template traffic intensity.

[0184] For example, assume that the cloud service system has obtained the implicit representation sequence of the time series distribution corresponding to each first template road monitoring data sequence, and these sequences have been merged according to the actual time order. Subsequently, the system inputs these sequences into an additional LSTM layer for further processing. This LSTM layer not only considers the time dependency within each sequence, but also captures the long-term dependency between sequences through its internal state transfer mechanism. After processing, the system generates a fixed-length vector as the first inference template traffic flow evolution implicit representation of the template traffic flow intensity. This vector contains the overall evolution trend information of the template traffic flow intensity within a week. After obtaining the implicit representation of the first inference template traffic flow evolution, in step S40, the cloud service system uses a mechanism to detect the evolution type corresponding to this representation. This step usually involves inputting the implicit representation into a classifier or regressor and determining the final evolution type label based on its output. Specifically, the cloud service system designs a classifier suitable for the current task. This classifier can be a simple fully connected layer plus a Softmax activation function (for multi-classification problems) or a Sigmoid activation function (for binary classification problems); or it can be a more complex structure such as a decision tree, random forest, or support vector machine. After the implicit representation of the first inference template traffic flow evolution is input into the classifier, the classifier calculates the probability or score of each evolution type based on its internal parameters and activation function. Then the system selects the type with the highest probability or the highest score as the final evolution type detection result based on these probabilities or scores.

[0185] For example, suppose that the cloud service system designs a classifier based on a fully connected layer and a Softmax activation function to detect the evolution type. This classifier accepts the implicit representation of the evolution of the first inference template traffic flow as input, and calculates the probability value of each evolution type (such as continuous growth, gradual decline, fluctuating changes, etc.) through a series of linear transformations and nonlinear activation operations. The system then selects the type with the highest probability value as the final evolution type detection result. For example, if the probability value of the "continuous growth" type is 0.8 and the probability values ​​of other types are all lower than 0.2, the system considers the evolution type of the template traffic flow intensity to be "continuous growth".

[0186] After obtaining the first inference evolution type detection result, in step S50, the cloud service system compares it with the first a priori evolution type detection result, and calculates the first training error of the first basic vehicle flow evolution trend detection network based on the comparison result. This error reflects the gap between the current performance of the network model and the ideal performance, and is an important basis for subsequent debugging and optimization of network parameters.

[0187] Specifically, the cloud service system uses a preset error calculation method (such as cross entropy loss, mean square error, etc.) to quantify the difference between the first inference evolution type detection result and the first prior evolution type detection result. Commonly used error calculation methods for classification problems include cross entropy loss (for multi-classification problems) and binary cross entropy loss (for binary classification problems). These loss functions can capture the inconsistency between the predicted label and the true label and give a specific numerical measurement.

[0188] After calculating the error value, the cloud service system will further analyze it to determine the main source and possible causes of the error. For example, the error may come from data noise, insufficient model complexity, or overfitting. By analyzing the error, the cloud service system can provide guidance for subsequent model debugging and optimization.

[0189] For example, suppose the cloud service system uses cross entropy loss as the error calculation method. It first constructs a loss function expression based on the first reasoning evolution type detection result and the first prior evolution type detection result, and substitutes the relevant parameters into the expression to calculate the specific loss value (i.e., training error). Then the system analyzes this loss value and finds that the error mainly comes from inaccurate prediction results in certain specific time periods. This may be due to the large data noise in these time periods or the model's insufficient feature representation ability for these time periods. By analyzing the error, the cloud service system decides to focus on data processing and model optimization issues in these time periods during subsequent debugging.

[0190] After calculating the first training error, in step S60, the cloud service system will perform cyclic debugging and optimization on the first basic traffic flow evolution trend detection network according to the error value to improve its performance until the preset requirements are met. This step usually involves multiple sub-steps such as parameter adjustment, structure optimization and regularization processing, and may require multiple iterations to complete.

[0191] For example, the cloud service system attempts to adjust parameters in the network, such as weights and biases, to improve the predictive performance of the model. Parameter adjustment can be achieved automatically through gradient descent methods (such as SGD, Adam and other optimization algorithms) or manually through trial and error. During the adjustment process, the system will continuously monitor the changes in training error and verification error to determine the best parameter configuration. If simple parameter adjustments cannot meet performance requirements, the cloud service system may need to consider optimizing the network structure. Structural optimization may include increasing or decreasing the number of network layers, changing the connection method between layers, or introducing new network components. Through structural optimization, the cloud service system can further improve the expressiveness and generalization capabilities of the model.

[0192] In order to prevent the model from overfitting, the cloud service system also needs to regularize the model. Regularization can be achieved by adding regularization terms (such as L1 regularization, L2 regularization, etc.) to the loss function or by using techniques such as dropout. These techniques help reduce the complexity of the model and improve its generalization ability for unseen data.

[0193] All the above optimization steps need to be repeated in the training process until the preset stop condition is reached (such as the training error is lower than a certain threshold, the verification error no longer decreases, or the preset number of iterations is reached). In each iteration, the cloud service system will recalculate the training error and update the model parameters and structure according to the error value to gradually approach the optimal solution.

[0194] For example, suppose that after calculating the first training error, the cloud service system finds that the model's prediction results in certain time periods are inaccurate, resulting in a high overall error. In order to reduce the error, the system first tried to adjust the weight parameters of the forget gate and input gate in the LSTM layer to improve the model's ability to capture time dependencies. However, after multiple adjustments, the error still did not show a clear downward trend. So the system decided to optimize the network structure and tried to add an additional LSTM layer to enhance the model's memory and expression capabilities. At the same time, in order to prevent the newly added layer from causing overfitting problems, the system also added an L2 regularization term to the loss function to constrain the weight parameters. After multiple rounds of iterative optimization, the cloud service system finally obtained a traffic flow evolution trend detection network model with performance that met the requirements, and determined it as the traffic flow evolution trend detection network that will eventually be used for actual road smart lighting control.

[0195] As an implementation method of the above training process, the method provided in the present application also includes a training method of contrastive learning, which specifically includes the following steps:

[0196] Step S1: outputting a template traffic flow intensity through a second basic traffic flow evolution trend detection network to implicitly represent the second inference template traffic flow evolution in a second template road time series monitoring data set; the network configuration variables in the second basic traffic flow evolution trend detection network are the same as the network configuration variables in the first basic traffic flow evolution trend detection network;

[0197] Step S2: performing evolution type detection on the implicit representation of the second inference template traffic flow evolution, and obtaining a second inference evolution type detection result corresponding to the template traffic flow intensity in the second template road time series monitoring data set;

[0198] Step S3: generating a second training error based on the second inference evolution type detection result and the second a priori evolution type detection result corresponding to the second template road time series monitoring data set;

[0199] Step S4: generating a third training error based on the first reasoning evolution type detection result and the second reasoning evolution type detection result;

[0200] Based on this, the aforementioned step S60, combined with the first training error, performs a loop debugging on the first basic vehicle flow evolution trend detection network, and determines the debugged first basic vehicle flow evolution trend detection network as the vehicle flow evolution trend detection network, which can specifically include: combining the first training error, the second training error and the third training error to obtain a fused training error, and based on the fused training error, performs a loop debugging on the first basic vehicle flow evolution trend detection network, and determines the debugged first basic vehicle flow evolution trend detection network as the vehicle flow evolution trend detection network.

[0201] The above implementation introduces a comparative learning training method, which can significantly improve the generalization ability and robustness of the traffic flow evolution trend detection network. By using two or more basic networks with the same network configuration but facing different data sets for comparative learning, the cloud service system can capture more subtle data differences and pattern changes, and then optimize the parameters and structure of the network model.

[0202] Specifically, in step S1, the cloud service system deploys a second basic traffic flow evolution trend detection network with the same structure as the first basic traffic flow evolution trend detection network but with different initial weight values. This network is used to process another second template road time series monitoring data set that is different from the first template road time series monitoring data set. Through this parallel processing process, the system aims to explore the impact of different data sets on network output and lay the foundation for subsequent comparative learning.

[0203] For example, suppose the first template road time series monitoring data set records the traffic flow change data of a city's main road within a week, while the second template road time series monitoring data set records the traffic flow change data of another secondary road in the same city during the same time period. Although the traffic flow change trends of the two roads may be similar (such as being affected by morning and evening peaks), their specific change patterns may also be different due to differences in road type, surrounding environment, traffic control and other factors.

[0204] The cloud service system first processes the second template road time series monitoring data set through the second basic traffic flow evolution trend detection network. This network also contains multiple LSTM layers, fully connected layers and other components to extract features from time series data and generate high-level representations. During the processing, the network encodes the traffic flow data at each time point, and finally outputs a second inference template traffic flow evolution implicit representation representing the traffic flow evolution trend of the secondary road within a week. This representation may be a fixed-length vector or matrix that captures the dynamic change characteristics of traffic flow over time.

[0205] After obtaining the implicit representation of the evolution of the traffic flow of the second inference template, in step S2, the cloud service system uses the same classifier or regressor as in step S4 to detect the evolution type of the representation. The purpose of this step is to determine the evolution type (such as continuous growth, gradual decline, fluctuating changes, etc.) of the traffic flow intensity represented by the second template road time series monitoring data set in practice.

[0206] For example, the cloud service system uses a trained classifier (such as a fully connected layer based on the Softmax activation function) to process the implicit representation of the traffic flow evolution of the second reasoning template. The classifier calculates the probability value of each evolution type based on the input feature representation, and selects the type with the highest probability as the final evolution type detection result. For example, if the classifier outputs a probability of 0.6 for the "continuous growth" type, a probability of 0.3 for the "gradual decline" type, and a probability of 0.1 for the "fluctuating change" type, the system considers the traffic flow evolution type of the secondary trunk road within a week to be "continuous growth".

[0207] In step S3, the cloud service system compares the second inference evolution type detection result with the second prior evolution type detection result corresponding to the second template road time series monitoring data set, and calculates the second training error. This error reflects the performance of the second basic traffic flow evolution trend detection network when processing the second template data set. For example, assume that the second prior evolution type detection result is obtained by traffic experts based on historical data and field observations, and clearly points out that the traffic flow on this secondary trunk road within a week shows an evolution trend of "first increase and then decrease". The cloud service system compares this result with the second inference evolution type detection result (such as "continuous growth") and calculates the difference between the two (such as through the cross entropy loss function). This difference quantifies the degree of deviation between the prediction accuracy of the network model when processing the second template data set and the actual situation.

[0208] In step S4, in order to further enhance the comparative learning effect, the cloud service system will also compare the differences between the inference evolution type detection results obtained by the two basic networks when processing different template data sets, and generate a third training error. This error is intended to evaluate the consistency of the two networks when processing similar but not identical tasks. For example, the cloud service system first obtains the first inference evolution type detection result (such as the "continuous growth" type for the first template data set) and the second inference evolution type detection result (such as the "continuous growth" or "first increase and then decrease" type for the second template data set). Then the system compares whether the two results are consistent or similar, and calculates the third training error accordingly. This error may be calculated based on a similarity metric (such as cosine similarity) or a direct category matching degree. If the two results are highly consistent, the error is small; otherwise, the error is large. After obtaining the first training error, the second training error, and the third training error, the cloud service system performs a weighted sum (or processes according to other fusion strategies) on these three error values ​​to generate a fusion training error. This fusion error comprehensively reflects the overall performance of the network model when processing different data sets and tasks. The system then uses this fusion error as an optimization target to perform cyclic debugging and optimization on the first basic traffic flow evolution trend detection network until the preset performance requirements are met.

[0209] For example, the cloud service system sets a certain weight coefficient (such as determined by experience or experiment) to balance the importance of the three training errors. For example, the weight of the first training error can be set to 0.5 to represent its main influencing factor; the weight of the second training error can be set to 0.3 to represent its secondary influencing factor; the weight of the third training error can be set to 0.2 to represent its auxiliary influencing factor. Then the three error values ​​are weighted and summed according to their respective weights to obtain the fused training error. By introducing the second basic traffic flow evolution trend detection network and using multiple training errors for joint optimization, the performance and generalization ability of the network model are significantly improved. This process not only helps to capture the differences and similarities between different data sets, but also helps to enhance the adaptability and robustness of the network when processing new tasks.

[0210] See also Figure 3 , is a schematic diagram of the structure of a cloud service system provided in an embodiment of the present application. Figure 3As shown, the cloud service system 10 may include: a processor 1001, a network interface 1004 and a memory 1005. In addition, the cloud service system 10 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 1005 may optionally be at least one storage device away from the aforementioned processor 1001. As Figure 3 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application. Figure 3 In the cloud service system 10 shown, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the method provided in the above embodiment.

[0211] It should be understood that the cloud service system 10 described in the embodiments of the present application can execute the above Figure 2 The description of the intelligent road lighting control method based on vehicle-road-cloud collaboration in the corresponding embodiment will not be repeated here.

Claims

1. A road intelligent lighting control method based on vehicle-road-cloud collaboration, characterized in that: Applied to a cloud service system, the method includes: Acquire a road time series monitoring data set containing a target vehicle flow intensity, perform vehicle flow identification on multiple road monitoring data sequences in the road time series monitoring data set, and obtain time series distribution information of the target vehicle flow intensity in each road monitoring data sequence; wherein the road time series monitoring data set is collected by road-side monitoring equipment and vehicle monitoring equipment; Performing feature extraction on the time series distribution information of the target vehicle flow intensity in each road monitoring data sequence to obtain a time series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence; Obtaining the time sequence between the multiple road monitoring data sequences, merging the time series distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence, extracting the traffic flow change characteristics of the implicit representation sequence time series set, and obtaining the implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set; Performing evolution type detection on the implicit representation of the traffic flow evolution to obtain a traffic flow evolution trend detection result corresponding to the target traffic flow intensity; Formulate a road lighting strategy for a future target time period based on the traffic flow evolution trend detection result; The plurality of road monitoring data sequences include a road monitoring data sequence S m , m is not greater than the number G of the plurality of road monitoring data sequences, 1≤m≤G; the vehicle flow identification is performed on the plurality of road monitoring data sequences in the road time series monitoring data set to obtain the time series distribution information of the target vehicle flow intensity in each road monitoring data sequence, including: The road monitoring data sequence S m Perform arbitrary segmentation to obtain K time series intervals to be selected, where K ≥ 1; Extracting data items from the K time series intervals to be selected respectively, to obtain interval data items corresponding to the K time series intervals to be selected respectively; According to the numerical range of the data item related to the target traffic flow intensity, the interval data items respectively contained in the K candidate time series intervals are divided into data items to obtain the target data item in each candidate time series interval, and the road monitoring data sequence S is divided based on the target data item. m Sequence segmentation is performed to obtain a binary tag set corresponding to the target traffic flow intensity; the data item value corresponding to the target data item belongs to the data item value range; Based on the binary tag set, it is determined that the target vehicle flow intensity is in the road monitoring data sequence S m The time series distribution information in ; The feature extraction of the time series distribution information of the target vehicle flow intensity in each road monitoring data sequence to obtain the time series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence includes: Merging the time series distribution information of the target vehicle flow intensity in multiple road monitoring data sequences to obtain a time series distribution array corresponding to the target vehicle flow intensity, inputting the time series distribution array into a feature extraction component in a vehicle flow evolution trend detection network, filtering the time series distribution array, and obtaining time series filtering information corresponding to the time series distribution array; Performing a normalization operation on the time series filtering information to obtain normalized time series filtering information; Based on the nonlinear transformation operator in the feature extraction component, the time series filtering information after the normalization operation is activated to generate a time series distribution implicit representation sequence corresponding to the time series distribution information of the target vehicle flow intensity in the multiple road monitoring data sequences; Based on the time series distribution implicit representation sequence, determining the time series distribution sub-implicit representation sequence corresponding to each road monitoring data sequence; The step of obtaining the time sequence between the plurality of road monitoring data sequences, merging the time series distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence, extracting the traffic flow variation characteristics of the implicit representation sequence time series set, and obtaining the implicit representation of traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set includes: Acquire a time sequence between the multiple road monitoring data sequences, and merge the time series distribution sub-implicit representation sequences corresponding to each road monitoring data sequence into an implicit representation sequence time series set according to the time sequence; Based on the time recursive component of the traffic flow evolution trend detection network, the implicit representation sequence time series set is recursively embedded to obtain an embedding array corresponding to the implicit representation sequence time series set; Based on the nonlinear transformation operator in the time recursive component, the embedded array is activated to obtain an implicit representation of the traffic flow evolution of the target traffic flow intensity in the road time series monitoring data set.

2. The method according to claim 1, characterized in that The nonlinear transformation operator in the feature extraction component is used to activate the time series filtering information after the normalization operation to generate a time series distribution implicit representation sequence corresponding to the time series distribution information of the target vehicle flow intensity in the multiple road monitoring data sequences, including: Based on the nonlinear transformation operator in the feature extraction component, the time series filtering information after the normalization operation is activated to obtain a candidate time series distribution implicit representation sequence corresponding to the time series distribution information in the multiple road monitoring data sequences; The output dimension and output scale in the feature extraction component are obtained, and based on the output dimension and the output scale, the candidate time series distribution implicit representation sequence is optimized to obtain a time series distribution implicit representation sequence corresponding to the time series distribution information of the target vehicle flow intensity in the multiple road monitoring data sequences; the scale corresponding to the time series distribution implicit representation sequence is the product of the output scale and the output dimension.

3. The method according to claim 1, characterized in that The step of performing a normalization operation on the time series filtering information to obtain the time series filtering information after the normalization operation includes: Based on the canonical operator in the feature extraction component, clustering the time series filtering information to obtain B time series filtering clusters, where B≥1; The expected values ​​and dispersions respectively corresponding to the B time series filter clusters are obtained, and based on the expected values ​​and dispersions respectively corresponding to the B time series filter clusters, the time series filter information in each time series filter cluster is standardized to obtain the time series filter information after the standardized operation.

4. The method according to claim 1, characterized in that: The plurality of road monitoring data sequences include a road monitoring data sequence H m and road monitoring data sequence H m+1 , the road monitoring data sequence H m With the road monitoring data sequence H m+1 adjacent, m+1 is not greater than the number G of the plurality of road monitoring data sequences, 1≤m≤G; The time recursive component based on the traffic flow evolution trend detection network recursively embeds the implicit representation sequence time series set to obtain an embedding array corresponding to the implicit representation sequence time series set, including: Obtain the road monitoring data sequence H in the implicit representation sequence time series set m The corresponding first memory state array stores the road monitoring data sequence H m The corresponding time series distribution sub-implicit representation sequence and the first memory state array are loaded into the time recursive component of the traffic flow evolution trend detection network, and the road monitoring data sequence H m The corresponding time series distribution implicit representation sequence is embedded to obtain the road monitoring data sequence H m The corresponding representation array; Based on the road monitoring data sequence H m The corresponding representation array determines the road monitoring data sequence H in the implicit representation sequence time series set m+1 The corresponding second memory state array stores the road monitoring data sequence H m+1 The corresponding time series distribution sub-implicit representation sequence and the second memory state array are loaded into the time recursive component of the traffic flow evolution trend detection network, and the road monitoring data sequence H m+1 The corresponding time series distribution implicit representation sequence is embedded to obtain the road monitoring data sequence H m+1 The corresponding representation array; If the road monitoring data sequence H m+1 When it is the last road monitoring data sequence among the multiple road monitoring data sequences, an embedding array corresponding to the implicit representation sequence time series set is generated based on the representation array corresponding to each road monitoring data sequence in the multiple road monitoring data sequences.

5. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: Obtaining a first basic traffic flow evolution trend detection network and a first template road time series monitoring data set corresponding to the template traffic flow intensity, and a first a priori evolution type detection result corresponding to the template traffic flow intensity in the first template road time series monitoring data set; By using the first basic traffic flow evolution trend detection network, feature extraction is performed on the time series distribution information corresponding to multiple first template road monitoring data sequences in the first template road time series monitoring data set to obtain a time series distribution sub-implicit representation sequence corresponding to each first template road monitoring data sequence; Acquire the template time sequence between the multiple first template road monitoring data sequences, merge the time series distribution sub-implicit representation sequences corresponding to each first template road monitoring data sequence into a template implicit representation sequence time series set according to the template time sequence, extract the traffic flow change features of the template implicit representation sequence time series set, and obtain the first inference template traffic flow evolution implicit representation of the template traffic flow intensity in the first template road time series monitoring data set; Performing an evolution type detection on the implicit representation of the first inference template traffic flow evolution to obtain a first inference evolution type detection result corresponding to the template traffic flow intensity; Determining a first training error corresponding to the first basic traffic flow evolution trend detection network based on a first a priori evolution type detection result and a first inference evolution type detection result corresponding to the template traffic flow intensity; In combination with the first training error, the first basic traffic flow evolution trend detection network is debugged in a loop, and the debugged first basic traffic flow evolution trend detection network is determined as the traffic flow evolution trend detection network; the traffic flow evolution trend detection network is used to obtain the traffic flow evolution trend detection result corresponding to the target traffic flow intensity in the road time series monitoring data set.

6. The method according to claim 5, characterized in that The method further comprises: Outputting the template traffic flow intensity in the second template road time series monitoring data set through the second basic traffic flow evolution trend detection network as a second inference template traffic flow evolution implicit representation; the network configuration variables in the second basic traffic flow evolution trend detection network are the same as the network configuration variables in the first basic traffic flow evolution trend detection network; Performing an evolution type detection on the implicit representation of the evolution of the second inference template traffic flow to obtain a second inference evolution type detection result corresponding to the template traffic flow intensity in the second template road time series monitoring data set; generating a second training error based on the second inference evolution type detection result and a second a priori evolution type detection result corresponding to the second template road time series monitoring data set; generating a third training error based on the first reasoning evolution type detection result and the second reasoning evolution type detection result; The step of cyclically debugging the first basic vehicle flow evolution trend detection network in combination with the first training error, and determining the debugged first basic vehicle flow evolution trend detection network as the vehicle flow evolution trend detection network, includes: The first training error, the second training error and the third training error are combined to obtain a fusion training error. The first basic vehicle flow evolution trend detection network is cyclically debugged based on the fusion training error, and the debugged first basic vehicle flow evolution trend detection network is determined as the vehicle flow evolution trend detection network.

7. A cloud service system, characterized in that: include: processor; and a memory, wherein the memory stores a computer-readable code, and when the computer-readable code is executed by the processor, the processor executes the method as claimed in any one of claims 1 to 6.

Images